fix: prevent LLM from hallucinating queries during contextualization

fix: preload models on startup and increase LLM timeout

fix: soften strict prompt for keyword queries

fix: switch to multilingual reranker

chore: remove IS_RENDER and gemini imports

feat(rag): use bkai huggingface model universally

fix(prompt): allow LLM to respond to chit-chat naturally

fix(rag): update prompt and reduce temperature to prevent hallucinations

fix: use models/gemini-embedding-2 with 768 dimensions for compatibility with local database

fix: correct GoogleGenerativeAIEmbeddings class name

fix: make heavy ML libraries import conditionally on Render to prevent OOM/timeouts

fix: strip carriage returns from database url

feat: tích hợp supabase database và cấu hình github actions ci

“…I”

“…I”

hoàn thiện Multimodal RAG chatbot với FastAPI Backend & Web UI