fix: score-threshold filtering + clear sources when answering from general knowledge

feat: deduplicated, human-readable source citations with page numbers

fix: sidebar checkbox visibility + LLM falls back to general knowledge

fix: strip DeepSeek R1 <think> reasoning tags from responses

fix: replace Google embeddings with HuggingFace sentence-transformers (no API key needed)

fix: pin langchain-google-genai<2.0 to restore working embedding-001 model

fix: update Google embedding model from embedding-001 to text-embedding-004

fix: robust Pydantic retriever + friendly error messages in UI

fix: use get_index() in ask endpoint to auto-create Pinecone index if missing

fix: replace deprecated langchain.text_splitter with langchain-text-splitters

fix: use absolute import path for load_vectorstore module

fix: update BaseRetriever import to langchain_core.retrievers

fix: lazy Pinecone init so FastAPI starts cleanly on HF Spaces

feat: production deployment setup for HF Spaces