Skip to content

FAQ

Does it need an API key or network access?

No. The default layout extractor and all matching rules run offline. Only the optional llm extractor calls a server, and you can point it at a local one such as Ollama or vLLM.

How accurate is it on my invoices?

The eval numbers measure the five layouts in the shipped dataset, which the layout extractor was built against. They catch regressions; they don't predict accuracy on new vendors. Add your own invoices to a dataset directory with a manifest.json and ground truth, then run invoice-agent eval --dataset DIR. For layouts the rules don't cover, try --extractor llm.

What happens with a scanned invoice?

A page without a text layer gets NO_TEXT_LAYER and the invoice goes to review instead of being guessed. Install the [ocr] extra and tesseract, then pass --ocr.

Why did an invoice I processed twice get rejected?

The ledger remembers every invoice it has seen. Processing the same invoice again is a DUPLICATE_INVOICE. Use a fresh ledger (or restart the service) to start over.

Can I change the tolerances?

Yes. Write a JSON file with the settings you want to change and pass --policy FILE. See Policy.

Which PO and receipt formats work?

JSON, CSV, and SQLite, picked by file extension. See the README.

Is the dataset real?

No. Every vendor, address, and product is fictional, and every page carries a footer that says it is a synthetic test document. invoice-agent generate-dataset rebuilds it byte for byte.

How do I verify a download?

Each release has a SHA256SUMS file and build provenance attestations. Run gh attestation verify invoice-agent-linux-x64 -R superintelligenceco/invoice-agent. The image is signed with cosign keyless signing; the release notes show the cosign verify command.