InsightDocs — ドキュメント横断の RAG

Overview

InsightDocs is a retrieval-augmented generation (RAG) application for querying Markdown documentation. You upload docs, ask questions in plain English, and get answers with inline citations that link back to the exact source chunks — so every answer is verifiable.

What it does

• Hybrid retrieval — combines dense (semantic) and sparse (BM25 keyword) search using Reciprocal Rank Fusion for better recall
• Header-aware chunking — preserves document hierarchy as metadata so retrieved context stays coherent
• Cross-encoder reranking — BAAI/bge-reranker-base reorders candidates for relevance before they reach the model
• Citations & refusal — an expandable source panel shows the retrieved text, and the system refuses to answer when context is insufficient
• Evaluation harness — an ablation setup compares four retrieval configurations

Stack

Python · Streamlit (UI) · ChromaDB (vector store) · BM25 · OpenRouter for embeddings and chat · deployed on Hugging Face Spaces.

Why it's interesting

The focus is on trustworthy retrieval: hybrid search plus reranking plus a refusal mechanism means the system prefers saying "I don't know" over hallucinating — and always shows its work.