DocRAG

RAG-based question answering over technical PDFs, with hybrid retrieval (BM25 + vector search).

Built to explore how retrieval quality changes when lexical and semantic search are fused, motivated by earlier work using LLMs to extract implementation-defined parameters from the RISC-V Privileged Spec, where pure semantic retrieval often missed exact-term matches (register names, CSR fields).

Pipeline

PDF → PyMuPDF parsing → sliding-window chunking (900 chars, 150 overlap)
    → MiniLM embeddings → ChromaDB (vector index)
    → BM25 index (rank_bm25)
    → hybrid retrieval: score = 0.6·vector + 0.4·BM25 (min-max normalized)
    → Gemini 3.5 Flash answer, grounded in retrieved chunks, with page citations

Why hybrid search

Technical documents are full of exact identifiers (stvec, MTVEC_MODE, part numbers) that embedding models treat as noise. BM25 catches exact-term matches; vector search catches paraphrases and conceptual questions. Fusing both gives better recall than either alone.

Run

pip install -r requirements.txt
export GEMINI_API_KEY=your_key   # answers need this; retrieval works without it
streamlit run app.py

Upload a PDF, index it, ask questions. Retrieved chunks are shown with pages and fusion scores so retrieval quality is inspectable.

Design choices

• Chunking: character-based sliding window with overlap. Simple, predictable, keeps page metadata for citations.
• Embeddings: all-MiniLM-L6-v2, runs locally, no API cost for indexing.
• Fusion: min-max normalization of both score distributions before weighted sum (α = 0.6 toward vector). Raw BM25 and cosine scores live on different scales, so normalization is required before fusing.
• Grounding: the LLM is instructed to answer only from context and cite page numbers, and to say when the context doesn't contain the answer.

Next steps

• Reranking retrieved chunks with a cross-encoder
• OCR path for scanned PDFs (Tesseract)
• Retrieval evaluation harness (recall@k on a labeled question set)