Document parsing and data extraction

• Images and scans (vision model)

  gemini-3.8-flashGoogleinput ¥1.5 / output ¥7.5 per 1M tokens· Standard group

  Reads invoice photos, scanned contracts and table screenshots and returns the fields as JSON. Images count as input tokens, and higher resolution means more tokens, so scale images down to the point where text is still clearly readable.

  AlternativesClaude Haiku 5.5in ¥0.4/1Mgpt-5.4-miniin ¥1.5/1Mqwen3-vl-flashin ¥0.196/1M

• Text documents (chat model)

  deepseek-v4.1-flashDeepSeekinput ¥2.8 / output ¥11.2 per 1M tokens· China Models group

  For PDFs with a text layer, Word files and email bodies, extract the text in code and let a text model pull out clauses, dates and amounts. No vision needed, a lower price, and a good fit for long contracts.

  Alternativesqwen3.5-plusin ¥1.12/1Mgpt-5.4-miniin ¥1.5/1M

Recommended models are picked from the live catalog; prices are each model's cheapest group, updated Nov 12, 2026, 8:17 PM. Click a model to see every group's price.

1. 1

   Classify and preprocess

   Identify the document type (invoice, contract, statement). Take the text layer where there is one; convert scans and photos to images at a sensible resolution.

2. 2

   Define the fields

   Write a field list per document type with names, types and formats (dates as YYYY-MM-DD, amounts as numbers) and put it in the prompt.

3. 3

   Call the model

   Images go to the vision model and text to the text model, with temperature 0 and response_format asking for JSON. Some models do not support response_format; ask for JSON only in the prompt and cut the JSON out in code.

4. 4

   Validate in code

   Check required fields and date and amount formats; for invoices, check that line items add up to the total; for contracts, check the order of dates. Retry once on failure.

5. 5

   Human review

   Documents that still fail, or exceed an amount threshold, go to a review queue. Write reviewed results back, and add typical mistakes to the prompt as examples.

Assumptions: 10,000 invoice and scan pages a month at about 2,000 input tokens each (about 1,500 for the image, 500 for instructions) and 400 output tokens; plus 5,000 contracts with a text layer at about 6,000 input and 600 output tokens each.

Calculated from live prices in each model's cheapest group, updated Nov 12, 2026, 8:17 PM.

Model usage only: preprocessing, storage and human review are not included. How many tokens an image becomes depends on the model and resolution; the table assumes about 1,500 per page, and the usage field in each response is what you are billed on. Twice the pages means roughly twice the cost.

Encode the image as base64 in image_url, ask for the listed fields as JSON, then check that every field is present.

import base64
import json
import os
from openai import OpenAI

client = OpenAI(base_url="https://open.tutujin.com/v1", api_key=os.environ["TUTUJIN_API_KEY"])

FIELDS = ["invoice_number", "issue_date", "seller_name", "buyer_name", "total_amount", "currency"]
PROMPT = (
    "Extract these fields from the invoice image and reply with one JSON object only: "
    + ", ".join(FIELDS)
    + ". Use null for anything you cannot read. Dates as YYYY-MM-DD, amounts as numbers."
)


def parse_json(text):
    """Some models add text around the JSON; keep the outermost object."""
    return json.loads(text[text.find("{"): text.rfind("}") + 1])


def extract(path):
    with open(path, "rb") as f:
        data_url = "data:image/png;base64," + base64.b64encode(f.read()).decode()
    resp = client.chat.completions.create(
        model="gemini-3.8-flash",
        messages=[{
            "role": "user",
            "content": [
                {"type": "text", "text": PROMPT},
                {"type": "image_url", "image_url": {"url": data_url}},
            ],
        }],
        response_format={"type": "json_object"},
        temperature=0,
    )
    record = parse_json(resp.choices[0].message.content)
    missing = [k for k in FIELDS if k not in record]
    if missing:
        raise ValueError("missing fields: " + ", ".join(missing))
    return record


print(extract("invoice.png"))
# Text-layer PDFs and Word files: extract the text yourself and send it to deepseek-v4.1-flash with the same field list.