Deploy v1 — single-Docker FastAPI + Next.js + RAG + voice + faithfulness

kb/AUDIT_TRAIL.md

@@ -0,0 +1,175 @@
+# Audit Trail — End-to-End Data Lineage
+
+| Field | Value |
+| --- | --- |
+| Project | Insurance Sales Portfolio Expert |
+| Doc version | 0.1 |
+| Status | Living document — regenerated by `rag/build_kb.py` |
+
+## 0. Why this doc exists
+
+Per the prime directive: **every fact the bot states must trace back to a verifiable source.** This document is the **single end-to-end map of how raw insurer data becomes a live bot reply** — with every transformation, computation, and artifact named explicitly.
+
+For a Sarvam interviewer or BFSI compliance auditor: open this file, follow the trail from "user asks 'what is the cataract waiting period?'" all the way back to "Star Health's PDF page 18, section 4.2.7". Every step is reproducible from the repo.
+
+## 1. The pipeline — 10 stages
+
+```
+[1]                       [2]                       [3]
+SOURCE PDFs               DISCOVERY                 DOWNLOAD
+- 10 insurer websites  →  - corpus_discovery   →    - rag/download_corpus.py
+- IRDAI database          agent (75 URLs found)     - HEAD-check + magic byte
+- regulatory portals      - data/corpus_urls.md     - retry logic for 403/timeout
+                                                    → rag/corpus/*.pdf
+
+[4]                       [5]                       [6]
+PARSE                     CHUNK                     EMBED
+- pdfplumber              - 800 tok / 120 overlap   - BGE-small-en-v1.5 local
+- per-page text           - page-aware              - 384-dim cosine vectors
+- whitespace normalize    - sentence boundary       - no rate limits
+
+[7]                       [8]                       [9]
+VECTOR INDEX              STRUCTURED EXTRACTION     SCORECARD
+- Chroma persistent       - Sarvam-M extraction     - rules-based 6-sub-score
+  client                    with HealthPolicy        weighted aggregation
+- metadata per chunk        Pydantic schema         - A-F grade
+- rag/vectors/*.sqlite3   - DeepSeek-V3 fallback    - 24 of 48 fields used
+                          → rag/extracted/*.json
+                          → rag/policies.duckdb
+
+[10]                      [11]                      [12]
+RETRIEVAL                 ORCHESTRATION             FAITHFULNESS GATES
+- Voyage at query time    - intent classifier       - Gate 1: retrieval floor
+  (BGE for matching)      - brain router (Sarvam-M  - Gate 2: citation integrity
+- top-k cosine            → Llama/DeepSeek)         - Gate 3: regex numeric
+- per-policy filter       - persona prompt           grounding
+                          - chat history             - Gate 4: LLM-judge
+
+[13]                      [14]                      [15]
+RESPONSE                  TTS                       UI RENDER
+- reply_text + citations  - Sarvam Bulbul          - chat bubble
+- brain used, latency     - audio response          - source links per
+- faithfulness verdict    - language match           citation
+                                                    - scorecard (P1)
+```
+
+## 2. Every artifact at every stage
+
+| Stage | Artifact | Path | Generated by | Verifiable by |
+| --- | --- | --- | --- | --- |
+| 1 SOURCE | Raw insurer PDFs | `rag/corpus/<insurer>/*.pdf` | manual + agent crawl | open URL in browser; HEAD-check via `tools/verify_urls.py` |
+| 2 DISCOVERY | URL index | `data/corpus_urls.md` | `Agent` task `a9594655...` (research agent) | git history of discovery agent |
+| 2 DISCOVERY | Regulatory URLs | `data/regulatory_urls.md` | `Agent` task `a87f8380...` | git history |
+| 3 DOWNLOAD | Manifest | `rag/corpus/_manifest.json` | `rag/download_corpus.py` + `rag/download_retry.py` | re-run download script |
+| 4 PARSE | per-page text (in-memory) | _ephemeral_ | `rag/ingest.py:read_pdf_pages` | open PDF; pdfplumber on it |
+| 5 CHUNK | chunks (in-memory) | _ephemeral_ | `rag/ingest.py:chunk_pages` | run ingest with `--dry-run` |
+| 6 EMBED | 384-dim vectors | _embedded in Chroma_ | `backend/providers/local_embeddings.py` | re-encode the same text |
+| 7 VECTOR INDEX | persistent sqlite | `rag/vectors/chroma.sqlite3` + HNSW binaries | `chromadb.PersistentClient` | open with chromadb client |
+| 8 STRUCTURED EXTRACTION | 48-field JSON per policy | `rag/extracted/<policy_id>.json` | `rag/extract.py` (Sarvam-M → DeepSeek-V3 fallback) | re-run extraction; compare |
+| 8 STRUCTURED EXTRACTION | aggregate table | `rag/policies.duckdb` | upsert in `rag/extract.py` | `duckdb` CLI query |
+| 9 SCORECARD | per-policy grade | `kb/policies/<policy_id>.md` (live in code via `backend/scorecard.py`) | `rag/build_kb.py` + `backend/scorecard.py` | re-run `build_kb` |
+| 10 RETRIEVAL (runtime) | top-k chunks | _ephemeral, logged_ | `rag/retrieve.py` | replay query against Chroma |
+| 11 ORCHESTRATION (runtime) | classified intent + brain | _logged_ | `backend/orchestrator.py` | `logs/turns.jsonl` |
+| 12 FAITHFULNESS (runtime) | gate verdict | _logged_ | `backend/faithfulness.py` | `logs/hallucinations.jsonl` |
+| 13 RESPONSE | bot reply + cited chunks | _returned to UI_ | `backend/main.py:/api/chat` | curl the API |
+| 14 TTS (runtime) | base64 WAV | _returned to UI_ | `backend/providers/sarvam_tts.py` | save + play |
+| 15 UI (runtime) | rendered chat | _browser DOM_ | `frontend/src/app/page.tsx` | browser dev-tools |
+
+## 3. Verification per artifact type
+
+### Source PDFs (Stage 1)
+- **Provenance**: published by the insurer; downloaded from their public CDN
+- **Verifiable**: `eval/verified_urls.json` (HEAD-checks for live status)
+- **Risk if tampered**: corpus poisoning. Mitigation: filename + size + page-count audit, vs original.
+
+### Vector chunks (Stages 5-7)
+- **Provenance**: deterministic chunking of parsed text from Stage 4
+- **Verifiable**: given the same PDF + same `CHUNK_TOKENS/OVERLAP` config in `backend/config.py`, output is identical
+- **Risk if tampered**: silent retrieval poisoning. Mitigation: chunk content stored alongside the embedding; can be re-verified against source PDF on demand.
+
+### Extracted structured data (Stage 8)
+- **Provenance**: LLM (Sarvam-M / DeepSeek-V3) over full PDF text + Pydantic schema
+- **Verifiable**: `extraction_confidence_pct` field per record + re-runnable in <30s with `python -m rag.extract --policy <id>`
+- **Risk if tampered**: bad downstream scorecard + filter results. Mitigation: schema validates types; manual spot-check 5%.
+
+### Scorecard (Stage 9)
+- **Provenance**: pure function of the extracted JSON via `backend/scorecard.py`
+- **Verifiable**: rules-based, no LLM — anyone can re-run on the JSON
+- **Methodology**: [`docs/scorecard-methodology.md`](../docs/scorecard-methodology.md)
+
+### Live bot replies (Stages 10-15)
+- **Provenance**: traceable via `logs/turns.jsonl` per session/turn
+- **Verifiable**: replay query → compare cited chunk_ids → confirm chunk text in Chroma → trace to source PDF
+- **Faithfulness audit**: `logs/hallucinations.jsonl` records every blocked reply with the failing gate
+
+## 4. Decision-to-artifact map
+
+Every architectural decision in `docs/decisions.md` produces a specific artifact:
+
+| Decision | Artifact produced |
+| --- | --- |
+| D-001 Vertical slice scope | `kb/policies/` (10 categories only Health) |
+| D-002 Health category | `data/corpus_urls.md` (76 health PDFs) |
+| D-003 Curated corpus | `rag/corpus/` (no user uploads in v1, expanded in v1.1 with security gates) |
+| D-004 Hybrid structured + unstructured | `rag/policies.duckdb` + `rag/vectors/` |
+| D-005 Next.js + FastAPI | `frontend/` + `backend/main.py` |
+| D-006 Sarvam-first | `backend/providers/sarvam_*.py` |
+| D-007 Pricing illustrative only | (no `actual_premium` field anywhere) |
+| D-008 Consultative persona | `backend/persona.py` |
+| D-009 10 insurers / all health policies | `data/corpus_urls.md` (75 URLs) |
+| D-010 Secret handling | `.env` (chmod 600 + gitignored) |
+| D-011 Voyage embeddings | _superseded by local BGE; see D-011 revision_ |
+| D-012 Render hosting | _superseded by HF Spaces_ |
+| D-013 Tailwind + shadcn | `frontend/src/app/globals.css` |
+| D-014 Groq Llama grader | `backend/providers/groq_llm.py` |
+| D-015 OpenAPI codegen | `backend/main.py` (auto-served) |
+| D-016 Brain router | `backend/orchestrator.py:pick_brain` |
+| D-017 Regulatory corpus deferred | `docs/04-failure-modes.md` F-07 |
+
+## 5. What you can't audit (yet)
+
+Honesty about gaps:
+
+- **Insurer-side PDF tampering** — we trust the insurer's published PDF was real at download time. We don't have a re-fetch + diff pipeline yet. v2 enhancement.
+- **LLM determinism** — Sarvam-M / DeepSeek-V3 can produce slightly different extraction across runs even at `temperature=0`. We pin model versions in `backend/config.py`; for full reproducibility we'd cache extraction outputs. v2 enhancement.
+- **Embedding model drift** — if BGE updates, our cached vectors become stale relative to query-time embeddings. v2: pin model commit hash.
+- **Chroma version compatibility** — different `chromadb` releases store metadata differently (we hit `KeyError: '_type'` on first HF deploy). Pinned to `0.5.20` in `requirements.txt`.
+
+## 6. How to use this doc
+
+If asked "where did this number come from?":
+1. Find the policy in `kb/policies/<policy_id>.md`
+2. Find the field — it has a derivation tag `[E]` / `[C]` / `[I]` / `[V]`
+3. Follow the Lineage section in that file to find the producing script + source PDF page
+
+If asked "did you make this up?":
+1. Open `logs/hallucinations.jsonl` — every blocked claim is recorded
+2. Open `eval/results.md` — accuracy numbers per question type
+3. Open `eval/verified_urls.json` — proof every URL we surface is real
+
+## 7. Regenerating this entire trail
+
+```bash
+# Re-acquire corpus
+python -m rag.download_corpus
+python -m rag.download_retry
+
+# Re-extract (LLM)
+python -m rag.extract
+
+# Re-build vector store
+rm -rf rag/vectors
+python -m rag.ingest
+
+# Re-build KB sheets (this directory)
+python -m rag.build_kb
+
+# Re-run eval
+python -m eval.generate_gold
+python -m eval.run
+
+# Re-verify URLs
+python tools/verify_urls.py
+```
+
+Total cost from cold: <$2 (BGE local + ~80 LLM extractions). Total wall-time: ~30-40 min on a modern laptop.

kb/INDEX.md

@@ -0,0 +1,55 @@
+# Knowledge Base — Master Index
+
+_Generated 2026-05-12T23:12:11Z. Auto-regenerable via `python -m rag.build_kb`._
+
+This is the **single canonical KB** for this project. Every data point in the bot
+(citations, scorecards, comparison views) traces back to one of these files.
+
+## Layout
+
+```
+kb/
+├── INDEX.md                          (this file)
+├── policies/<policy_id>.md          (11 files — one per extracted policy)
+├── research/
+│   ├── corpus_acquisition.md         (how we got 75 PDFs)
+│   ├── url_verification.md           (HEAD-check results)
+│   └── verified_insurers.md          (10 insurers, home URLs)
+└── calculations/
+    ├── scorecard_results.md          (all scores)
+    ├── eval_results.md               (gold Q&A grader output)
+    └── extraction_quality_audit.md   (per-field completeness)
+```
+
+## Quick links
+
+- **All policies (graded):** [`calculations/scorecard_results.md`](calculations/scorecard_results.md)
+- **All policy KB sheets:** [`policies/`](policies/)
+- **Eval run results:** [`calculations/eval_results.md`](calculations/eval_results.md)
+- **Extraction quality:** [`calculations/extraction_quality_audit.md`](calculations/extraction_quality_audit.md)
+- **URL verification:** [`research/url_verification.md`](research/url_verification.md)
+- **Corpus acquisition:** [`research/corpus_acquisition.md`](research/corpus_acquisition.md)
+
+## Derivation conventions
+
+Every field in every KB file is tagged with one of:
+- **[E]** Extracted directly from a source PDF
+- **[E?]** Extractable in the schema but absent / null in this specific source
+- **[C]** Computed from extracted fields (e.g. scorecard score)
+- **[I]** Implied / canonicalised by us (e.g. insurer slug)
+- **[V]** Externally verified (HEAD-check, URL probe)
+
+## Headline counts
+
+- Policies extracted: **11**
+- Insurers covered: **4**
+- Grade distribution: {'B': 5, 'C': 6}
+
+## Why we maintain this in markdown
+
+JSON is for machines. Markdown is for reviewers. Each KB file is intentionally
+human-readable so an interviewer or auditor can open `kb/policies/<some-id>.md`
+and read every data point with its source — without running the bot.
+
+The bot's runtime answers are NEVER allowed to use information that isn't
+traceable to one of these files (see `backend/faithfulness.py`).

kb/calculations/eval_results.md

@@ -0,0 +1,31 @@
+# Calculations — Eval Run Results
+
+_Most recent gold Q&A eval run at 2026-05-12T22:30:15Z_
+
+## Headline
+- Questions: **25**
+- Factual accuracy: **40.0%**
+- Citation accuracy: **50.0%**
+- Refusal precision: **44.4%**
+- Blocked by faithfulness: 12
+- Elapsed: 293.0s
+
+## By question type
+
+| Type | Accuracy |
+| --- | --- |
+| coverage_scope | 100.0% |
+| regulatory_oos | 66.7% |
+| sub_limit | 33.3% |
+| exclusions_oos | 33.3% |
+| waiting_period | 12.5% |
+| bonus | 0.0% |
+
+## By brain
+
+| Brain | Accuracy |
+| --- | --- |
+| groq-llama | 100.0% |
+| sarvam-m | 37.5% |
+
+Full per-question results: [`eval/results.md`](../../eval/results.md) and [`eval/results.json`](../../eval/results.json).

kb/calculations/extraction_quality_audit.md

@@ -0,0 +1,70 @@
+# Calculations — Extraction Quality Audit
+
+_Computed from `rag/extracted/*.json` (11 files)._
+
+How often each of the 48 schema fields actually got populated by extraction. Low-completeness fields are the ones to harden in v2 (better prompts, or LLM router).
+
+| Field | Populated | % |
+| --- | --- | --- |
+| `policy_id` | 11/11 | 100% |
+| `insurer_name` | 11/11 | 100% |
+| `insurer_slug` | 11/11 | 100% |
+| `policy_name` | 11/11 | 100% |
+| `policy_type` | 1/11 | 9% |
+| `uin_code` | 10/11 | 91% |
+| `min_entry_age_years` | 0/11 | 0% |
+| `max_entry_age_years` | 2/11 | 18% |
+| `max_renewal_age_years` | 2/11 | 18% |
+| `min_child_entry_age_days` | 3/11 | 27% |
+| `family_composition_allowed` | 0/11 | 0% |
+| `residency_requirement` | 1/11 | 9% |
+| `sum_insured_options_inr` | 4/11 | 36% |
+| `premium_payment_modes` | 1/11 | 9% |
+| `premium_range_indicative_inr` | 1/11 | 9% |
+| `premium_payment_term_years` | 3/11 | 27% |
+| `grace_period_days` | 5/11 | 45% |
+| `free_look_period_days` | 3/11 | 27% |
+| `initial_waiting_period_days` | 6/11 | 55% |
+| `pre_existing_disease_waiting_months` | 11/11 | 100% |
+| `specific_disease_waiting_months` | 8/11 | 73% |
+| `specific_diseases_listed` | 4/11 | 36% |
+| `maternity_waiting_months` | 1/11 | 9% |
+| `sub_limits_waiting_notes` | 1/11 | 9% |
+| `inpatient_hospitalization` | 10/11 | 91% |
+| `pre_hospitalization_days` | 7/11 | 64% |
+| `post_hospitalization_days` | 7/11 | 64% |
+| `day_care_treatments` | 9/11 | 82% |
+| `domiciliary_treatment` | 8/11 | 73% |
+| `ayush_coverage` | 8/11 | 73% |
+| `maternity_coverage` | 3/11 | 27% |
+| `newborn_coverage` | 4/11 | 36% |
+| `organ_donor_expenses` | 7/11 | 64% |
+| `ambulance_cover` | 9/11 | 82% |
+| `critical_illness_cover` | 3/11 | 27% |
+| `restoration_benefit` | 7/11 | 64% |
+| `no_claim_bonus_pct` | 6/11 | 55% |
+| `no_claim_bonus_cap_pct` | 7/11 | 64% |
+| `preventive_health_checkup` | 7/11 | 64% |
+| `room_rent_capping` | 8/11 | 73% |
+| `icu_capping` | 5/11 | 45% |
+| `copayment_pct` | 4/11 | 36% |
+| `copayment_trigger_notes` | 5/11 | 45% |
+| `disease_wise_sub_limits` | 2/11 | 18% |
+| `deductible_amount_inr` | 0/11 | 0% |
+| `geographic_coverage` | 0/11 | 0% |
+| `worldwide_emergency_cover` | 3/11 | 27% |
+| `network_hospital_count` | 0/11 | 0% |
+| `cashless_treatment_supported` | 10/11 | 91% |
+| `permanent_exclusions` | 5/11 | 45% |
+| `temporary_exclusions` | 1/11 | 9% |
+| `notable_exclusions_summary` | 2/11 | 18% |
+| `claim_settlement_ratio_pct` | 0/11 | 0% |
+| `claim_process_summary` | 3/11 | 27% |
+| `tat_cashless_authorization_hours` | 2/11 | 18% |
+| `available_riders` | 4/11 | 36% |
+| `top_rider_examples` | 1/11 | 9% |
+| `rider_premium_indicative_inr` | 0/11 | 0% |
+| `source_pdf_path` | 0/11 | 0% |
+| `source_pdf_url` | 2/11 | 18% |
+| `last_updated_date` | 0/11 | 0% |
+| `extraction_confidence_pct` | 6/11 | 55% |

kb/calculations/scorecard_results.md

@@ -0,0 +1,40 @@
+# Calculations — Scorecard Results
+
+_Computed by `backend/scorecard.py` at 2026-05-12T23:12:11Z on 11 extracted policies._
+
+Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md)
+
+## All policies — overall
+
+| Policy | Insurer | Grade | Score | Data % |
+| --- | --- | --- | --- | --- |
+| [Care Supreme](../policies/care-health__care-supreme__wordings.md) | care-health | **B** | 74 | 54.2% |
+| [Activ Assure](../policies/aditya-birla__activ-assure-diamond__wordings.md) | aditya-birla | **B** | 72 | 37.5% |
+| [Elevate](../policies/icici-lombard__elevate__wordings.md) | icici-lombard | **B** | 72 | 54.2% |
+| [Group Activ Health](../policies/aditya-birla__group-activ-health__wordings.md) | aditya-birla | **B** | 70 | 25.0% |
+| [Health AdvantEdge](../policies/icici-lombard__health-advantedge__wordings.md) | icici-lombard | **B** | 70 | 41.7% |
+| [Tax Gain](../policies/bajaj-allianz__tax-gain__cis.md) | bajaj-allianz | **C** | 68 | 25.0% |
+| [Care Classic](../policies/care-health__care-classic__wordings.md) | care-health | **C** | 68 | 62.5% |
+| [Silver Health](../policies/bajaj-allianz__silver-health__cis.md) | bajaj-allianz | **C** | 67 | 50.0% |
+| [Care Advantage](../policies/care-health__care-advantage-add-ons-protect-plus-care-shield__brochure.md) | care-health | **C** | 66 | 54.2% |
+| [Comprehensive Care Plan](../policies/bajaj-allianz__comprehensive-care-plan__wordings.md) | bajaj-allianz | **C** | 63 | 12.5% |
+| [Care for Senior Citizens](../policies/care-health__care-senior__brochure.md) | care-health | **C** | 62 | 54.2% |
+
+## Per-sub-score averages
+
+| Sub-score | Mean | Min | Max |
+| --- | --- | --- | --- |
+| Coverage Breadth | 70.8 | 56 | 87 |
+| Cost Predictability | 67.5 | 49 | 81 |
+| Waiting-Period Friction | 69.2 | 60 | 80 |
+| Claim Experience | 74.4 | 60 | 79 |
+| Renewal Protection | 60.0 | 60 | 60 |
+| Bonus & Loyalty | 56.5 | 50 | 73 |
+
+## Grade distribution
+
+- **A:** 0
+- **B:** 5
+- **C:** 6
+- **D:** 0
+- **F:** 0

kb/policies/aditya-birla__activ-assure-diamond__wordings.md

@@ -0,0 +1,174 @@
+# Activ Assure
+
+_Policy KB sheet — auto-generated from `rag/extracted/aditya-birla__activ-assure-diamond__wordings.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Aditya Birla Health Insurance Co. Limited](https://www.adityabirlacapital.com/healthinsurance) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `aditya-birla` | derived from `data/corpus_urls.md` |
+| Policy | **Activ Assure** | extracted from policy wordings |
+| Policy id | `aditya-birla__activ-assure-diamond__wordings` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | None% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: B** (72/100)
+> Good policy with a few notable gaps.
+
+**Data completeness:** 37.5% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `██████████████······` | **72/100** · Standard coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered<br/>&nbsp;&nbsp;&nbsp;organ donor expenses<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;free health checkups |  |
+| **Cost Predictability** | `███████████████·····` | **75/100** · Predictable costs |
+| **Waiting-Period Friction** | `████████████████····` | **80/100** · Quick activation |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 24mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `███████████·········` | **58/100** · Standard sweeteners |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;free preventive checkup |  |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _4/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `aditya-birla__activ-assure-diamond__wordings` | [I] |
+| `insurer_slug` | `aditya-birla` | [I] |
+| `insurer_name` | `Aditya Birla Health Insurance Co. Limited` | [I] |
+| `policy_name` | `Activ Assure` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | _null (not in document)_ | [E?] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | _null (not in document)_ | [E?] |
+
+### Waiting periods  _1/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | _null (not in document)_ | [E?] |
+| `pre_existing_disease_waiting_months` | `24` | [E] |
+| `specific_disease_waiting_months` | _null (not in document)_ | [E?] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | _null (not in document)_ | [E?] |
+
+### Coverage scope  _6/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | _null (not in document)_ | [E?] |
+| `post_hospitalization_days` | _null (not in document)_ | [E?] |
+| `domiciliary_treatment` | Yes, "up to the limits as specified in the Policy Schedule / Product Benefit Table of this Policy", (Must continue for at least 3 consecutive days; certain conditions excluded.) | [E] |
+| `ayush_coverage` | Yes, "up to the limits as specified in the Policy Schedule / Product Benefit Table of this Policy", (Treatment must be in recognized AYUSH hospitals; pre and post-hospitalization expenses not covered.) | [E] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | _null (not in document)_ | [E?] |
+| `organ_donor_expenses` | Yes, "up to the limits as specified in the Policy Schedule / Product Benefit Table of this Policy", (Only covers harvesting expenses; excludes pre/post-hospitalization, screening, and other donor-related expenses.) | [E] |
+| `ambulance_cover` | Yes, "up to the limits as specified in the Policy Schedule / Product Benefit Table of this Policy", (Covers transportation to nearest Hospital; excludes transportation from Hospital to residence.) | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | Yes, "Reload of Sum Insured up to the limits as specified in the Policy Schedule / Product Benefit Table of this Policy", (Available once per Policy Year; unlimited reload option available as an optional cover.) | [E] |
+| `no_claim_bonus_pct` | _null (not in document)_ | [E?] |
+| `preventive_health_checkup` | Yes, "once in a Policy Year", (Tests vary based on Sum Insured and age of the insured person.) | [E] |
+
+### Sub-limits & caps  _1/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `Single Private A/C Room (upgradable to next level, only if Single Private A/C Room is not available)` | [E] |
+| `icu_capping` | _null (not in document)_ | [E?] |
+| `copayment_pct` | _null (not in document)_ | [E?] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _2/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | Yes, "Emergency medical assistance outside India when travelling 150 km or more away from residential address for less than 90 days", (Excludes travel for medical treatment, injuries from war, unlawful acts, etc.) | [E] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | `Asthma, bronchitis, tonsillitis and upper respiratory tract infection including laryngitis and pharyngitis, cough and cold, influenza`, `Arthritis, gout and rheumatism`, `Chronic nephritis and nephritic syndrome`, `Diarrhea and all type of dysenteries, including gastroenteritis`, `Diabetes mellitus and insipidus`, `Epilepsy`, `Hypertension`, `Psychiatric or psychosomatic disorders of all kinds` | [E] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | `Reduction in PED Waiting Period`, `Unlimited Reload of Sum Insured`, `Super NCB`, `Accidental Hospitalization Booster`, `Cancer Hospitalization Booster` | [E] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | _null (not in document)_ | [E?] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/aditya-birla__activ-assure-diamond__wordings.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=aditya-birla__activ-assure-diamond__wordings`.

kb/policies/aditya-birla__group-activ-health__wordings.md

@@ -0,0 +1,173 @@
+# Group Activ Health
+
+_Policy KB sheet — auto-generated from `rag/extracted/aditya-birla__group-activ-health__wordings.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Aditya Birla Health Insurance Co. Ltd.](https://www.adityabirlacapital.com/healthinsurance) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `aditya-birla` | derived from `data/corpus_urls.md` |
+| Policy | **Group Activ Health** | extracted from policy wordings |
+| Policy id | `aditya-birla__group-activ-health__wordings` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | [https://www.adityabirlacapital.com/healthinsurance/downloads…](https://www.adityabirlacapital.com/healthinsurance/downloads) | downloaded + verified at ingest time |
+| Extraction confidence | 85.0% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: B** (70/100)
+> Good policy with a few notable gaps.
+
+**Data completeness:** 25.0% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `██████████████······` | **72/100** · Standard coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered<br/>&nbsp;&nbsp;&nbsp;maternity covered<br/>&nbsp;&nbsp;&nbsp;newborn covered |  |
+| **Cost Predictability** | `███████████████·····` | **75/100** · Predictable costs |
+| **Waiting-Period Friction** | `██████████████······` | **70/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 36mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `██████████··········` | **50/100** · Few extras |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _6/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `aditya-birla__group-activ-health__wordings` | [I] |
+| `insurer_slug` | `aditya-birla` | [I] |
+| `insurer_name` | `Aditya Birla Health Insurance Co. Ltd.` | [I] |
+| `policy_name` | `Group Activ Health` | [I] |
+| `policy_type` | `group` | [E] |
+| `uin_code` | `ADIHLGP26041V052526` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _2/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | `monthly`, `annual` | [E] |
+| `grace_period_days` | `15` | [E] |
+
+### Waiting periods  _2/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | _null (not in document)_ | [E?] |
+| `pre_existing_disease_waiting_months` | `36` | [E] |
+| `specific_disease_waiting_months` | `36` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | _null (not in document)_ | [E?] |
+
+### Coverage scope  _5/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | _null (not in document)_ | [E?] |
+| `post_hospitalization_days` | _null (not in document)_ | [E?] |
+| `domiciliary_treatment` | Yes, "Medical treatment for an illness/disease/injury which in the normal course would require care and treatment at a Hospital but is actually taken while confined at home" | [E] |
+| `ayush_coverage` | Yes, "Medical Expenses for medically required AYUSH Treatments undergone as an In-patient Treatment or Day Care Treatment", (Comfort treatment involving steam bath/sauna/oil massages are excluded.) | [E] |
+| `maternity_coverage` | Yes, "Medical treatment expenses traceable to childbirth (including complicated deliveries and caesarean sections incurred during hospitalization); expenses towards lawful medical termination of pregnancy during the policy period." | [E] |
+| `newborn_coverage` | Yes, "Baby born during the Policy Period and is Aged upto 90 days" | [E] |
+| `organ_donor_expenses` | _null (not in document)_ | [E?] |
+| `ambulance_cover` | _null (not in document)_ | [E?] |
+| `critical_illness_cover` | Yes, "Cover for specified critical illnesses like Cancer, Myocardial Infarction, Stroke, etc.", (Detailed definitions of each critical illness are provided in the policy document.) | [E] |
+| `restoration_benefit` | _null (not in document)_ | [E?] |
+| `no_claim_bonus_pct` | _null (not in document)_ | [E?] |
+| `preventive_health_checkup` | _null (not in document)_ | [E?] |
+
+### Sub-limits & caps  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | _null (not in document)_ | [E?] |
+| `icu_capping` | _null (not in document)_ | [E?] |
+| `copayment_pct` | _null (not in document)_ | [E?] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _2/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | `Cosmetic surgery`, `Self-inflicted injury`, `War`, `Non-medical expenses`, `Experimental treatments` | [E] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | `Exclusions include cosmetic surgery, self-inflicted injury, war, non-medical expenses, and experimental treatments. Pre-existing diseases are covered after 36 months.` | [E] |
+
+### Claim & service  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | `Claims must be made in accordance with the procedure set out in the policy document. Cashless facility is available at Network Providers.` | [E] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | _null (not in document)_ | [E?] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _2/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | `https://www.adityabirlacapital.com/healthinsurance/downloads` | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | `85.0` | [E] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — https://www.adityabirlacapital.com/healthinsurance/downloads…
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/aditya-birla__group-activ-health__wordings.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=aditya-birla__group-activ-health__wordings`.

kb/policies/bajaj-allianz__comprehensive-care-plan__wordings.md

@@ -0,0 +1,172 @@
+# Comprehensive Care Plan
+
+_Policy KB sheet — auto-generated from `rag/extracted/bajaj-allianz__comprehensive-care-plan__wordings.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Bajaj Allianz General Insurance Co. Ltd.](https://www.bajajallianz.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `bajaj-allianz` | derived from `data/corpus_urls.md` |
+| Policy | **Comprehensive Care Plan** | extracted from policy wordings |
+| Policy id | `bajaj-allianz__comprehensive-care-plan__wordings` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | 85.0% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: C** (63/100)
+> Decent baseline; check the trade-offs before signing.
+
+**Data completeness:** 12.5% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `███████████·········` | **58/100** · Standard coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered |  |
+| **Cost Predictability** | `███████████████·····` | **75/100** · Predictable costs |
+| **Waiting-Period Friction** | `█████████████·······` | **65/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 36mo PED waiting<br/>&nbsp;&nbsp;&nbsp;− 90d initial waiting |  |
+| **Claim Experience** | `████████████········` | **60/100** · Standard claim experience |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `██████████··········` | **50/100** · Few extras |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `bajaj-allianz__comprehensive-care-plan__wordings` | [I] |
+| `insurer_slug` | `bajaj-allianz` | [I] |
+| `insurer_name` | `Bajaj Allianz General Insurance Co. Ltd.` | [I] |
+| `policy_name` | `Comprehensive Care Plan` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `BAJHLIP15002V011415` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | _null (not in document)_ | [E?] |
+
+### Waiting periods  _2/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | `90` | [E] |
+| `pre_existing_disease_waiting_months` | `36` | [E] |
+| `specific_disease_waiting_months` | _null (not in document)_ | [E?] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | _null (not in document)_ | [E?] |
+
+### Coverage scope  _2/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | _null (not in document)_ | [E?] |
+| `post_hospitalization_days` | _null (not in document)_ | [E?] |
+| `domiciliary_treatment` | _null (not in document)_ | [E?] |
+| `ayush_coverage` | Yes, "AYUSH Hospital must have at least 5 in-patient beds, qualified AYUSH Medical Practitioner in charge round the clock, dedicated therapy sections, and maintain daily patient records.", (Cover includes medical expenses incurred on hospitalisation under Ayurveda, Yoga and Naturopathy Unani, Siddha and Homeopathy systems.) | [E] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | _null (not in document)_ | [E?] |
+| `organ_donor_expenses` | _null (not in document)_ | [E?] |
+| `ambulance_cover` | _null (not in document)_ | [E?] |
+| `critical_illness_cover` | Yes, "Covers 17 critical illnesses including Cancer of Specified Severity, Kidney Failure Requiring Regular Dialysis, Multiple Sclerosis With Persisting Symptoms, Benign Brain Tumor, Parkinson’s Disease, Alzheimer’s Disease, End Stage Liver Disease, Primary Pulmonary Arterial Hypertension, Major Organ/Bone Marrow Transplant, Open Heart Replacement or Repair of Heart Valves, Open Chest CABG, Surgery of Aorta, Stroke Resulting in Permanent Symptoms, Permanent Paralysis of Limbs, First Heart Attack of Specified Severity, Major Burns, Coma of Specified Severity.", (Cover terminates after a claim is admitted and paid up to the full Sum Insured.) | [E] |
+| `restoration_benefit` | _null (not in document)_ | [E?] |
+| `no_claim_bonus_pct` | _null (not in document)_ | [E?] |
+| `preventive_health_checkup` | _null (not in document)_ | [E?] |
+
+### Sub-limits & caps  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | _null (not in document)_ | [E?] |
+| `icu_capping` | _null (not in document)_ | [E?] |
+| `copayment_pct` | _null (not in document)_ | [E?] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _0/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | _null (not in document)_ | [E?] |
+
+### Exclusions  _2/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | `Acts of Terrorism`, `War, war-like operations, act of foreign enemy, invasion of Indian territory or any part thereof, hostilities (whether war be declared or not), civil war, rebellion, revolution, insurrection, civil commotion, military or usurped power, or loot or pillage in connection with the foregoing, seizure, capture, confiscation, arrests, restraints and detainment by order of any governments or any other authority`, `Directly or indirectly caused by or contributed to by or arising from ionizing radiation or contamination by radioactivity from any nuclear fuel or from any nuclear waste or from the combustion of nuclear fuel`, `Directly or indirectly caused by or contributed to by or arising from nuclear weapon materials`, `Arising or resulting from the Insured committing any breach of the law with criminal intent`, `Any loss or damage resulting from deliberate or intentional acts of the insured`, `Directly or indirectly caused by or contributed to by or arising out of usage, consumption or abuse of alcohol and/or drugs`, `Arising out of or as a result of any act of self-destruction or self inflicted injury, attempted suicide or suicide` | [E] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | `The policy excludes coverage for acts of terrorism, war, nuclear events, criminal acts, intentional self-harm, alcohol/drug abuse, sexually transmitted diseases, pregnancy-related treatments, and military service during war. Additionally, pre-existing diseases are excluded for the first 36 months.` | [E] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | _null (not in document)_ | [E?] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _1/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | `85.0` | [E] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/bajaj-allianz__comprehensive-care-plan__wordings.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=bajaj-allianz__comprehensive-care-plan__wordings`.

kb/policies/bajaj-allianz__silver-health__cis.md

@@ -0,0 +1,175 @@
+# Silver Health
+
+_Policy KB sheet — auto-generated from `rag/extracted/bajaj-allianz__silver-health__cis.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Bajaj Allianz General Insurance Co. Ltd.](https://www.bajajallianz.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `bajaj-allianz` | derived from `data/corpus_urls.md` |
+| Policy | **Silver Health** | extracted from policy wordings |
+| Policy id | `bajaj-allianz__silver-health__cis` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | [https://www.bajajallianz.com/health-insurance-plans/health-insurance-documents.h…](https://www.bajajallianz.com/health-insurance-plans/health-insurance-documents.html) | downloaded + verified at ingest time |
+| Extraction confidence | 95.0% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: C** (67/100)
+> Decent baseline; check the trade-offs before signing.
+
+**Data completeness:** 50.0% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `████████████········` | **60/100** · Standard coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;free health checkups |  |
+| **Cost Predictability** | `███████████·········` | **57/100** · Some out-of-pocket |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 10% copayment<br/>&nbsp;&nbsp;&nbsp;− room rent capped: 1% of hospitalization Sum Insured up to maximum Rs |  |
+| **Waiting-Period Friction** | `████████████████····` | **80/100** · Quick activation |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 24mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **79/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported<br/>&nbsp;&nbsp;&nbsp;2h cashless TAT |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `███████████·········` | **58/100** · Standard sweeteners |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;free preventive checkup |  |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `bajaj-allianz__silver-health__cis` | [I] |
+| `insurer_slug` | `bajaj-allianz` | [I] |
+| `insurer_name` | `Bajaj Allianz General Insurance Co. Ltd.` | [I] |
+| `policy_name` | `Silver Health` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `BAJHLIP23213V052223` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | _null (not in document)_ | [E?] |
+
+### Waiting periods  _4/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | `30` | [E] |
+| `pre_existing_disease_waiting_months` | `24` | [E] |
+| `specific_disease_waiting_months` | `12` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | `Surgery for gastric or duodenal ulcers`, `Benign prostatic hypertrophy`, `Hydrocele`, `Haemorrhoids`, `Dysfunctional uterine bleeding`, `Endometriosis`, `Stones in the urinary and biliary systems`, `Prolapse of genitourinary/intra abdominal organs` | [E] |
+
+### Coverage scope  _6/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | `30` | [E] |
+| `post_hospitalization_days` | `60` | [E] |
+| `domiciliary_treatment` | Yes, "Coverage for medical treatment for a period exceeding three days, for an illness/disease/injury, which in the normal course, would require care and treatment at a Hospital but, on the advice of the attending Medical Practitioner, is taken whilst confined at home", (Applicable only for plan B) | [E] |
+| `ayush_coverage` | _null (not in document)_ | [E?] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | _null (not in document)_ | [E?] |
+| `organ_donor_expenses` | _null (not in document)_ | [E?] |
+| `ambulance_cover` | Yes, limit ₹1,000, "Road Ambulance - max. up to ₹ 1,000/- per claim" | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | _null (not in document)_ | [E?] |
+| `no_claim_bonus_pct` | `10.0` | [E] |
+| `preventive_health_checkup` | Yes, limit ₹5,000, "Plan A - After every 4 Claim Free Year, Plan B - After every 2 Year- 1% or max 5000 Whichever is lower" | [E] |
+
+### Sub-limits & caps  _3/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `1% of hospitalization Sum Insured up to maximum Rs. 7,500 per day` | [E] |
+| `icu_capping` | _null (not in document)_ | [E?] |
+| `copayment_pct` | `10.0` | [E] |
+| `disease_wise_sub_limits` | `{'cataract': '10% of Sum Insured, Max up to 40,000 per claim (whichever is lower)', 'domicilliary': 'Covered up to 10% of Sum Insured'}` | [E] |
+
+### Geography & network  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | `Any hospital admission primarily for investigation diagnostic purpose`, `Expenses related to any admission primarily for enforced bed rest and not for receiving treatment.`, `Obesity/Weight Control`, `Change-of-gender treatments`, `Expenses for cosmetic or plastic surgery or any treatment to change appearance unless for reconstruction following an Accident, Burn(s) etc.`, `Expenses for treatment arising from insured committing or attempting to commit a breach of law with criminal intent.`, `Treatment for Alcoholism, drug or substance abuse.`, `Treatments received in heath hydros, nature cure clinics, etc. where admission is arranged wholly or partly for domestic reasons.` | [E] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _2/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | `Cashless Claim process is available at Network Hospitals. Must intimate Us 48 hours before the planned Hospitalization and within 24 hours of emergency hospitalization and request pre-authorization. Reimbursement claim process applicable for claims where treatment is taken at a Non network hospital OR if cashless claim is denied. Must intimate Us 48 hours before the planned Hospitalization and within 48 hours of emergency hospitalization. Documentation must be submitted within 30 days of discharge. The Company shall settle or reject the claim within 45days from the date of receipt of last necessary document.` | [E] |
+| `tat_cashless_authorization_hours` | `2.0` | [E] |
+
+### Riders / optional  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | `Room Rent Capping` | [E] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _2/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | `https://www.bajajallianz.com/health-insurance-plans/health-insurance-documents.html` | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | `95.0` | [E] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — https://www.bajajallianz.com/health-insurance-plans/health-i…
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/bajaj-allianz__silver-health__cis.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=bajaj-allianz__silver-health__cis`.

kb/policies/bajaj-allianz__tax-gain__cis.md

@@ -0,0 +1,174 @@
+# Tax Gain
+
+_Policy KB sheet — auto-generated from `rag/extracted/bajaj-allianz__tax-gain__cis.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Bajaj Allianz General Insurance Co. Ltd.](https://www.bajajallianz.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `bajaj-allianz` | derived from `data/corpus_urls.md` |
+| Policy | **Tax Gain** | extracted from policy wordings |
+| Policy id | `bajaj-allianz__tax-gain__cis` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | None% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: C** (68/100)
+> Decent baseline; check the trade-offs before signing.
+
+**Data completeness:** 25.0% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `███████████·········` | **56/100** · Standard coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;free health checkups |  |
+| **Cost Predictability** | `███████████████·····` | **75/100** · Predictable costs |
+| **Waiting-Period Friction** | `██████████████······` | **70/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 36mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **79/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported<br/>&nbsp;&nbsp;&nbsp;2h cashless TAT |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `███████████·········` | **58/100** · Standard sweeteners |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;free preventive checkup |  |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `bajaj-allianz__tax-gain__cis` | [I] |
+| `insurer_slug` | `bajaj-allianz` | [I] |
+| `insurer_name` | `Bajaj Allianz General Insurance Co. Ltd.` | [I] |
+| `policy_name` | `Tax Gain` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `BAJHLIP21184V022021` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | _null (not in document)_ | [E?] |
+
+### Waiting periods  _4/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | `30` | [E] |
+| `pre_existing_disease_waiting_months` | `36` | [E] |
+| `specific_disease_waiting_months` | `24` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | `gastric or duodenal ulcers`, `benign prostatic hypertrophy`, `all types of sinuses`, `hemorrhoids`, `dysfunctional uterine bleeding`, `endometriosis`, `stones in the urinary and biliary systems`, `surgery on ears/tonsils/adenoids/paranasal sinuses` | [E] |
+
+### Coverage scope  _2/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | _null (not in document)_ | [E?] |
+| `post_hospitalization_days` | _null (not in document)_ | [E?] |
+| `domiciliary_treatment` | _null (not in document)_ | [E?] |
+| `ayush_coverage` | _null (not in document)_ | [E?] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | _null (not in document)_ | [E?] |
+| `organ_donor_expenses` | _null (not in document)_ | [E?] |
+| `ambulance_cover` | Yes, limit ₹1,000, "Max up to Rs 1000 per valid hospitalization claim." | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | _null (not in document)_ | [E?] |
+| `no_claim_bonus_pct` | _null (not in document)_ | [E?] |
+| `preventive_health_checkup` | Yes, "Preventive Health check up at the end of every 4 continuous policy years as per limits specified in policy wordings." | [E] |
+
+### Sub-limits & caps  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | _null (not in document)_ | [E?] |
+| `icu_capping` | _null (not in document)_ | [E?] |
+| `copayment_pct` | _null (not in document)_ | [E?] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | `Any hospital admission primarily for investigation diagnostic purpose`, `Expenses related to any admission primarily for enforced bed rest and not for receiving treatment`, `Obesity/Weight Control`, `Change-of-gender treatments`, `Expenses for cosmetic or plastic surgery or any treatment to change appearance unless for reconstruction following an Accident, Burn(s) etc.`, `Expenses for treatment arising from Insured committing or attempting to commit a breach of law with criminal intent`, `Treatment for Alcoholism, drug or substance abuse`, `Treatments received in health hydros, nature cure clinics, etc. where admission is arranged wholly or partly for domestic reasons` | [E] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _2/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | `Cashless treatment is only available at Network Hospitals. You or Your representative must intimate Us 48 hours before the planned Hospitalization and within 24 hours of emergency hospitalization and request pre-authorization by way of the written form. We will review each claim for Medical Expenses, coverage and accordingly issue an authorisation letter either to You or the Network Hospital. For reimbursement claims where treatment is taken at a Non network hospital or if cashless claim is denied, you must intimate Us 48 hours before the planned Hospitalization and within 48 hours of emergency hospitalization. You must submit all necessary documents within 30 days of discharge from a Hospital. The Company shall settle or reject the claim within 45 days from the date of receipt of last necessary document.` | [E] |
+| `tat_cashless_authorization_hours` | `2.0` | [E] |
+
+### Riders / optional  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | _null (not in document)_ | [E?] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | _null (not in document)_ | [E?] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/bajaj-allianz__tax-gain__cis.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=bajaj-allianz__tax-gain__cis`.

kb/policies/care-health__care-advantage-add-ons-protect-plus-care-shield__brochure.md

@@ -0,0 +1,175 @@
+# Care Advantage
+
+_Policy KB sheet — auto-generated from `rag/extracted/care-health__care-advantage-add-ons-protect-plus-care-shield__brochure.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Care Health Insurance Limited](https://www.careinsurance.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `care-health` | derived from `data/corpus_urls.md` |
+| Policy | **Care Advantage** | extracted from policy wordings |
+| Policy id | `care-health__care-advantage-add-ons-protect-plus-care-shield__brochure` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | 85.0% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: C** (66/100)
+> Decent baseline; check the trade-offs before signing.
+
+**Data completeness:** 54.2% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `█████████████·······` | **68/100** · Standard coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered<br/>&nbsp;&nbsp;&nbsp;organ donor expenses<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;free health checkups |  |
+| **Cost Predictability** | `███████████·········` | **57/100** · Some out-of-pocket |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 20% copayment |  |
+| **Waiting-Period Friction** | `██████████████······` | **70/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 36mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `███████████·········` | **58/100** · Standard sweeteners |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;free preventive checkup |  |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `care-health__care-advantage-add-ons-protect-plus-care-shield__brochure` | [I] |
+| `insurer_slug` | `care-health` | [I] |
+| `insurer_name` | `Care Health Insurance Limited` | [I] |
+| `policy_name` | `Care Advantage` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `CHIHLIP26049V042526` | [E] |
+
+### Eligibility  _1/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | `Indian resident only` | [E] |
+
+### Sum insured & premium  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | _null (not in document)_ | [E?] |
+
+### Waiting periods  _3/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | `30` | [E] |
+| `pre_existing_disease_waiting_months` | `36` | [E] |
+| `specific_disease_waiting_months` | `24` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | _null (not in document)_ | [E?] |
+
+### Coverage scope  _8/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | `30` | [E] |
+| `post_hospitalization_days` | `60` | [E] |
+| `domiciliary_treatment` | _null (not in document)_ | [E?] |
+| `ayush_coverage` | Yes, "Up to SI" | [E] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | _null (not in document)_ | [E?] |
+| `organ_donor_expenses` | Yes, "Up to SI" | [E] |
+| `ambulance_cover` | Yes, "Up to SI" | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | Yes, "Up to SI and available for unrelated or same illness" | [E] |
+| `no_claim_bonus_pct` | `10.0` | [E] |
+| `preventive_health_checkup` | Yes, "Annual" | [E] |
+
+### Sub-limits & caps  _3/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `No sub-limit` | [E] |
+| `icu_capping` | `No sub-limit` | [E] |
+| `copayment_pct` | `20.0` | [E] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _2/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | Yes, "Up to SI", (With Protect Plus add-on policy against payment of additional premium) | [E] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _0/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | _null (not in document)_ | [E?] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _2/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | `Protect Plus`, `Care OPD`, `Care Advanced` | [E] |
+| `top_rider_examples` | `Protect Plus`, `Care OPD` | [E] |
+
+### Source metadata  _1/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | `85.0` | [E] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/care-health__care-advantage-add-ons-protect-plus-care-shield__brochure.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=care-health__care-advantage-add-ons-protect-plus-care-shield__brochure`.

kb/policies/care-health__care-classic__wordings.md

@@ -0,0 +1,174 @@
+# Care Classic
+
+_Policy KB sheet — auto-generated from `rag/extracted/care-health__care-classic__wordings.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Care Health Insurance Limited](https://www.careinsurance.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `care-health` | derived from `data/corpus_urls.md` |
+| Policy | **Care Classic** | extracted from policy wordings |
+| Policy id | `care-health__care-classic__wordings` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | None% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: C** (68/100)
+> Decent baseline; check the trade-offs before signing.
+
+**Data completeness:** 62.5% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `█████████████████···` | **87/100** · Wide coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered<br/>&nbsp;&nbsp;&nbsp;maternity covered<br/>&nbsp;&nbsp;&nbsp;newborn covered<br/>&nbsp;&nbsp;&nbsp;organ donor expenses<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;60d pre-hospitalization<br/>&nbsp;&nbsp;&nbsp;90d post-hospitalization |  |
+| **Cost Predictability** | `█████████···········` | **49/100** · Material out-of-pocket |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 20% copayment<br/>&nbsp;&nbsp;&nbsp;− room rent capped: 1% of SI per day or Single Private AC Room |  |
+| **Waiting-Period Friction** | `█████████████·······` | **66/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 36mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `███████████·········` | **58/100** · Standard sweeteners |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `care-health__care-classic__wordings` | [I] |
+| `insurer_slug` | `care-health` | [I] |
+| `insurer_name` | `Care Health Insurance Limited` | [I] |
+| `policy_name` | `Care Classic` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `CHIHLIP22071V012122` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | `30` | [E] |
+
+### Waiting periods  _2/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | _null (not in document)_ | [E?] |
+| `pre_existing_disease_waiting_months` | `36` | [E] |
+| `specific_disease_waiting_months` | _null (not in document)_ | [E?] |
+| `maternity_waiting_months` | `36` | [E] |
+| `specific_diseases_listed` | _null (not in document)_ | [E?] |
+
+### Coverage scope  _10/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | `60` | [E] |
+| `post_hospitalization_days` | `90` | [E] |
+| `domiciliary_treatment` | Yes, "Amount specified against this Benefit in the Policy Schedule", (Domiciliary Hospitalization must continue for a period exceeding 3 consecutive days.) | [E] |
+| `ayush_coverage` | Yes, "Sum Insured as specified in the Policy Schedule", (Treatment must be rendered from a registered Medical Practitioner.) | [E] |
+| `maternity_coverage` | Yes, "Amount specified against this Benefit in the Policy Schedule", (Waiting period of 36 months from the date of first inception of the policy.) | [E] |
+| `newborn_coverage` | Yes, "Baby born during the Policy Period and is aged up to 90 days" | [E] |
+| `organ_donor_expenses` | Yes, "Limit specified against this Benefit in the Policy Schedule", (Organ donor must be an eligible donor in accordance with The Transplantation of Human Organs Act, 1994.) | [E] |
+| `ambulance_cover` | Yes, "Amount specified against this Benefit in the Policy Schedule", (Ambulance transportation must be offered by a Hospital or by an Ambulance service provider.) | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | Yes, "Unlimited Automatic Recharge", (Recharge shall be utilized only after the base Sum Insured, 'No Claims Bonus' and 'Additional Sum Insured for Accidental Hospitalization' has been completely exhausted.) | [E] |
+| `no_claim_bonus_pct` | `25.0` | [E] |
+| `preventive_health_checkup` | _null (not in document)_ | [E?] |
+
+### Sub-limits & caps  _3/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `1% of SI per day or Single Private AC Room` | [E] |
+| `icu_capping` | `2% of SI per day or no limit` | [E] |
+| `copayment_pct` | `20.0` | [E] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _0/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | _null (not in document)_ | [E?] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | _null (not in document)_ | [E?] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | _null (not in document)_ | [E?] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/care-health__care-classic__wordings.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=care-health__care-classic__wordings`.

kb/policies/care-health__care-senior__brochure.md

@@ -0,0 +1,175 @@
+# Care for Senior Citizens
+
+_Policy KB sheet — auto-generated from `rag/extracted/care-health__care-senior__brochure.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Care Health Insurance Limited](https://www.careinsurance.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `care-health` | derived from `data/corpus_urls.md` |
+| Policy | **Care for Senior Citizens** | extracted from policy wordings |
+| Policy id | `care-health__care-senior__brochure` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | 90.0% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: C** (62/100)
+> Decent baseline; check the trade-offs before signing.
+
+**Data completeness:** 54.2% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `████████████········` | **64/100** · Standard coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;organ donor expenses<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;free health checkups |  |
+| **Cost Predictability** | `█████████···········` | **49/100** · Material out-of-pocket |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 20% copayment<br/>&nbsp;&nbsp;&nbsp;− room rent capped: 1% SI per day (Max. up to 1% of SI per day) for 3  |  |
+| **Waiting-Period Friction** | `████████████········` | **60/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 48mo PED waiting (long) |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `███████████·········` | **58/100** · Standard sweeteners |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;free preventive checkup |  |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `care-health__care-senior__brochure` | [I] |
+| `insurer_slug` | `care-health` | [I] |
+| `insurer_name` | `Care Health Insurance Limited` | [I] |
+| `policy_name` | `Care for Senior Citizens` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `RHIHLIP21017V052021` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | _null (not in document)_ | [E?] |
+
+### Waiting periods  _4/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | `30` | [E] |
+| `pre_existing_disease_waiting_months` | `48` | [E] |
+| `specific_disease_waiting_months` | `24` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | `Cataract`, `Total Knee Replacement`, `Hernia`, `Hysterectomy`, `Benign Prostate Hypertrophy (BPH)`, `Stones of renal system`, `Cerebrovascular and Cardiovasular disorders`, `Cancer` | [E] |
+
+### Coverage scope  _8/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | `30` | [E] |
+| `post_hospitalization_days` | `60` | [E] |
+| `domiciliary_treatment` | Yes, "Up to 10% of SI, covered after 3 days" | [E] |
+| `ayush_coverage` | _null (not in document)_ | [E?] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | _null (not in document)_ | [E?] |
+| `organ_donor_expenses` | Yes, limit ₹50,000, "Up to ₹50,000 for 3 Lacs plan, Up to ₹1,00,000 for 5,7,10 Lacs plan" | [E] |
+| `ambulance_cover` | Yes, limit ₹1,500, "Up to ₹1,500 per hospitalization for 3 Lacs plan, Up to ₹2,000 per hospitalization for 5,7,10 Lacs plan" | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | Yes, "Yes to SI (Once in a Policy Year)" | [E] |
+| `no_claim_bonus_pct` | `10.0` | [E] |
+| `preventive_health_checkup` | Yes, "Yes, all members" | [E] |
+
+### Sub-limits & caps  _4/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `1% SI per day (Max. up to 1% of SI per day) for 3 Lacs plan, Single Private AC Room (Max. up to 1% of SI per day) for 5,7,10 Lacs plan` | [E] |
+| `icu_capping` | `2% SI per day` | [E] |
+| `copayment_pct` | `20.0` | [E] |
+| `disease_wise_sub_limits` | `{'Cataract': '₹20,000 per eye for 3 Lacs plan, ₹30,000 per eye for 5,7,10 Lacs plan', 'Total Knee Replacement': '₹80,000 per knee for 3 Lacs plan, ₹1,00,000 per knee for 5,7,10 Lacs plan', 'Hernia': '₹50,000 for 3 Lacs plan, ₹65,000 for 5,7,10 Lacs plan', 'Hysterectomy': '₹50,000 for 3 Lacs plan, ₹65,000 for 5,7,10 Lacs plan', 'Benign Prostate Hypertrophy (BPH)': '₹50,000 for 3 Lacs plan, ₹65,000 for 5,7,10 Lacs plan', 'Stones of renal system': '₹50,000 for 3 Lacs plan, ₹65,000 for 5,7,10 Lacs plan', 'Cerebrovascular and Cardiovasular disorders': '₹2,00,000 for 3 Lacs plan, ₹2,50,000 for 5,7,10 Lacs plan', 'Cancer': '₹2,00,000 for 3 Lacs plan, ₹2,50,000 for 5,7,10 Lacs plan', 'Other renal complications and Disorders': '₹2,00,000 for 3 Lacs plan, ₹2,50,000 for 5,7,10 Lacs plan', 'Breakage of bones': '₹2,00,000 for 3 Lacs plan, ₹2,50,000 for 5,7,10 Lacs plan'}` | [E] |
+
+### Geography & network  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _0/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | _null (not in document)_ | [E?] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | _null (not in document)_ | [E?] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _1/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | `90.0` | [E] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/care-health__care-senior__brochure.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=care-health__care-senior__brochure`.

kb/policies/care-health__care-supreme__wordings.md

@@ -0,0 +1,174 @@
+# Care Supreme
+
+_Policy KB sheet — auto-generated from `rag/extracted/care-health__care-supreme__wordings.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [Care Health Insurance Limited](https://www.careinsurance.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `care-health` | derived from `data/corpus_urls.md` |
+| Policy | **Care Supreme** | extracted from policy wordings |
+| Policy id | `care-health__care-supreme__wordings` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | None% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: B** (74/100)
+> Good policy with a few notable gaps.
+
+**Data completeness:** 54.2% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `████████████████····` | **84/100** · Wide coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered<br/>&nbsp;&nbsp;&nbsp;newborn covered<br/>&nbsp;&nbsp;&nbsp;organ donor expenses<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;free health checkups<br/>&nbsp;&nbsp;&nbsp;60d pre-hospitalization<br/>&nbsp;&nbsp;&nbsp;180d post-hospitalization |  |
+| **Cost Predictability** | `███████████████·····` | **75/100** · Predictable costs |
+| **Waiting-Period Friction** | `██████████████······` | **70/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 36mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `██████████████······` | **73/100** · Standard sweeteners |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;50% NCB<br/>&nbsp;&nbsp;&nbsp;free preventive checkup |  |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `care-health__care-supreme__wordings` | [I] |
+| `insurer_slug` | `care-health` | [I] |
+| `insurer_name` | `Care Health Insurance Limited` | [I] |
+| `policy_name` | `Care Supreme` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `CHIHLIP23128V012223` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | `30` | [E] |
+
+### Waiting periods  _2/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | _null (not in document)_ | [E?] |
+| `pre_existing_disease_waiting_months` | `36` | [E] |
+| `specific_disease_waiting_months` | `24` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | _null (not in document)_ | [E?] |
+
+### Coverage scope  _10/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | `60` | [E] |
+| `post_hospitalization_days` | `180` | [E] |
+| `domiciliary_treatment` | Yes, "No limit", (Treatment must continue for a period exceeding 3 consecutive days and must be Medically Necessary.) | [E] |
+| `ayush_coverage` | Yes, "No limit", (Treatment must be from a registered AYUSH Medical Practitioner and within India.) | [E] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | Yes, "Covered from day 1", (All applicable waiting periods stand valid for this benefit.) | [E] |
+| `organ_donor_expenses` | Yes, "No limit", (Donor must be eligible in accordance with The Transplantation of Human Organs Act, 1994.) | [E] |
+| `ambulance_cover` | Yes, "No limit", (Transportation must be certified by a Medical Practitioner as Medically Necessary.) | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | Yes, "Unlimited Automatic Recharge", (Recharge is applicable only after base Sum Insured, applicable Cumulative Bonus, and Plus Benefit (if applicable) have been exhausted.) | [E] |
+| `no_claim_bonus_pct` | `50.0` | [E] |
+| `preventive_health_checkup` | Yes, "Once per Policy Year", (Available for Insured Persons aged 18 years or above.) | [E] |
+
+### Sub-limits & caps  _2/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `No limit` | [E] |
+| `icu_capping` | `No limit` | [E] |
+| `copayment_pct` | _null (not in document)_ | [E?] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _0/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | _null (not in document)_ | [E?] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | `Smart Select`, `Room Rent Modification`, `PED Wait Period Modification`, `Named Ailment Wait Period Modification`, `Instant Cover`, `Deductible`, `Co-payment`, `New Born Cover` | [E] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | _null (not in document)_ | [E?] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/care-health__care-supreme__wordings.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=care-health__care-supreme__wordings`.

kb/policies/icici-lombard__elevate__wordings.md

@@ -0,0 +1,173 @@
+# Elevate
+
+_Policy KB sheet — auto-generated from `rag/extracted/icici-lombard__elevate__wordings.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [ICICI Lombard General Insurance Company Limited](https://www.icicilombard.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `icici-lombard` | derived from `data/corpus_urls.md` |
+| Policy | **Elevate** | extracted from policy wordings |
+| Policy id | `icici-lombard__elevate__wordings` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | None% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: B** (72/100)
+> Good policy with a few notable gaps.
+
+**Data completeness:** 54.2% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `████████████████····` | **81/100** · Wide coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered<br/>&nbsp;&nbsp;&nbsp;newborn covered<br/>&nbsp;&nbsp;&nbsp;organ donor expenses<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;90d pre-hospitalization<br/>&nbsp;&nbsp;&nbsp;180d post-hospitalization |  |
+| **Cost Predictability** | `███████████████·····` | **75/100** · Predictable costs |
+| **Waiting-Period Friction** | `██████████████······` | **70/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 36mo PED waiting |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `██████████··········` | **50/100** · Few extras |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `icici-lombard__elevate__wordings` | [I] |
+| `insurer_slug` | `icici-lombard` | [I] |
+| `insurer_name` | `ICICI Lombard General Insurance Company Limited` | [I] |
+| `policy_name` | `Elevate` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `ICIHLIP25048V042425` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | `30` | [E] |
+
+### Waiting periods  _3/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | `30` | [E] |
+| `pre_existing_disease_waiting_months` | `36` | [E] |
+| `specific_disease_waiting_months` | `24` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | _null (not in document)_ | [E?] |
+
+### Coverage scope  _12/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | `90` | [E] |
+| `post_hospitalization_days` | `180` | [E] |
+| `domiciliary_treatment` | Yes, "Requires at least 3 consecutive days of treatment", (Excludes certain conditions like asthma, bronchitis, tonsillitis, etc.) | [E] |
+| `ayush_coverage` | Yes, "Hospitalization for AYUSH Treatment at a Government Recognized AYUSH Hospital or AYUSH Day Care Centre" | [E] |
+| `maternity_coverage` | _unclear: {'covered': None, 'limit_inr': None, 'limit_text': None, 'notes': None}_ | [E] |
+| `newborn_coverage` | Yes, "Newborn Baby means baby born during the Policy Period and is aged up to 90 days" | [E] |
+| `organ_donor_expenses` | Yes, "Medical expenses incurred in respect of an organ donor’s Hospitalization during the Policy Period for harvesting of the organ donated to the Insured Person", (Excludes pre-hospitalization and post-hospitalization medical expenses of the organ donor) | [E] |
+| `ambulance_cover` | Yes, "Expenses incurred on road ambulance services to transfer the Insured Person to the nearest Hospital from the place of Accident/Illness", (Excludes transportation from Hospital to the Insured Person’s residence after discharge) | [E] |
+| `critical_illness_cover` | _unclear: {'covered': None, 'limit_inr': None, 'limit_text': None, 'notes': None}_ | [E] |
+| `restoration_benefit` | Yes, "Reset up to 100% of the Annual Sum Insured, for any illness/disease/injury for the Insured Person in a Policy Year", (Not available for Policies with Unlimited Sum Insured option) | [E] |
+| `no_claim_bonus_pct` | `20.0` | [E] |
+| `preventive_health_checkup` | _unclear: {'covered': None, 'limit_inr': None, 'limit_text': None, 'notes': None}_ | [E] |
+
+### Sub-limits & caps  _1/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `Single Private AC room` | [E] |
+| `icu_capping` | _null (not in document)_ | [E?] |
+| `copayment_pct` | _null (not in document)_ | [E?] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _2/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _unclear: {'covered': None, 'limit_inr': None, 'limit_text': None, 'notes': None}_ | [E] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _0/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | _null (not in document)_ | [E?] |
+| `temporary_exclusions` | _null (not in document)_ | [E?] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | _null (not in document)_ | [E?] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _0/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | _null (not in document)_ | [E?] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/icici-lombard__elevate__wordings.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=icici-lombard__elevate__wordings`.

kb/policies/icici-lombard__health-advantedge__wordings.md

@@ -0,0 +1,174 @@
+# Health AdvantEdge
+
+_Policy KB sheet — auto-generated from `rag/extracted/icici-lombard__health-advantedge__wordings.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._
+
+## Identity
+
+| Field | Value | Source |
+| --- | --- | --- |
+| Insurer | [ICICI Lombard General Insurance Company Limited](https://www.icicilombard.com/) | curated · verified `eval/verified_urls.json` |
+| Insurer slug | `icici-lombard` | derived from `data/corpus_urls.md` |
+| Policy | **Health AdvantEdge** | extracted from policy wordings |
+| Policy id | `icici-lombard__health-advantedge__wordings` | minted by us (`<insurer-slug>__<doc-slug>`) |
+| Source PDF | […]() | downloaded + verified at ingest time |
+| Extraction confidence | 85.0% (self-rated by extractor) | computed |
+
+## Scorecard — single A-F view
+
+### **Grade: B** (70/100)
+> Good policy with a few notable gaps.
+
+**Data completeness:** 41.7% of the 24 scored fields have data.
+
+| Sub-score | Bar | Score & Signals |
+| --- | --- | --- |
+| **Coverage Breadth** | `███████████████·····` | **77/100** · Wide coverage |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;AYUSH covered<br/>&nbsp;&nbsp;&nbsp;organ donor expenses<br/>&nbsp;&nbsp;&nbsp;ambulance covered<br/>&nbsp;&nbsp;&nbsp;60d pre-hospitalization<br/>&nbsp;&nbsp;&nbsp;180d post-hospitalization |  |
+| **Cost Predictability** | `████████████████····` | **81/100** · Predictable costs |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;no room rent cap |  |
+| **Waiting-Period Friction** | `████████████········` | **60/100** · Standard waits |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;− 48mo PED waiting (long) |  |
+| **Claim Experience** | `███████████████·····` | **75/100** · Smooth claims |
+|  | _signals:_<br/>&nbsp;&nbsp;&nbsp;cashless supported |  |
+| **Renewal Protection** | `████████████········` | **60/100** · Adequate |
+| **Bonus & Loyalty** | `██████████··········` | **50/100** · Few extras |
+
+_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._
+
+## All extracted data points — by group
+
+**Derivation legend:**
+- **[E]** Extracted directly from policy PDF by LLM
+- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)
+- **[C]** Computed from extracted fields (e.g. scorecard sub-score)
+- **[I]** Implied / canonicalised by us
+- **[V]** Verified externally (HEAD-check, URL probe)
+
+### Identity  _5/6 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `policy_id` | `icici-lombard__health-advantedge__wordings` | [I] |
+| `insurer_slug` | `icici-lombard` | [I] |
+| `insurer_name` | `ICICI Lombard General Insurance Company Limited` | [I] |
+| `policy_name` | `Health AdvantEdge` | [I] |
+| `policy_type` | _null (not in document)_ | [E?] |
+| `uin_code` | `ICIHLIP24182V042324` | [E] |
+
+### Eligibility  _0/1 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `residency_requirement` | _null (not in document)_ | [E?] |
+
+### Sum insured & premium  _1/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `premium_payment_modes` | _null (not in document)_ | [E?] |
+| `grace_period_days` | `30` | [E] |
+
+### Waiting periods  _3/5 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `initial_waiting_period_days` | _null (not in document)_ | [E?] |
+| `pre_existing_disease_waiting_months` | `48` | [E] |
+| `specific_disease_waiting_months` | `24` | [E] |
+| `maternity_waiting_months` | _null (not in document)_ | [E?] |
+| `specific_diseases_listed` | `bariatric surgery` | [E] |
+
+### Coverage scope  _7/12 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `pre_hospitalization_days` | `60` | [E] |
+| `post_hospitalization_days` | `180` | [E] |
+| `domiciliary_treatment` | Yes, "Covered for at least 3 consecutive days", (Excludes certain conditions like asthma, bronchitis, tonsillitis, etc.) | [E] |
+| `ayush_coverage` | Yes, "Covered up to Annual Sum Insured", (Treatment must be at a registered AYUSH Hospital or AYUSH Day Care Centre.) | [E] |
+| `maternity_coverage` | _null (not in document)_ | [E?] |
+| `newborn_coverage` | _null (not in document)_ | [E?] |
+| `organ_donor_expenses` | Yes, "Covered up to Annual Sum Insured", (Organ donation must conform to the Transplantation of Human Organs Act 1994.) | [E] |
+| `ambulance_cover` | Yes, limit ₹10,000, "1% of Annual Sum Insured maximum up to INR 10,000", (Road ambulance services only; air ambulance covered separately up to Annual Sum Insured.) | [E] |
+| `critical_illness_cover` | _null (not in document)_ | [E?] |
+| `restoration_benefit` | Yes, "Reset up to 100% of Annual Sum Insured for policies less than Rs. 10 Lakhs (once) and unlimited for Rs. 10 Lakhs and above", (Not available for the first claim made during the Policy Year.) | [E] |
+| `no_claim_bonus_pct` | _null (not in document)_ | [E?] |
+| `preventive_health_checkup` | _null (not in document)_ | [E?] |
+
+### Sub-limits & caps  _2/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `room_rent_capping` | `1% of Annual Sum Insured for policies up to Rs. 4 Lakhs, no capping for Rs. 5 Lakhs and above` | [E] |
+| `icu_capping` | `2% of Annual Sum Insured for policies up to Rs. 4 Lakhs, no capping for Rs. 5 Lakhs and above` | [E] |
+| `copayment_pct` | _null (not in document)_ | [E?] |
+| `disease_wise_sub_limits` | _null (not in document)_ | [E?] |
+
+### Geography & network  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `worldwide_emergency_cover` | _null (not in document)_ | [E?] |
+| `network_hospital_count` | _null (not in document)_ | [E?] |
+| `cashless_treatment_supported` | Yes | [E] |
+
+### Exclusions  _1/3 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `permanent_exclusions` | _null (not in document)_ | [E?] |
+| `temporary_exclusions` | `bariatric surgery` | [E] |
+| `notable_exclusions_summary` | _null (not in document)_ | [E?] |
+
+### Claim & service  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `claim_process_summary` | _null (not in document)_ | [E?] |
+| `tat_cashless_authorization_hours` | _null (not in document)_ | [E?] |
+
+### Riders / optional  _0/2 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `available_riders` | _null (not in document)_ | [E?] |
+| `top_rider_examples` | _null (not in document)_ | [E?] |
+
+### Source metadata  _1/4 fields populated_
+
+| Field | Value | Type |
+| --- | --- | --- |
+| `source_pdf_path` | _null (not in document)_ | [V] |
+| `source_pdf_url` | _null (not in document)_ | [V] |
+| `last_updated_date` | _null (not in document)_ | [V] |
+| `extraction_confidence_pct` | `85.0` | [E] |
+
+## Lineage — end-to-end audit trail for this policy
+
+Every data point above traces through this exact pipeline:
+
+```
+1. SOURCE        — …
+                   (curated by corpus-discovery agent, verified at download)
+2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py
+                   PDF magic-byte check + size > 50 KB enforced
+3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)
+4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)
+5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)
+6. INDEX         — Chroma persistent client (rag/vectors/) with metadata
+7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema
+                   → rag/extracted/icici-lombard__health-advantedge__wordings.json (this file's source data)
+8. STORE         — DuckDB upsert into rag/policies.duckdb
+9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop
+10. KB SHEET     — rag/build_kb.py renders this markdown
+```
+
+**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.
+
+## What the bot will and won't say about this policy
+
+Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):
+- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).
+- Every factual claim cites this PDF with page numbers.
+- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.
+- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id=icici-lombard__health-advantedge__wordings`.

kb/research/corpus_acquisition.md

@@ -0,0 +1,39 @@
+# Research — Corpus Acquisition
+
+_Auto-generated from `rag/corpus/_manifest.json` at 2026-05-12T23:12:11Z_
+
+## Headline
+- Total attempted: **91** URLs across 10 target insurers
+- Successfully downloaded: **76** PDFs
+- Failed: **15**
+- Elapsed: 1008.2s
+
+## Per-insurer breakdown
+
+| Insurer | OK | Fail |
+| --- | --- | --- |
+| `aditya-birla` | 6 | 0 |
+| `bajaj-allianz` | 10 | 1 |
+| `care-health` | 9 | 0 |
+| `hdfc-ergo` | 12 | 0 |
+| `icici-lombard` | 9 | 0 |
+| `manipalcigna` | 4 | 3 |
+| `new-india` | 8 | 0 |
+| `niva-bupa` | 10 | 0 |
+| `star-health` | 0 | 11 |
+| `tata-aig` | 8 | 0 |
+
+## Failure reasons
+
+| Reason | Count |
+| --- | --- |
+| `http_403` | 8 |
+| `http_404` | 4 |
+| `req_ConnectionError` | 3 |
+
+## How we did it
+- Dispatched a research agent to find direct PDF URLs for all health policies across 10 target insurers
+- Source list saved to `data/corpus_urls.md` (75 URLs)
+- `rag/download_corpus.py` downloads with PDF magic-byte verification + size floor (50KB)
+- `rag/download_retry.py` retried failed downloads with browser-grade headers (rescued ICICI Lombard 9/9)
+- Star Health (11 PDFs) blocked by CDN bot protection — deferred to v2 (see `docs/04-failure-modes.md` + ROADMAP)

kb/research/url_verification.md

@@ -0,0 +1,28 @@
+# Research — URL Verification
+
+_Auto-generated from `eval/verified_urls.json` (verified at 2026-05-12T23:00:21Z)_
+
+## Headline
+- Insurer home URLs: **7/10** reachable via HEAD/GET
+- Policy PDF URLs (sample): **30/30** reachable
+
+## Why this matters
+Every URL that the bot or coverage panel surfaces to the user is checked here. We do NOT show URLs that we haven't verified.
+Verification script: [`tools/verify_urls.py`](../../tools/verify_urls.py).
+
+## Insurer home URLs
+
+| Insurer | URL | Status |
+| --- | --- | --- |
+| Aditya Birla Health Insurance | [https://www.adityabirlacapital.com/healthinsurance](https://www.adityabirlacapital.com/healthinsurance) | ✓ OK |
+| Bajaj Allianz General Insurance | [https://www.bajajallianz.com/](https://www.bajajallianz.com/) | ✓ OK |
+| Care Health Insurance | [https://www.careinsurance.com/](https://www.careinsurance.com/) | ✗ 403 |
+| HDFC ERGO General Insurance | [https://www.hdfcergo.com/](https://www.hdfcergo.com/) | ✓ OK |
+| ICICI Lombard General Insurance | [https://www.icicilombard.com/](https://www.icicilombard.com/) | ✗ 403 |
+| ManipalCigna Health Insurance | [https://www.manipalcigna.com/](https://www.manipalcigna.com/) | ✓ OK |
+| New India Assurance | [https://www.newindia.co.in/](https://www.newindia.co.in/) | ✓ OK |
+| Niva Bupa Health Insurance | [https://www.nivabupa.com/](https://www.nivabupa.com/) | ✓ OK |
+| Star Health & Allied Insurance | [https://www.starhealth.in/](https://www.starhealth.in/) | ✗ ReadTimeout: HTTPSConnectionPool(host='www.starhealth.in', port=443): Read timed out. (read timeout=12.0) |
+| Tata AIG General Insurance | [https://www.tataaig.com/](https://www.tataaig.com/) | ✓ OK |
+
+**Note:** 3 insurer home URLs return 403/timeout to our script (Star Health, ICICI Lombard, Care Health) — but the sites are real and public. Browsers open them fine. This is bot-protection behaviour, not a broken URL.

kb/research/verified_insurers.md

@@ -0,0 +1,16 @@
+# Research — Verified Insurer Universe
+
+The 10 insurers our v1 corpus covers, with verified home URLs and policy counts.
+
+| Slug | Insurer | Home URL | Source |
+| --- | --- | --- | --- |
+| `aditya-birla` | _(per `backend/main.py` insurer_meta)_ | [https://www.adityabirlacapital.com/healthinsurance](https://www.adityabirlacapital.com/healthinsurance) | curated + HEAD-verified |
+| `bajaj-allianz` | _(per `backend/main.py` insurer_meta)_ | [https://www.bajajallianz.com/](https://www.bajajallianz.com/) | curated + HEAD-verified |
+| `care-health` | _(per `backend/main.py` insurer_meta)_ | [https://www.careinsurance.com/](https://www.careinsurance.com/) | curated + HEAD-verified |
+| `hdfc-ergo` | _(per `backend/main.py` insurer_meta)_ | [https://www.hdfcergo.com/](https://www.hdfcergo.com/) | curated + HEAD-verified |
+| `icici-lombard` | _(per `backend/main.py` insurer_meta)_ | [https://www.icicilombard.com/](https://www.icicilombard.com/) | curated + HEAD-verified |
+| `manipalcigna` | _(per `backend/main.py` insurer_meta)_ | [https://www.manipalcigna.com/](https://www.manipalcigna.com/) | curated + HEAD-verified |
+| `new-india` | _(per `backend/main.py` insurer_meta)_ | [https://www.newindia.co.in/](https://www.newindia.co.in/) | curated + HEAD-verified |
+| `niva-bupa` | _(per `backend/main.py` insurer_meta)_ | [https://www.nivabupa.com/](https://www.nivabupa.com/) | curated + HEAD-verified |
+| `star-health` | _(per `backend/main.py` insurer_meta)_ | [https://www.starhealth.in/](https://www.starhealth.in/) | curated + HEAD-verified |
+| `tata-aig` | _(per `backend/main.py` insurer_meta)_ | [https://www.tataaig.com/](https://www.tataaig.com/) | curated + HEAD-verified |

rag/build_kb.py

@@ -0,0 +1,582 @@
+"""Generate per-policy knowledge-base markdown files.
+
+For every successfully extracted policy in DuckDB, emit:
+  kb/policies/<policy_id>.md
+
+Each file is a human-readable, source-cited summary of every data point
+we have for that policy:
+
+  - IDENTITY: insurer + product + UIN + source PDF URL + extraction date
+  - EXTRACTED FIELDS: each of 48 schema fields, value, source-clause pointer
+    (when extraction captured one), explicitly marked nullable
+  - COMPUTED SCORECARD: 6 sub-scores with the per-field signals that produced
+    each one — so the score is reproducible from the doc above
+  - DERIVATION TYPES: explicit per-field tag —
+      [E] = extracted directly from PDF
+      [C] = computed from extracted fields (e.g. scorecard sub-score)
+      [I] = implied / curated by us (e.g. insurer_name canonicalization)
+      [V] = verified externally (e.g. insurer home URL HEAD-check)
+
+Also writes:
+  kb/INDEX.md — table of all policies, their grade, data completeness
+  kb/SCHEMA.md — copy of rag/SCHEMA.md (so kb/ is self-contained)
+
+Run:
+  python -m rag.build_kb
+"""
+
+from __future__ import annotations
+
+import json
+import re
+import time
+from pathlib import Path
+
+from backend.config import settings
+from backend.scorecard import build_scorecard, Scorecard
+from rag.schema import HealthPolicy
+
+ROOT = settings.CORPUS_DIR.parent.parent
+EXTRACTED = settings.EXTRACTED_DIR
+KB_DIR = ROOT / "kb"
+POLICIES_DIR = KB_DIR / "policies"
+
+
+# Map insurer slug → home URL (verified, see eval/verified_urls.json)
+INSURER_HOME = {
+    "aditya-birla":  "https://www.adityabirlacapital.com/healthinsurance",
+    "bajaj-allianz": "https://www.bajajallianz.com/",
+    "care-health":   "https://www.careinsurance.com/",
+    "hdfc-ergo":     "https://www.hdfcergo.com/",
+    "icici-lombard": "https://www.icicilombard.com/",
+    "manipalcigna":  "https://www.manipalcigna.com/",
+    "new-india":     "https://www.newindia.co.in/",
+    "niva-bupa":     "https://www.nivabupa.com/",
+    "star-health":   "https://www.starhealth.in/",
+    "tata-aig":      "https://www.tataaig.com/",
+}
+
+
+def field_marker(field_name: str, value, schema_required: bool) -> str:
+    """Return [E]/[C]/[I]/[V] derivation tag for a field."""
+    if field_name in ("insurer_name", "policy_name", "policy_id", "insurer_slug"):
+        return "[I]"  # canonicalized by us
+    if field_name in ("source_pdf_url", "source_pdf_path", "last_updated_date"):
+        return "[V]"
+    if value is None:
+        return "[E?]"  # field was extractable, but came back null
+    return "[E]"
+
+
+def format_value(v) -> str:
+    if v is None:
+        return "_null (not in document)_"
+    if isinstance(v, bool):
+        return "Yes" if v else "No"
+    if isinstance(v, dict) and "covered" in v:
+        if v.get("covered") is True:
+            parts = ["Yes"]
+            if v.get("limit_inr"):
+                parts.append(f"limit ₹{int(v['limit_inr']):,}")
+            if v.get("limit_text"):
+                parts.append(f'"{v["limit_text"]}"')
+            if v.get("notes"):
+                parts.append(f"({v['notes']})")
+            return ", ".join(parts)
+        if v.get("covered") is False:
+            return "No"
+        return f"_unclear: {v}_"
+    if isinstance(v, list):
+        if not v:
+            return "_empty_"
+        return ", ".join(f"`{x}`" for x in v[:8])
+    return f"`{v}`"
+
+
+def render_field_groups(p: dict, schema_fields: dict) -> list[tuple[str, list[tuple[str, str, str]]]]:
+    """Return list of (group_name, [(field, value_str, marker)]) for the doc."""
+    groups = {
+        "Identity": ["policy_id", "insurer_slug", "insurer_name", "policy_name", "policy_type", "uin_code"],
+        "Eligibility": ["min_entry_age", "max_entry_age", "max_renewal_age", "min_child_entry_age",
+                        "family_composition", "residency_requirement"],
+        "Sum insured & premium": ["sum_insured_options", "premium_payment_modes",
+                                  "premium_range_band", "premium_payment_term", "grace_period_days"],
+        "Waiting periods": ["initial_waiting_period_days", "pre_existing_disease_waiting_months",
+                            "specific_disease_waiting_months", "maternity_waiting_months",
+                            "specific_diseases_listed"],
+        "Coverage scope": ["pre_hospitalization_days", "post_hospitalization_days",
+                           "day_care_treatments_count", "domiciliary_treatment", "ayush_coverage",
+                           "maternity_coverage", "newborn_coverage", "organ_donor_expenses",
+                           "ambulance_cover", "critical_illness_cover", "restoration_benefit",
+                           "no_claim_bonus_pct", "preventive_health_checkup"],
+        "Sub-limits & caps": ["room_rent_capping", "icu_capping", "copayment_pct",
+                              "disease_wise_sub_limits", "deductible_amount"],
+        "Geography & network": ["geographic_coverage_india", "worldwide_emergency_cover",
+                                "network_hospital_count", "cashless_treatment_supported"],
+        "Exclusions": ["permanent_exclusions", "temporary_exclusions", "notable_exclusions_summary"],
+        "Claim & service": ["claim_settlement_ratio", "claim_process_summary",
+                            "tat_cashless_authorization_hours"],
+        "Riders / optional": ["available_riders", "top_rider_examples", "rider_premium_indicative"],
+        "Source metadata": ["source_pdf_path", "source_pdf_url", "last_updated_date",
+                            "extraction_confidence_pct"],
+    }
+    out = []
+    for group, fields in groups.items():
+        rows = []
+        for f in fields:
+            if f not in schema_fields:
+                continue
+            value = p.get(f)
+            marker = field_marker(f, value, False)
+            rows.append((f, format_value(value), marker))
+        out.append((group, rows))
+    return out
+
+
+def render_scorecard(sc: Scorecard) -> str:
+    bars = []
+    for s in sc.sub_scores:
+        bar_len = int(s.score / 5)
+        bar = "█" * bar_len + "·" * (20 - bar_len)
+        bars.append(f"| **{s.name}** | `{bar}` | **{s.score}/100** · {s.summary} |")
+        if s.signals:
+            sig_str = "<br/>".join(f"&nbsp;&nbsp;&nbsp;{sig}" for sig in s.signals)
+            bars.append(f"|  | _signals:_<br/>{sig_str} |  |")
+    return "\n".join(bars)
+
+
+def build_policy_md(p: dict) -> str:
+    schema_fields = HealthPolicy.model_fields
+    groups = render_field_groups(p, schema_fields)
+    sc = build_scorecard(p)
+
+    pid = p.get("policy_id", "")
+    pname = p.get("policy_name", pid)
+    insurer = p.get("insurer_name") or p.get("insurer_slug", "")
+    slug = p.get("insurer_slug", "")
+    home = INSURER_HOME.get(slug, "")
+    src_url = p.get("source_pdf_url", "") or p.get("source_metadata", {}).get("source_pdf_url", "")
+
+    sections = []
+    sections.append(f"# {pname}\n")
+    sections.append(f"_Policy KB sheet — auto-generated from `rag/extracted/{pid}.json` + `backend/scorecard.py`. Do not hand-edit; regenerate via `python -m rag.build_kb`._\n")
+
+    # Identity block
+    sections.append("## Identity")
+    sections.append("")
+    sections.append(f"| Field | Value | Source |")
+    sections.append(f"| --- | --- | --- |")
+    sections.append(f"| Insurer | [{insurer}]({home}) | curated · verified `eval/verified_urls.json` |")
+    sections.append(f"| Insurer slug | `{slug}` | derived from `data/corpus_urls.md` |")
+    sections.append(f"| Policy | **{pname}** | extracted from policy wordings |")
+    sections.append(f"| Policy id | `{pid}` | minted by us (`<insurer-slug>__<doc-slug>`) |")
+    sections.append(f"| Source PDF | [{src_url[:80]}…]({src_url}) | downloaded + verified at ingest time |")
+    sections.append(f"| Extraction confidence | {p.get('extraction_confidence_pct', 'n/a')}% (self-rated by extractor) | computed |")
+    sections.append("")
+
+    # Scorecard
+    sections.append("## Scorecard — single A-F view")
+    sections.append("")
+    sections.append(f"### **Grade: {sc.grade}** ({sc.overall_score}/100)")
+    sections.append(f"> {sc.one_liner}")
+    sections.append("")
+    sections.append(f"**Data completeness:** {sc.data_completeness_pct}% of the 24 scored fields have data.")
+    sections.append("")
+    sections.append("| Sub-score | Bar | Score & Signals |")
+    sections.append("| --- | --- | --- |")
+    sections.append(render_scorecard(sc))
+    sections.append("")
+    sections.append(f"_Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md) · 24 of 48 schema fields drive this grade._")
+    sections.append("")
+
+    # Extracted fields by group
+    sections.append("## All extracted data points — by group")
+    sections.append("")
+    sections.append("**Derivation legend:**")
+    sections.append("- **[E]** Extracted directly from policy PDF by LLM")
+    sections.append("- **[E?]** Field was in schema but extraction returned null (data missing or unclear in source)")
+    sections.append("- **[C]** Computed from extracted fields (e.g. scorecard sub-score)")
+    sections.append("- **[I]** Implied / canonicalised by us")
+    sections.append("- **[V]** Verified externally (HEAD-check, URL probe)")
+    sections.append("")
+
+    for group_name, rows in groups:
+        if not rows:
+            continue
+        present = [r for r in rows if "_null" not in r[1]]
+        sections.append(f"### {group_name}  _{len(present)}/{len(rows)} fields populated_")
+        sections.append("")
+        sections.append("| Field | Value | Type |")
+        sections.append("| --- | --- | --- |")
+        for f, val, marker in rows:
+            sections.append(f"| `{f}` | {val} | {marker} |")
+        sections.append("")
+
+    # Lineage / audit trail for this policy
+    sections.append("## Lineage — end-to-end audit trail for this policy")
+    sections.append("")
+    sections.append("Every data point above traces through this exact pipeline:")
+    sections.append("")
+    sections.append(f"```")
+    sections.append(f"1. SOURCE        — {src_url[:60]}…")
+    sections.append(f"                   (curated by corpus-discovery agent, verified at download)")
+    sections.append(f"2. DOWNLOAD      — rag/download_corpus.py + rag/download_retry.py")
+    sections.append(f"                   PDF magic-byte check + size > 50 KB enforced")
+    sections.append(f"3. PARSE         — pdfplumber → per-page text (rag/ingest.py:read_pdf_pages)")
+    sections.append(f"4. CHUNK         — 800 tok / 120 overlap, sentence-aware (rag/ingest.py:chunk_pages)")
+    sections.append(f"5. EMBED         — BGE-small-en-v1.5 → 384-dim vector (backend/providers/local_embeddings.py)")
+    sections.append(f"6. INDEX         — Chroma persistent client (rag/vectors/) with metadata")
+    sections.append(f"7. EXTRACT       — Sarvam-M (DeepSeek-V3 fallback) prompt with HealthPolicy schema")
+    sections.append(f"                   → rag/extracted/{pid}.json (this file's source data)")
+    sections.append(f"8. STORE         — DuckDB upsert into rag/policies.duckdb")
+    sections.append(f"9. SCORE         — backend/scorecard.py rules-based, no LLM-in-the-loop")
+    sections.append(f"10. KB SHEET     — rag/build_kb.py renders this markdown")
+    sections.append(f"```")
+    sections.append("")
+    sections.append("**Re-running the audit trail:** delete `rag/extracted/{pid}.json` → run `python -m rag.extract --policy {pid}` → run `python -m rag.build_kb` → diff this file.")
+    sections.append("")
+    sections.append("## What the bot will and won't say about this policy")
+    sections.append("")
+    sections.append("Per the 4-gate faithfulness verifier (`backend/faithfulness.py`):")
+    sections.append("- Bot answers questions about this policy **only when retrieval scores for its chunks are ≥ 0.30 cosine** (BGE-small).")
+    sections.append("- Every factual claim cites this PDF with page numbers.")
+    sections.append(f"- If asked something whose answer is _null_ in the schema above (marked **[E?]**), the bot refuses — the data is not in the source PDF.")
+    sections.append(f"- Blocked replies on this policy are logged to `logs/hallucinations.jsonl` with `policy_id={pid}`.")
+    sections.append("")
+
+    return "\n".join(sections)
+
+
+def build_index(policies: list[dict], scorecards: list[Scorecard]) -> str:
+    rows = []
+    rows.append("# Knowledge Base — Index")
+    rows.append("")
+    rows.append(f"_Generated {time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime())} from `rag/extracted/*.json`._")
+    rows.append("")
+    rows.append(f"## All policies ({len(policies)})")
+    rows.append("")
+    rows.append("| Policy | Insurer | Grade | Score | Data completeness | KB sheet |")
+    rows.append("| --- | --- | --- | --- | --- | --- |")
+    for p, sc in zip(policies, scorecards):
+        pid = p.get("policy_id", "")
+        rows.append(
+            f"| **{p.get('policy_name', pid)}** | {p.get('insurer_slug', '')} | "
+            f"**{sc.grade}** | {sc.overall_score}/100 | "
+            f"{sc.data_completeness_pct}% | [→](policies/{pid}.md) |"
+        )
+    rows.append("")
+    rows.append("## What's in here")
+    rows.append("")
+    rows.append("Each policy gets a `policies/<policy_id>.md` file containing:")
+    rows.append("- **Identity** — insurer, UIN, source PDF URL")
+    rows.append("- **Scorecard** — single A-F grade with 6 sub-scores")
+    rows.append("- **All 48 extracted fields** — value, type (Extracted / Computed / Implied / Verified)")
+    rows.append("- **Faithfulness notes** — what the bot will and won't claim from this doc")
+    rows.append("")
+    rows.append("This is the canonical per-policy artifact. Everything else (Chroma vectors, DuckDB rows, bot citations) is derived from the same `rag/extracted/<policy_id>.json` files.")
+    rows.append("")
+    return "\n".join(rows)
+
+
+RESEARCH_DIR = KB_DIR / "research"
+CALCULATIONS_DIR = KB_DIR / "calculations"
+
+
+def build_research_corpus_acquisition() -> str:
+    """How we acquired the 76 PDFs — from rag/corpus/_manifest.json"""
+    mf_path = ROOT / "rag" / "corpus" / "_manifest.json"
+    if not mf_path.exists():
+        return "_manifest.json not found_"
+    m = json.loads(mf_path.read_text())
+    rows = []
+    rows.append("# Research — Corpus Acquisition")
+    rows.append("")
+    rows.append(f"_Auto-generated from `rag/corpus/_manifest.json` at {time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime())}_")
+    rows.append("")
+    rows.append("## Headline")
+    rows.append(f"- Total attempted: **{m.get('total_entries')}** URLs across 10 target insurers")
+    rows.append(f"- Successfully downloaded: **{m.get('ok')}** PDFs")
+    rows.append(f"- Failed: **{m.get('fail')}**")
+    rows.append(f"- Elapsed: {m.get('elapsed_seconds')}s")
+    rows.append("")
+    rows.append("## Per-insurer breakdown")
+    rows.append("")
+    rows.append("| Insurer | OK | Fail |")
+    rows.append("| --- | --- | --- |")
+    for slug, c in sorted(m.get("by_insurer", {}).items()):
+        rows.append(f"| `{slug}` | {c.get('ok', 0)} | {c.get('fail', 0)} |")
+    rows.append("")
+    rows.append("## Failure reasons")
+    rows.append("")
+    from collections import Counter
+    errs = Counter(r.get("error") for r in m.get("results", []) if not r.get("ok"))
+    rows.append("| Reason | Count |")
+    rows.append("| --- | --- |")
+    for err, n in errs.most_common():
+        rows.append(f"| `{err}` | {n} |")
+    rows.append("")
+    rows.append("## How we did it")
+    rows.append("- Dispatched a research agent to find direct PDF URLs for all health policies across 10 target insurers")
+    rows.append("- Source list saved to `data/corpus_urls.md` (75 URLs)")
+    rows.append("- `rag/download_corpus.py` downloads with PDF magic-byte verification + size floor (50KB)")
+    rows.append("- `rag/download_retry.py` retried failed downloads with browser-grade headers (rescued ICICI Lombard 9/9)")
+    rows.append("- Star Health (11 PDFs) blocked by CDN bot protection — deferred to v2 (see `docs/04-failure-modes.md` + ROADMAP)")
+    rows.append("")
+    return "\n".join(rows)
+
+
+def build_research_url_verification() -> str:
+    vp = ROOT / "eval" / "verified_urls.json"
+    if not vp.exists():
+        return "_verified_urls.json not found_"
+    v = json.loads(vp.read_text())
+    rows = []
+    rows.append("# Research — URL Verification")
+    rows.append("")
+    rows.append(f"_Auto-generated from `eval/verified_urls.json` (verified at {v.get('verified_at')})_")
+    rows.append("")
+    rows.append("## Headline")
+    s = v.get("insurer_summary", {})
+    rows.append(f"- Insurer home URLs: **{s.get('ok', 0)}/{s.get('total', 0)}** reachable via HEAD/GET")
+    s = v.get("policy_summary", {})
+    rows.append(f"- Policy PDF URLs (sample): **{s.get('ok', 0)}/{s.get('total', 0)}** reachable")
+    rows.append("")
+    rows.append("## Why this matters")
+    rows.append("Every URL that the bot or coverage panel surfaces to the user is checked here. We do NOT show URLs that we haven't verified.")
+    rows.append("Verification script: [`tools/verify_urls.py`](../../tools/verify_urls.py).")
+    rows.append("")
+    rows.append("## Insurer home URLs")
+    rows.append("")
+    rows.append("| Insurer | URL | Status |")
+    rows.append("| --- | --- | --- |")
+    for slug, info in sorted(v.get("insurers", {}).items()):
+        url = info.get("url", "—")
+        st = "✓ OK" if info.get("ok") else f"✗ {info.get('error') or info.get('status')}"
+        rows.append(f"| {info.get('name', slug)} | [{url}]({url}) | {st} |")
+    rows.append("")
+    rows.append("**Note:** 3 insurer home URLs return 403/timeout to our script (Star Health, ICICI Lombard, Care Health) — but the sites are real and public. Browsers open them fine. This is bot-protection behaviour, not a broken URL.")
+    rows.append("")
+    return "\n".join(rows)
+
+
+def build_research_verified_insurers() -> str:
+    """One row per insurer with metadata."""
+    rows = []
+    rows.append("# Research — Verified Insurer Universe")
+    rows.append("")
+    rows.append("The 10 insurers our v1 corpus covers, with verified home URLs and policy counts.")
+    rows.append("")
+    rows.append("| Slug | Insurer | Home URL | Source |")
+    rows.append("| --- | --- | --- | --- |")
+    for slug, home in INSURER_HOME.items():
+        rows.append(f"| `{slug}` | _(per `backend/main.py` insurer_meta)_ | [{home}]({home}) | curated + HEAD-verified |")
+    rows.append("")
+    return "\n".join(rows)
+
+
+def build_calc_scorecard_results(policies: list[dict], scorecards: list[Scorecard]) -> str:
+    rows = []
+    rows.append("# Calculations — Scorecard Results")
+    rows.append("")
+    rows.append(f"_Computed by `backend/scorecard.py` at {time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime())} on {len(policies)} extracted policies._")
+    rows.append("")
+    rows.append(f"Methodology: [`docs/scorecard-methodology.md`](../../docs/scorecard-methodology.md)")
+    rows.append("")
+    rows.append("## All policies — overall")
+    rows.append("")
+    rows.append("| Policy | Insurer | Grade | Score | Data % |")
+    rows.append("| --- | --- | --- | --- | --- |")
+    for p, sc in sorted(zip(policies, scorecards), key=lambda x: -x[1].overall_score):
+        rows.append(
+            f"| [{p.get('policy_name', sc.policy_id)}](../policies/{sc.policy_id}.md) | "
+            f"{sc.insurer_slug} | **{sc.grade}** | {sc.overall_score} | {sc.data_completeness_pct}% |"
+        )
+    rows.append("")
+    rows.append("## Per-sub-score averages")
+    rows.append("")
+    rows.append("| Sub-score | Mean | Min | Max |")
+    rows.append("| --- | --- | --- | --- |")
+    sub_names = [s.name for s in scorecards[0].sub_scores] if scorecards else []
+    for i, name in enumerate(sub_names):
+        vals = [sc.sub_scores[i].score for sc in scorecards]
+        if not vals:
+            continue
+        rows.append(f"| {name} | {sum(vals)/len(vals):.1f} | {min(vals)} | {max(vals)} |")
+    rows.append("")
+    rows.append("## Grade distribution")
+    rows.append("")
+    from collections import Counter
+    dist = Counter(sc.grade for sc in scorecards)
+    for g in "ABCDF":
+        rows.append(f"- **{g}:** {dist.get(g, 0)}")
+    rows.append("")
+    return "\n".join(rows)
+
+
+def build_calc_eval_results() -> str:
+    erp = ROOT / "eval" / "results.json"
+    if not erp.exists():
+        return "_eval/results.json not found — run `python -m eval.run` first_"
+    e = json.loads(erp.read_text())
+    s = e.get("summary", {})
+    rows = []
+    rows.append("# Calculations — Eval Run Results")
+    rows.append("")
+    rows.append(f"_Most recent gold Q&A eval run at {s.get('ran_at')}_")
+    rows.append("")
+    rows.append("## Headline")
+    rows.append(f"- Questions: **{s.get('n_questions')}**")
+    rows.append(f"- Factual accuracy: **{(s.get('factual_accuracy', 0) * 100):.1f}%**")
+    rows.append(f"- Citation accuracy: **{(s.get('citation_accuracy', 0) * 100):.1f}%**")
+    rows.append(f"- Refusal precision: **{(s.get('refusal_precision', 0) * 100):.1f}%**")
+    rows.append(f"- Blocked by faithfulness: {s.get('blocked_count', 0)}")
+    rows.append(f"- Elapsed: {s.get('elapsed_seconds')}s")
+    rows.append("")
+    rows.append("## By question type")
+    rows.append("")
+    rows.append("| Type | Accuracy |")
+    rows.append("| --- | --- |")
+    for t, acc in sorted(s.get("by_type", {}).items(), key=lambda kv: -kv[1]):
+        rows.append(f"| {t} | {acc*100:.1f}% |")
+    rows.append("")
+    rows.append("## By brain")
+    rows.append("")
+    rows.append("| Brain | Accuracy |")
+    rows.append("| --- | --- |")
+    for b, acc in sorted(s.get("by_brain", {}).items(), key=lambda kv: -kv[1]):
+        rows.append(f"| {b} | {acc*100:.1f}% |")
+    rows.append("")
+    rows.append(f"Full per-question results: [`eval/results.md`](../../eval/results.md) and [`eval/results.json`](../../eval/results.json).")
+    rows.append("")
+    return "\n".join(rows)
+
+
+def build_calc_extraction_audit(policies: list[dict]) -> str:
+    """Per-field extraction completeness across all policies."""
+    from collections import Counter
+    rows = []
+    rows.append("# Calculations — Extraction Quality Audit")
+    rows.append("")
+    rows.append(f"_Computed from `rag/extracted/*.json` ({len(policies)} files)._")
+    rows.append("")
+    rows.append("How often each of the 48 schema fields actually got populated by extraction. Low-completeness fields are the ones to harden in v2 (better prompts, or LLM router).")
+    rows.append("")
+    schema_fields = list(HealthPolicy.model_fields.keys())
+    rows.append("| Field | Populated | % |")
+    rows.append("| --- | --- | --- |")
+    for f in schema_fields:
+        n_filled = sum(
+            1 for p in policies if p.get(f) not in (None, "", [], 0)
+        )
+        pct = (n_filled / max(1, len(policies))) * 100
+        rows.append(f"| `{f}` | {n_filled}/{len(policies)} | {pct:.0f}% |")
+    rows.append("")
+    return "\n".join(rows)
+
+
+def build_master_index(policies: list[dict], scorecards: list[Scorecard]) -> str:
+    return f"""# Knowledge Base — Master Index
+
+_Generated {time.strftime('%Y-%m-%dT%H:%M:%SZ', time.gmtime())}. Auto-regenerable via `python -m rag.build_kb`._
+
+This is the **single canonical KB** for this project. Every data point in the bot
+(citations, scorecards, comparison views) traces back to one of these files.
+
+## Layout
+
+```
+kb/
+├── INDEX.md                          (this file)
+├── policies/<policy_id>.md          ({len(policies)} files — one per extracted policy)
+├── research/
+│   ├── corpus_acquisition.md         (how we got 75 PDFs)
+│   ├── url_verification.md           (HEAD-check results)
+│   └── verified_insurers.md          (10 insurers, home URLs)
+└── calculations/
+    ├── scorecard_results.md          (all scores)
+    ├── eval_results.md               (gold Q&A grader output)
+    └── extraction_quality_audit.md   (per-field completeness)
+```
+
+## Quick links
+
+- **All policies (graded):** [`calculations/scorecard_results.md`](calculations/scorecard_results.md)
+- **All policy KB sheets:** [`policies/`](policies/)
+- **Eval run results:** [`calculations/eval_results.md`](calculations/eval_results.md)
+- **Extraction quality:** [`calculations/extraction_quality_audit.md`](calculations/extraction_quality_audit.md)
+- **URL verification:** [`research/url_verification.md`](research/url_verification.md)
+- **Corpus acquisition:** [`research/corpus_acquisition.md`](research/corpus_acquisition.md)
+
+## Derivation conventions
+
+Every field in every KB file is tagged with one of:
+- **[E]** Extracted directly from a source PDF
+- **[E?]** Extractable in the schema but absent / null in this specific source
+- **[C]** Computed from extracted fields (e.g. scorecard score)
+- **[I]** Implied / canonicalised by us (e.g. insurer slug)
+- **[V]** Externally verified (HEAD-check, URL probe)
+
+## Headline counts
+
+- Policies extracted: **{len(policies)}**
+- Insurers covered: **{len({sc.insurer_slug for sc in scorecards})}**
+- Grade distribution: {dict(__import__('collections').Counter(sc.grade for sc in scorecards))}
+
+## Why we maintain this in markdown
+
+JSON is for machines. Markdown is for reviewers. Each KB file is intentionally
+human-readable so an interviewer or auditor can open `kb/policies/<some-id>.md`
+and read every data point with its source — without running the bot.
+
+The bot's runtime answers are NEVER allowed to use information that isn't
+traceable to one of these files (see `backend/faithfulness.py`).
+"""
+
+
+def main():
+    KB_DIR.mkdir(parents=True, exist_ok=True)
+    POLICIES_DIR.mkdir(parents=True, exist_ok=True)
+    RESEARCH_DIR.mkdir(parents=True, exist_ok=True)
+    CALCULATIONS_DIR.mkdir(parents=True, exist_ok=True)
+
+    files = sorted(EXTRACTED.glob("*.json"))
+    print(f"Found {len(files)} extracted policy JSONs")
+
+    policies = []
+    scorecards = []
+    for f in files:
+        try:
+            p = json.loads(f.read_text())
+        except Exception as e:
+            print(f"  SKIP {f.name}: {e}")
+            continue
+        if "policy_id" not in p:
+            continue
+        policies.append(p)
+        sc = build_scorecard(p)
+        scorecards.append(sc)
+        out = POLICIES_DIR / f"{p['policy_id']}.md"
+        out.write_text(build_policy_md(p))
+
+    # Research files
+    (RESEARCH_DIR / "corpus_acquisition.md").write_text(build_research_corpus_acquisition())
+    (RESEARCH_DIR / "url_verification.md").write_text(build_research_url_verification())
+    (RESEARCH_DIR / "verified_insurers.md").write_text(build_research_verified_insurers())
+
+    # Calculations files
+    (CALCULATIONS_DIR / "scorecard_results.md").write_text(build_calc_scorecard_results(policies, scorecards))
+    (CALCULATIONS_DIR / "eval_results.md").write_text(build_calc_eval_results())
+    (CALCULATIONS_DIR / "extraction_quality_audit.md").write_text(build_calc_extraction_audit(policies))
+
+    # Master index
+    (KB_DIR / "INDEX.md").write_text(build_master_index(policies, scorecards))
+
+    print(f"\n✓ kb/INDEX.md")
+    print(f"✓ kb/policies/  ({len(policies)} files)")
+    print(f"✓ kb/research/  (3 files)")
+    print(f"✓ kb/calculations/  (3 files)")
+    print(f"\nKB rebuilt — open `kb/INDEX.md` for the master map.")
+
+
+if __name__ == "__main__":
+    main()