// the take
A LangGraph-orchestrated multi-agent chatbot that combines RAG over medical PDFs, web search, and a few PyTorch vision models (chest X-ray classification, skin lesion segmentation) behind a FastAPI app. It's aimed at developers who want a reference implementation for agentic RAG and agent-to-agent handoff patterns, not at anyone looking for a clinically validated diagnostic tool.
The retrieval pipeline is genuinely thought out: Qdrant hybrid search (BM25 + dense vectors), cross-encoder reranking, and LLM-generated semantic chunking that respects document structure instead of naive fixed-size splitting. The confidence-based handoff between the RAG agent and the web-search agent, using log-probability thresholds rather than a second LLM call to decide when to fall back, is a nice touch most RAG demos skip. Docling-based parsing embeds LLM summaries of extracted images and tables alongside text, which actually handles multimodal PDFs better than text-only extraction.
Brain tumor detection is listed as a feature but marked '(TBD)' in the README — it's not implemented, just planned. The vision models (chest X-ray, skin lesion) ship as .pth files with no stated validation metrics, datasets, or accuracy figures, which matters a lot when the pitch is medical diagnosis. The stack is hardwired to Azure OpenAI (gpt-4o, ada-002 embeddings) by default; switching to OpenAI direct or local models requires code changes 'throughout the codebase,' per the author's own README, not config changes. Last push was May 2025 with no activity since, so despite 1k stars this looks like a frozen solo portfolio project rather than something with ongoing maintenance behind it.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →