A hosted API¶

The best quality, and the only option here that costs money. Read Without a key first if you have not: for most people the assistant route is better and free.

Anthropic¶

uv pip install -e ".[llm]"
export ANTHROPIC_API_KEY=sk-ant-...
uv run mimem doctor --live

uv run mimem build paper.pdf --out out/paper --llm --provider anthropic

--provider anthropic is the default when you pass --llm, so it can be left off. --model picks a different one; the default is the capable model rather than the cheap one, deliberately — downgrading should be a decision you measure, not one the tool makes for you.

OpenAI, and everything shaped like it¶

export OPENAI_API_KEY=sk-...
uv run mimem build paper.pdf --out out/paper --llm --provider openai --model gpt-4o-mini

The same adapter reaches anything serving POST /chat/completions, which is most of the market. Point --base-url at it:

uv run mimem build paper.pdf --llm --provider openai \
  --base-url https://openrouter.ai/api/v1 --model anthropic/claude-3.5-sonnet

The key comes from OPENAI_API_KEY whatever the host, so export the one that provider issued.

Knowing the cost before you pay it¶

uv run mimem elaborate out/paper/doc.ir.json --dry-run

prints every planned call and an estimate, and makes none of them. Then cap the real run:

uv run mimem build paper.pdf --llm --budget 2.00

--budget is a hard stop, not a warning: the run ends and keeps the artefacts it had already produced.

Two things make a paper cheaper than the call count suggests. The source document is sent as a cached prefix, so the second task on a document is a fraction of the first. And answers are cached by content, so re-running after an edit re-asks only about what changed.

Record once, replay for ever¶

uv run mimem record paper.pdf --out fixtures/paper --provider openai
uv run mimem build paper.pdf --fixtures fixtures/paper

The second command is free and gives byte-identical elaborations. Worth doing on any document you will build more than twice, and the only honest way to compare two changes to the rest of the pipeline — otherwise the model moves underneath the comparison.

What is checked before anything reaches you¶

Every generated sentence goes through the grounding gate:

• numbers — every digit-bearing token must appear in the source sentences it was written from;
• years — a wrong one is impossible to catch by ear;
• directions — if the model says a quantity increases and the source says it decreases, the claim is rejected. This is the check the rule exists for, and the only one that catches a failure which reads perfectly;
• names — a capitalised word the source never uses, reported as a warning.

Rejections are not failures. The task takes its degradation path, the manifest records it, and the elaboration report tells you how many were rejected and why. A run with rejections is the system working.

When it goes wrong¶