Skip to content
Atmane Ayoub
← All work

Open source · 2025

Multi-agent healthcare assistant

An open-source multi-agent system for medical questions: agentic RAG with hybrid search and reranking, medical-image analysis agents, web search, confidence-based routing and human-in-the-loop validation.

Response time vs search
−40%

Role

Author

What it is

A chatbot for medical diagnosis support, research and patient questions, built as a set of specialised agents rather than one prompt.

How it works

  • Multi-agent orchestration with LangGraph: separate agents for retrieval, reasoning, medical-image analysis and web search, with agent-to-agent handoff.
  • Agentic RAG: Docling-based parsing of PDFs (text, tables and images), LLM-based semantic chunking that respects document structure, and LLM query expansion with related medical terms.
  • Hybrid retrieval in Qdrant — BM25 sparse search combined with dense vectors — followed by cross-encoder reranking.
  • Confidence-based routing: low-confidence answers are routed differently instead of being returned as if certain.
  • Human-in-the-loop validation for AI-based medical-image results.
  • Input and output guardrails to keep answers safe and on topic.

Result

Compared with searching the sources manually, the assistant cut response time by about 40%.

A research and learning project — not a medical device, and not for clinical use.

Multi-agentRAGLangGraphComputer vision

Working on something similar?

I'm open to senior and staff roles and to select advisory engagements.