E-will-era/easy-welfare-guide

Deduplicate RAG results by parent document ID

Switch embedding model to bge-m3 (multilingual) for Korean document support

Generate RAG search queries in English to match bge-en embedding model

Increase query gen keywords from 3 to 5 for better RAG recall

Limit query gen output to 3 keywords and apply max_tokens to API calls

Relax RAG retrieval to surface more related documents

Catch all LLM errors for retry and add call start logging

Add lock timeout and reduce call interval to prevent infinite hang

Release lock during retry backoff to unblock other requests

Serialize concurrent LLM calls with asyncio Lock

Increase rate limit guard interval and retry resilience

Filter old MCP results and boost recent ones in scoring

Fix MCP search using metadata instead of program name

Fix expiry check missing dates without year in end date

Add rate limit guard with minimum interval between LLM calls

Fix document ID cache init hitting SQLite variable limit

Fix BM25 SQL variable limit error and align embedding model

Update frontend/src/components/input/GuideSection.jsx

Update backend/app/rag/engine.py

Update backend/app/rag/engine.py

Update backend/app/rag/engine.py

Fix document duplication, optimize eligibility flow, and improve UX

Clean initial import