Lagomy

Snap a supplement label and get a structured record of what it says, with UK-sourced evidence for each ingredient and a printable report you can hand to a clinician.

Log, don't advise. Show, link, source — never conclude. Lagomy records what a label says and what published UK sources say. It does not rank products, reassure, diagnose, or tell you what to take. Named for lagom, the Swedish idea of just enough.

Live API: api.lagomy.com/docs

The crew needed a reliable source of UK supplement compositions, and there wasn't one. The available options were US-regime databases, retailer scrapes, or aggregator sites, none of which are trustworthy for UK products.

So the dataset came first: 50 products, 149 ingredients and 952 ingredient rows, transcribed by hand from physical labels and brand-published nutrition tables, with every field traceable to its source. It's published openly on Hugging Face, and everything in this repository is built on top of it.

• Reads a label — Claude Vision transcribes exactly what's printed, flagging anything unclear in a needs_review list rather than guessing.
• Normalises ingredient names — keeps the printed name and adds a canonical one, so "Methylcobalamin" and "Vitamin B12" don't fragment the log.
• Retrieves UK evidence — a CrewAI agent searches NHS, NICE and BNF via Tavily and returns statements with their sources.
• Checks and structures — a second agent verifies each claim against the retrieved evidence, distinguishing what a nutrient does from what a deficiency causes, and emits JSON where every statement carries its source URL, authority and retrieval date.
• Serves it over an API — POST /evidence runs the crew and returns the structured record.

The no-advice rule is tested, not just intended. guardrail_cases.yaml holds adversarial probes in three directions: prompts designed to make the crew rank, reassure or diagnose; a crisis case that must hand off to emergency services; and positive controls that must still surface sourced regulatory facts, so the crew can't pass by refusing everything.

The same phrase checks run on live API responses. If advice-like language appears in the prose, the prose is withheld, and the sourced evidence is returned without it — the record survives, the risky rendering doesn't.

Requires Python 3.12 and uv.

uv sync

Create a .env file with:

ANTHROPIC_API_KEY=your-key
TAVILY_API_KEY=your-key

Run the crew directly:

crewai run

Or serve the API:

uvicorn api:app --reload

Interactive docs at http://127.0.0.1:8000/docs.

Run the guardrail suite:

python run_guardrail_eval.py

Note: each eval run executes the full crew against six cases with live searches, so it takes several minutes and costs API credit. Results are written to eval_results.json.

In active development, built in the open.

Working: the dataset, label reading, ingredient normalisation, evidence retrieval, synthesis, the PDF report, the API, and the guardrail suite.

Not yet built: the front end, search by product name, and subjective tracking over time.

The dataset is a working sample rather than a finished corpus, and is growing.

The code in this repository is MIT licensed — see LICENSE.

The Lagomy UK Supplements dataset is published separately on Hugging Face under CC BY-NC 4.0. The MIT licence above does not cover the dataset.