Overview
Built for research and study of medical textbooks.
Medical Agentic RAG turns scanned Vietnamese medical textbooks into searchable sections and finds the passages that match a question, each with its book, section, page range and original page image. The agent steps on top of the search are planned (see Next).
Dai's figures, stated as goals and not a measured result of the code: about 2,000 books and 99.99%.
Not for clinical or diagnostic use. Answers come from textbooks, which can be dated or wrong.
Who it is for
- Medical students, lecturers and researchers who study from Vietnamese textbooks.
- Readers who need the exact page a passage comes from.
How it works
- AI transcription of scanned pages with code checks: a layout model crops each page, sections follow the book's own table of contents, hard pages go to a stronger model, and Vietnamese diacritics are checked.
- Sections and page images are kept in a corpus store.
- Hybrid search: two embedding models plus BM25 keyword search on accented and folded Vietnamese, fused with RRF, then reranked.
What is in the code
- A search API that returns ranked sections with book, section, page range and the original page image.
- A small test page to try the search in a browser.
- Every step traced, from the question to the ranked results.
Tech stack
Python, FastAPI, PostgreSQL, MinIO, Qdrant, OpenSearch, Langfuse, Docker Compose, PyMuPDF, DocLayout-YOLO, vision LLMs, Qwen3 embeddings and reranker.
Next
Planned, not built yet:
- An agent that plans the search: rewrites or splits a question and restores missing diacritics.
- Result checks: search again, or say the answer is not in the books.
- Written answers with citations to book, section and page.