docs: per-package reference + showcase site (GitHub Pages) (#107)

ci: Trivy security scan — deps (pinned), secrets, misconfig → Security tab (#104)

retrieval: remove the cross-encoder reranker (ablation showed no earned keep) (#100)

ci: Backfill content workflow (skip crawl) + empty-snapshot purge guard (#98)

ci: Calibrate guard workflow to re-derive the topicality threshold (#95)

eval: built-in rerank ablation, judge on paid hy3, answer-latency measurement (#94)

enrichment: paid hy3, batch 50, and per-batch checkpoint push (#93)

Follow client-side redirects + read_page URL guard: real API content (#92)

Make the deploy usable: strip/redact API keys + gate non-English + faster QuOTE (#91)

QuOTE questions in the enrichment pipeline + one-click eval default (#88)

Eval knobs as workflow inputs: rerank on/off and k (#86)

Reranker as the precision stage + fold glossing into Build Index (#85)

Answer-quality eval (LLM-judge) + hy3-led OpenRouter default (#81)

Consolidate the Actions list: 9 workflows -> 5, one Eval button

Save eval/gloss results even when main moves mid-run; fit glosses to real quotas

Switch the LLM workflows to Gemini by default (own free quota)

Harden the OpenRouter path: no-choices 200s retry, dead deepseek slug removed

Remove superseded manual workflows from the Actions list

Contextual Retrieval: LLM glosses for api reference pages

Parametrize the embedding model for a proper bake-off on v1

Make the agentic eval completable: bounded sample + 120m timeout

Add one-off retrieval diagnostic: locate expected page in per-kind pools

Add the agentic benchmark: coverage via citations, agentic vs single-shot

Add the v1 eval dataset: 100 valid + 100 invalid + 20 agentic

Ground the eval in EXTERNAL docs truth, not our own index

Add an index-manifest dump to ground the eval set in verified sources

Revert the embedder to bge-small: the bge-base A/B showed no gain

Swap the embedder: bge-small (384d) -> bge-base (768d)

Add a retrieval benchmark: recall@k + MRR over the v0 questions

Guard: replace the injection classifier with embedding-space membership

OpenRouter free-tier resilience + GEMINI alias + live HF Space smoke test (#41)

Sync workflow: read HF token from the GITHUBPUSH secret

Auto-sync main to the HF Space on every push (no manual git dance)

M3.4: LangGraph rewrite of the agent + loop-vs-langgraph comparison

openai-compat: fall back across a chain of free models on rate-limit

Eval workflow: default to grounded=true (the real M2 test)

Add non-English query translation (via OpenRouter) before retrieval

M2 wiring: grounded answers with validated citations + grounded eval mode

Add search CLI + Search workflow: retrieval acceptance testing from the Actions UI

Switch embeddings to a local CPU model (bge-small): free tier APIs can't index the corpus

Workflows: read secrets by their actual repo names (GOOGLE_API, NEON)

Build Index: fail fast on missing secrets, stream progress unbuffered

Build Index workflow: full pipeline on Actions + weekly scheduled recrawl

Eval workflow: model as input (free slugs rotate); fail fast on 404 model-gone

Eval workflow: read the OpenRouter key from the 'OPENROUTER' repo secret

Add manual Eval workflow: run v0 eval on Actions runners, commit results back

M2 ingestion: discover (objects.inv + sitemap), crawl (hash-diff snapshot), heading chunker, CI