30 km — the wall

Retrieval-augmented, cited, free

Marathon Strategy RAG

A marathon coach that cites its sources. Ask a race-strategy question and get practical advice grounded in real marathon science — and in my own data analyses of 30,000+ Boston finishers.

By Jeremy Lee · 14 marathons, 2:48 PB

You
Should I try to negative-split my first marathon?
Coach
Probably not as a goal. In 31,912 Boston finishers only 2.5% actually ran a negative split, and ~94% slowed — median +14% [1]. Aim for an even effort instead: hold back slightly through halfway so you're not part of the 27% who fall apart after 30 km [2].

Why a citing coach?

Running advice is everywhere and a plain chatbot will invent statistics with total confidence. This coach answers only from a curated knowledge base and cites every claim with a [n] that traces back to a real source — established training science, or one of my own published marathon analyses.

📊

Grounded in real data

Cards distilled from my analyses of the Boston field — negative splits, the wall, the heat tax, course difficulty, super-shoes.

🔎

Every claim cited

Answers carry [n] citations back to the exact knowledge card — no invented numbers.

🆓

Free & local

Embeddings and search run on your machine with sentence-transformers + Chroma. The knowledge base ships in the repo.

🧩

Easily extended

Add a markdown card, re-index, and the coach can cite it — no retraining.

How it works

  1. Your question is embedded locally and matched against the knowledge cards.
  2. The most relevant passages are retrieved, each keeping its title and source.
  3. Claude answers using only those passages, citing each fact it uses.
  4. You see the advice plus the exact cards it was built from.
Under the hood: generation runs on the Claude CLI by default (uses your Claude subscription, no per-token cost); set ANTHROPIC_API_KEY to use the API instead. Retrieval is always local and free.
Get it on GitHub Read the setup guide