This project is currently in testing mode. It uses local hard disk files for storage (BM25 index, Qdrant data, and document state). This is not production-ready and is intended for local experimentation.

Adaptive RAG is a structure-aware retrieval-augmented generation (RAG) system that:

• Parses documents (PDF, DOCX, Markdown)
• Detects document structure and sections
• Chooses a retrieval strategy based on document structure and question type
• Retrieves via hybrid vector + keyword search
• Reranks results and generates answers using an LLM

1. Ingestion: Upload a document to /upload. The file is parsed and chunked.
2. Indexing: Chunks are stored in:
   • Qdrant (local): Vector search index
   • BM25 (local): Keyword search index
3. Querying: Ask a question at /ask.
4. Routing: Strategy is selected by document structure and question type.
5. Retrieval + Rerank: Hybrid retrieval and cross-encoder reranking.
6. Answering: LLM generates the final answer using retrieved context.

• GET / - Health message
• GET /health - Basic health check
• POST /upload - Upload and index a document
• POST /ask - Ask a question

These files and folders are created on disk during testing:

• document_state.json - Document profile and sections
• bm25s_index.pkl - BM25 keyword index
• qdrant_data/ - Local Qdrant vector store

See requirements.txt for dependencies.

uvicorn app.main:app --reload

• This build is not hardened for concurrent users.
• Indexing recreates the vector collection per upload.
• Use only for local testing until storage, concurrency, and security are improved.