A local RAG system that

A local RAG system that doesn't just retrieve and answer — it evaluates whether its own retrieval was good enough, and retries if not.

Upload a PDF, DOCX, or TXT file and ask questions against it. Instead of blindly trusting the first retrieval, the pipeline runs a critic/grounding stage that checks whether the retrieved context actually supports the answer — and if it doesn't, it rewrites the query and tries again (up to 3 times) before gracefully refusing.

LangGraph-orchestrated self-healing workflow: retrieve → generate → critique → rewrite → re-retrieve

Local embeddings (qwen3-embedding:0.6b) and generation (qwen2.5-coder:7b) via Ollama — fully offline inference

Persistent vector storage with ChromaDB

OCR fallback (Tesseract + pdf2image) for scanned PDFs

Transparent execution trace: similarity scores, retry count, grounding status, and rewritten queries all visible in the UI

Explicitly refuses to answer out-of-context questions instead of hallucinating