Series · 11 Parts

MedGemma CKD — a clinical RAG assistant, three ways

I built the same chronic-kidney-disease assistant at three levels — simple, agentic, multi-agent — on 24 real clinical guideline documents, for the Kaggle MedGemma Impact Challenge. This series walks the whole thing: the data pipeline, each level, and an honest evaluation.

  1. MedGemma CKD: building a clinical RAG assistant three ways
    Part 1

    MedGemma CKD: building a clinical RAG assistant three ways

    Why I built the same chronic-kidney-disease assistant at three levels — simple, agentic, multi-agent — and the shared foundation underneath. Part 1 of the MedGemma CKD series.
  2. See it answer — grounded, cited, and willing to say "no"
    Part 2

    See it answer — grounded, cited, and willing to say "no"

    Real questions through the CKD assistant — grounded, cited answers with their sources, and a clean refusal when the guidelines are silent. A demo from the MedGemma CKD series.
  3. From 24 clinical PDFs to a vector store
    Part 3

    From 24 clinical PDFs to a vector store

    The unglamorous half of RAG: OCR with Docling, section splitting, block-aware chunking, and embedding 24 clinical guideline documents into ChromaDB. Part 2 of the MedGemma CKD series.
  4. Chunking clinical guidelines — where recall is won or lost
    Part 4

    Chunking clinical guidelines — where recall is won or lost

    Recall was the ceiling — and recall is decided at chunk time. Docling OCR, section-first splitting, 2000-token chunks, the 512-token embedding catch, and a duplicate corpus. A MedGemma CKD deep-dive.
  5. Level 1 — Simple RAG, the honest baseline
    Part 5

    Level 1 — Simple RAG, the honest baseline

    Retrieve, answer, cite — and the retriever choices (flat, tree, RAPTOR, contextual) that decide whether any of it works. Part 3 of the MedGemma CKD series.
  6. The retriever zoo — three bets scored, two that didn't make it
    Part 6

    The retriever zoo — three bets scored, two that didn't make it

    Flat vs tree vs RAPTOR vs contextual vs hybrid — how each retriever works, a RAGAS scoreboard where RAPTOR wins, and the tree retriever that was silently broken for months. A MedGemma CKD deep-dive.
  7. Where the model runs — local, hosted, or a GPU box
    Part 7

    Where the model runs — local, hosted, or a GPU box

    Local embeddings, a provider-agnostic generator endpoint, and three places to run it — a laptop, AI Studio, and a self-hosted g6/vLLM GPU box — with the honest tradeoffs of each. A MedGemma CKD deep-dive.
  8. The knobs that move the numbers (and the ones that don't)
    Part 8

    The knobs that move the numbers (and the ones that don't)

    Top-k, similarity threshold, chunk size, embedding dimensions, RAPTOR and tree internals, generation settings — which parameters actually move the score, and which are set-and-forget. A MedGemma CKD deep-dive.
  9. Level 2 — Agentic RAG with LangGraph
    Part 9

    Level 2 — Agentic RAG with LangGraph

    Wrapping retrieval in a LangGraph state machine: PII redaction with Presidio, intent routing, and self-scoring with RAGAS. Part 4 of the MedGemma CKD series.
  10. Level 3 — A multi-agent clinical assistant
    Part 10

    Level 3 — A multi-agent clinical assistant

    A router and four specialists — Diet, Medication, Lifestyle, Knowledge — because a diet question and a drug-interaction question are different jobs. Part 5 of the MedGemma CKD series.
  11. Evaluating a clinical RAG — the honest results
    Part 11

    Evaluating a clinical RAG — the honest results

    RAGAS, a retriever comparison, and middling scores I'm not going to dress up. Why retrieval over dense clinical text is the hard part. Part 6 of the MedGemma CKD series.
← All blog posts