How do I extract fields from documents with an LLM API?

To extract fields from documents with an LLM API, the usual pattern is:

1. Convert the document to text
2. Define the fields you want
3. Prompt the model to return structured output
4. Parse and validate the result
5. Handle edge cases and retries

1) Extract text from the document

If your input is:

• PDF / Word / image: use OCR or document parsing first
• plain text / HTML / JSON: you may be able to send it directly

Common tools:

• PDFs: pymupdf, pdfplumber, unstructured
• Images: OCR like tesseract, cloud OCR, or a vision-capable model
• Scanned docs: OCR first, then LLM

2) Decide the schema

Be explicit about what fields you want, for example:

{
  "invoice_number": "",
  "date": "",
  "vendor_name": "",
  "total_amount": "",
  "currency": ""
}

3) Prompt the LLM for structured output

Ask for only JSON and specify the schema.

Example prompt:

Extract the following fields from the document text:

• invoice_number
• date
• vendor_name
• total_amount
• currency
  Return valid JSON only. If a field is missing, use null.

4) Example API call

Here’s a generic example in Python:

import requests
import json

document_text = open("doc.txt").read()

prompt = f"""
Extract these fields from the document:
- invoice_number
- date
- vendor_name
- total_amount
- currency

Return valid JSON only. If missing, use null.

Document:
{document_text}
"""

response = requests.post(
    "https://api.your-llm.com/v1/chat/completions",
    headers={"Authorization": "Bearer YOUR_API_KEY"},
    json={
        "model": "your-model",
        "messages": [
            {"role": "system", "content": "You extract structured data from documents."},
            {"role": "user", "content": prompt}
        ],
        "temperature": 0
    }
)

print(response.json())

5) Prefer schema-constrained output if available

If your API supports structured outputs, function calling, or JSON schema mode, use that. It reduces parsing errors a lot.

Example idea:

• Define a JSON schema for the fields
• The model must fill it
• Your code validates the result

6) Validate the output

Check:

• JSON parses correctly
• Required fields exist
• Dates are in a standard format
• Amounts are numeric
• Confidence or source text is reasonable

If validation fails, retry with a stricter prompt or fallback logic.

7) For better accuracy

A few practical tips:

• Use temperature=0
• Keep the prompt short and specific
• Provide field definitions and examples
• Include a few sample documents + expected outputs if possible
• Chunk long documents and merge results if needed
• Ask the model to cite the exact text span it used

Example with source evidence:

{
  "invoice_number": "INV-12345",
  "invoice_number_source": "Invoice #: INV-12345"
}

8) If you have many document types

Use a pipeline:

• classify document type
• choose a schema for that type
• extract fields
• validate against rules

---

If you want, I can give you:

1. a Python example using OpenAI-style structured outputs,
2. a JSON schema template for document extraction, or
3. a full invoice/receipt extraction pipeline.