Hawser

Pull data out of documents and images with AI, and get proof for every value.

You give Hawser a file (a PDF, a scan, a photo) and describe the data you want. An AI model reads the file and fills in your fields, and Hawser checks the answer before you get it:

• Every value comes with proof: the page and the exact spot on the page where it was found.
• No guessing: a value that isn't in the document is marked unknown, never invented.
• A clear verdict: a result is accepted only when every check passes; otherwise it is needs_review and a person can fix and approve it.

Open source (Apache-2.0); runs on your own machine or server; works with Claude, GPT, Gemini and local models.

Install

pip install "hawser-ai[ocr]"

The command and the Python import are both hawser. [ocr] adds a local OCR engine for scans and photos (on Linux it also needs sudo apt-get install libgl1). Python 3.12+.

Try it (no API key needed)

mkdir my-project && cd my-project
hawser init          # creates a sample invoice, its schema and a saved model answer
hawser quickstart    # processes it offline, prints each field with its page and position

Use it in your code

import hawser
from pydantic import BaseModel


class Invoice(BaseModel):
    invoice_number: str
    vendor_name: str
    total: float


result = hawser.run(
    "Extract the invoice.",
    inputs=["invoice.pdf"],
    output_schema=Invoice,
    models="openrouter/anthropic/claude-haiku-5.5",  # needs OPENROUTER_API_KEY
)

if result.status == "accepted":
    print(result.parse_as(Invoice).total)
    for proof in result.evidence_for("/total"):
        print("Found on page", proof.page, "as", repr(proof.text))
else:
    for issue in result.validation.issues:  # status is needs_review, rejected or failed
        print(issue.field, issue.message)

Set ANTHROPIC_API_KEY, OPENROUTER_API_KEY or OPENAI_API_KEY for the provider you use; hawser models list shows every model ID and its price.

More

• Step-by-step guide: https://github.com/hawser-ai/hawser-ai#get-started
• Documentation: https://github.com/hawser-ai/hawser-ai/tree/main/docs
• Measure accuracy on your own documents: hawser eval run (included; see the evaluations guide)
• Server and web app: pip install hawser-server (see hawser-server)
• Changelog: https://github.com/hawser-ai/hawser-ai/blob/main/CHANGELOG.md