Add manual Fast and Deep mode switch

Fix medical math rendering and chat session continuity

Move managed data sync off API startup

Update Gemini model for current API access

Deploy guideline-first DeepMed with Gemini rotation

fix: BM25 case-insensitive + add price to drug prefix

fix: remove recursive self.initialize() causing infinite loop

fix: syntax error in initialize() - corrupted text

fix: drug-aware chunking, streaming cleanup, conditional refs

Gemini thinking: budget 1024 tokens instead of fully disabled

Perf: add SSE streaming, disable Gemini thinking, skip query rewrite when no history

Upgrade Deep mode: 5-step clinical reasoning (guidelines -> pharmacovigilance -> drug matching -> treatment plan); expand auto-detect keywords

Update: NVIDIA fallback to Llama-3.3-70B, skip drug folder in RAG, use consolidated drug reference

fix: include NVIDIA_API_KEY in bot initialization check

feat: Gemini multi-key rotation (key1 → key2 → key3 → NVIDIA → OpenRouter)

feat: move ChromaDB to HF Dataset (PBThuong/deepmed-chroma-db)

feat: switch fallback LLM to DeepSeek V3.2 via NVIDIA Build (free)

Thiết lập lại thư viện y khoa sạch và cập nhật chroma_db