MaxModel

LLM answers grounded in your own sources — with verbatim citations and hallucinations removed. One deterministic core, a whole toolkit on top: no LLM judging an LLM.

MaxModel is the developer layer of the maxmodel.com gateway. Same key, same OpenAI-compatible API, one new capability: verified output. You pass your own sources (docs, wiki pages, retrieved RAG chunks); the API returns the model’s answer with verbatim citations and a list of any claim it could not ground — stripped out before it ever reaches your users. Because that check is a string match, everything built on it — eval, extraction, a CI gate — is deterministic too.

npm i maxmodel      # TypeScript
pip install maxmodel  # Python

Why it’s different

Most “citation” and faithfulness tools ask another model to grade the answer — an LLM judge, an NLI/embedding score, a fine-tuned classifier. Those are probabilistic: they can be wrong, and they give a different answer on a re-run.

MaxModel’s grounding check is deterministic code — your groundedness signal is a string match, not an opinion. Every cited quote must appear character-for-character in the source you supplied, or the claim is dropped. No quote, no claim.

This makes a specific failure structurally impossible: a model can’t “post-rationalize” a citation — attach a real-looking source to a claim the source doesn’t actually contain. (Research finds up to 57% of RAG citations are correct-but-not-faithful; verbatim matching rejects all of them.) The result is machine-verifiable — reproducible by any third party using only your source text and the output, no model weights needed.

import { MaxModel } from 'maxmodel'
 
const mx = new MaxModel({ apiKey: process.env.MAXMODEL_KEY! })
 
const out = await mx.verified.create({
  model: 'gpt-5.5-pro',
  messages: [{ role: 'user', content: 'What is our refund policy?' }],
  sources: [
    { id: 'refunds.md', text: 'Full refunds within 30 days. No refunds after 30 days.' },
  ],
})
 
console.log(out.text)
// "Full refunds are available within 30 days [refunds.md]. ..."
out.citations   // each quote is a verbatim substring of refunds.md
out.unsupported // claims that could not be grounded — caught, not shipped

One core, a whole toolkit

The same verbatim check is the floor under everything — so every piece is deterministic and re-runnable, none of it an LLM judging an LLM:

	What it does
`verified.create`	Grounded answers with verbatim citations + char offsets; ungrounded claims dropped.
`verified.eval` ⭐	Score groundedness over a dataset with the verbatim check — same number every run, unlike RAGAS-style LLM judges.
`verified.extract`	Per-field JSON where each value is a verbatim substring of a source, or `null`.
`checkNumbers: 'row'`	A number must sit in the table cell whose row + column the claim names — catches “right number, wrong row.”
CI quality gate	A GitHub Action that fails a PR when groundedness drops below a threshold.
Audit-grade eval	Reproducible by a third party from source text + output alone — defensible for legal / finance / healthcare.
Integrations	Vercel AI SDK middleware (`@maxmodel/vercel-provider`) + a one-line OpenAI `baseURL` swap.

Start here

Quickstart — install and make your first verified call in two minutes.
How verification works — the deterministic moat, explained honestly.
Migrating from OpenAI — change the import, add sources.
Python SDK — pip install maxmodel, same API in snake_case.
Errors — typed errors and the HTTP table.

What MaxModel is not (v1)

It does not retrieve for you — you pass the sources (post-retrieval chunks).
/v1/verified is non-streaming: verification needs the full claim set before it can answer.
No GUI. It’s an SDK + API for developers.

Quickstart