feat: strict Graph-RAG reasoning engine with JSON path output + validation script

feat: switch retrieval to knowledge-graph GraphRAG

fix: auto-detect qdrant creds and make gemini dependency optional

feat: add /vector_store_status endpoint for qdrant health and point count

feat: add optional Qdrant Cloud vector persistence with memory fallback

fix: collapse single newlines to spaces in PDF extraction so chunk overlap works correctly

fix: rejoin hyphenated line-breaks in PDF text extraction

feat: add GET /chunks endpoint with pagination and keyword filter

feat: add /chunk_context endpoint to expose neighbouring chunks

fix: filter weak retrieval chunks and recalibrate confidence

feat: grounded answering with confidence metadata and strict mode

feat: streaming LLM responses via SSE /query/stream endpoint

fix: prevent embedding dimension mismatch when provider falls back mid-ingest

fix: track active embed provider to prevent dimension mismatch on query

fix: skip retries, instant HF fallback on Gemini failure

fix: use InferenceClient for HF embeddings (old endpoint 410 Gone)

fix: auto-fallback to HF embeddings when Gemini rate-limited

fix: retry with backoff on Gemini 429 rate limit

migrate to google-genai SDK + update LLM models

fix: embedding shape for query + add LLM error logging

fix: use gemini-embedding-001 model

initial deploy