Revert all changes to app/services/rag_pipeline.py to match original RAG v1

Add pipeline phase timing logs to help debug Hugging Face space speed

Integrate RAG v1.2 admin panel and handoff features

Fix intent map: compound keywords, follow-up starters, prorata robustness

Fix follow-up detection: add prepositional starters for context continuity

Fix prorata and single-word HR term retrieval robustness

Fix inconsistent leave and per-annum responses

Fix acceptability retrieval; shorten redirect responses

Fix retrieval for prorata queries

Improve bot: self-identity, non-English redirect, privacy-tone for personal data, no hallucination on people/org, varied fallbacks

Fix: Broad HR queries + Anti-hallucination hardening (no fake careers page)

Fix: Conversational Intelligence - Natural greetings, varied responses, no robot loop

Production Intelligence: Graceful Guidance + Full Intent Coverage + Urdu Support

Upgrade: Hybrid Intelligence (LLM Query Rewriter + Local Expansion)

Upgrade: Intelligence & Reasoning (Added Format D + Expanded Intent Map)

Fix follow-up typo handling (then/them) + prevent specific leave type listing other types

Add conversational follow-up resolver: pronouns (them/it/this) now resolve to last topic for accurate retrieval

Hardening RAG Intelligence: Restoration of perfect format decision table, optimized thresholds, and query expansion logic for GPT-4o-mini

feat: institutional-grade intelligence upgrade with hybrid expansion, logical deduction, and core reference data

Restore: Revert codebase to stable May 14th 6 PM state

feat: hardcode leave entitlements in system prompt for guaranteed completeness

feat: further intelligence boost via multi-intent matching and relaxed rerank thresholds

feat: increase HR assistant intelligence via logical deduction and broader retrieval

Restore to latest backup from 13 May - 2026-05-15 17:56

Generic: Replace hardcoded intent map with universal keyword extractor, fix 'can i resign'

Fix: Disable stale cache, add ul/li CSS reset, unify post-processing with _clean_html

Fix: Replace LLM expander with instant hybrid normalizer, fix re-index, fix prorata

Optimization: Lowered relevance threshold for better recall

Performance: Professional batched search and latency optimizations

Upgrade: Professional intelligent retrieval pipeline with robust tokenization and LLM query expansion

Fix: Add specific intent rules for 'half day' and 'unauthorized' leaves to prevent generic catch-all matching

Fix: add clean topic names for LLM while preserving rich keyword search variants

Refactor intent map with human-friendly topic names to fix short query ambiguity

Fix: rewrite short/vague queries for LLM to prevent false no-info responses

Enhance sandwich deduction intent mapping and query expansion logic

Eliminate first Groq API call: replace LLM query rewrite with instant local keyword expander (~3-6s saved per request)

Optimize RAG pipeline for speed: persistent HTTP client, faster reranker, and result caching

Fix reranker filtering out casual/sick leaves by enriching rerank query with expanded LLM keywords

feat: enhance RAG intelligence - intent-first protocol, /no_think for Qwen3, smart rerank fallback, context word cap

Implement DeepSeek-R1, BGE-Large, and Hybrid Search

Upgraded RAG pipeline for ChatGPT-level depth and intelligence (Multi-query, larger context, smarter prompts)

Upload 47 files

wipe space

Implement Deep Intelligence: Query Expansion, Reranking, and High-Precision Chunking

Fix Admin sidebar sync and increase relevance threshold to 0.45

Enhance AI intelligence and personalization with user names

feat: add secure admin panel, session store, localStorage tracking