Document Classification

Classify documents (PDF, images, Word) with Gemma 3 27B. Under the hood, it uses structured outputs to match the labels given.

• Endpoint
• Request
  • Supported document types
• Process flow
• Response
• Default labels
• API Key Setup
• Examples
  • Quick Test with Sample PDF
  • Converting PDF to Base64
  • TypeScript (Node)
  • Python
• API Behavior and Limits
  • Performance Expectations & Service Limits
  • Key Limitations
  • Error Handling
  • Error Response Format
• Notes & best practices

Note This endpoint does not use the OpenAI SDK. Call it directly with HTTPS.

Endpoint

• Method: POST
• Path: https://api.inference.net/classify/document

Request

{
  "document": "base64_encoded_document_string",
  "fileType": "pdf",
  "additionalLabels": ["Custom Label 1", "Custom Label 2"]
}

• document: Required. Base64-encoded contents of the file. Data URLs are also supported; the API strips any prefix like data:application/pdf;base64, automatically.
• fileType: Required. File type of the document ("pdf" or "docx").
• additionalLabels: Optional. Array of extra labels to merge with the default set.

Note, currently this endpoint only accepts pdf and docx files.

Supported document types

• PDFs (application/pdf) - specify "fileType": "pdf"
• Office documents (DOCX) - specify "fileType": "docx"

If the document cannot be parsed or converted to an image, the endpoint returns a 400 error.

Process flow

When you submit a Base64 document, the API:

1. Receives document upload as Base64
2. Converts document pages to one or more images
3. Sends images and combined labels (default + additional) to Gemma 3 27B
4. Parses model output and returns matched labels

Response

The API returns a JSON object with the following structure:

{
  "labels": ["Memo", "Email Correspondence"],
  "metadata": {
    "processingTimeMs": 6748,
    "pageCount": 4,
    "pdfSizeMB": 0.1361713409423828,
    "pagesProcessed": 4,
    "correlationId": "1758222412597-eg945o7hq",
    "timestamp": "2025-09-18T19:06:59.345Z"
  }
}

• labels: Array of strings containing the matched document classification labels from the predefined set
• metadata: Object containing processing information:
  • processingTimeMs: Time taken to process the document in milliseconds
  • pageCount: Total number of pages in the PDF
  • pdfSizeMB: Size of the PDF file in megabytes
  • pagesProcessed: Number of pages that were successfully processed
  • correlationId: Unique identifier for the request
  • timestamp: ISO 8601 timestamp when the processing completed

Default labels

The following labels are included by default. You can extend or narrow this set using additionalLabels.

Invoice
Purchase Order
Receipt
Expense Report
Budget Report
Financial Statement
Balance Sheet
Income Statement
Cash Flow Statement
Tax Return
W2 Form
1099 Form
Pay Stub
Credit Note
Debit Note
Contract
Service Level Agreement(SLA)
Non - Disclosure Agreement(NDA)
Terms of Service
Privacy Policy
Memorandum of Understanding(MOU)
Letter of Intent(LOI)
Power of Attorney
Legal Notice
Compliance Report
Audit Report
Risk Assessment
Statement of Work(SOW)
Request for Proposal(RFP)
Request for Information(RFI)
Request for Quotation(RFQ)
Business Plan
Strategic Plan
Operational Procedure
Standard Operating Procedure(SOP)
Work Order
Change Request
Incident Report
Product Requirements Document(PRD)
Technical Specification
Architecture Document
API Documentation
User Manual
Installation Guide
Release Notes
Bug Report
Test Plan
Code Review
Database Schema
Network Diagram
Mermaid Diagram
UML Diagram
Project Charter
Project Plan
Gantt Chart
Status Report
Meeting Minutes
Action Items List
Risk Register
Stakeholder Analysis
Lessons Learned
Post - Mortem Report
Resume / CV
Job Description
Offer Letter
Employment Contract
Performance Review
Training Material
Employee Handbook
Organizational Chart
Leave Request
Timesheet
Benefits Summary
Marketing Plan
Campaign Brief
Brand Guidelines
Press Release
Case Study
White Paper
Product Brochure
Sales Proposal
Customer Testimonial
Market Research Report
Competitor Analysis
Sales Forecast
Email Correspondence
Memo
Newsletter
Presentation
Meeting Agenda
Conference Notes
Webinar Materials
FAQ Document
Quality Assurance Report
ISO Certification
Certificate of Compliance
Inspection Report
Warranty Document
Service Agreement

API Key Setup

To use this API, you need an Inference.net API key. You can get one by:

1. Visiting inference.net
2. Signing up for an account
3. Generating an API key in your dashboard

Examples

Set your API key first:

export INFERENCE_API_KEY=<your-inference-api-key>

Prepare a Base64 file (macOS/Linux):

base64 -i ./examples/sample.pdf -o ./examples/sample.b64

Quick Test with Sample PDF

This curl requests takes in a sample invoice that we have already converted to base64.

# Test with additional labels
curl -X POST https://api.inference.net/classify/document \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $INFERENCE_API_KEY" \
  -d "{\"document\": \"$(cat examples/sample.b64)\", \"fileType\": \"pdf\", \"additionalLabels\": [\"Receipt\", \"Bill\"]}" | jq .

Converting PDF to Base64

# macOS/Linux
base64 -i document.pdf -o document.b64

TypeScript (Node)

import * as fs from "fs";

const fileBase64 = fs.readFileSync("examples/sample.pdf", { encoding: "base64" });

const response = await fetch("https://api.inference.net/classify/document", {
  method: "POST",
  headers: {
    "Authorization": "Bearer YOUR_INFERENCE_API_KEY",
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    document: fileBase64,
    fileType: "pdf",
    additionalLabels: ["Invoice", "Contract"],
  }),
});

const result = await response.json();
console.log(result.labels, result.metadata);

Python

import base64
import requests

with open("examples/sample.pdf", "rb") as f:
    file_b64 = base64.b64encode(f.read()).decode("utf-8")

response = requests.post(
    "https://api.inference.net/classify/document",
    json={
        "document": file_b64,
        "fileType": "pdf",
        "additionalLabels": ["Invoice", "Contract"]
    },
    headers={
        "Authorization": "Bearer YOUR_INFERENCE_API_KEY",
        "Content-Type": "application/json",
    }
)

result = response.json()
print(result["labels"], result["metadata"])

API Behavior and Limits

Performance Expectations & Service Limits

Note This API is is not yet ready for production workloads. It is intended for evaluation and development purposes only.

API Timeouts:

• since we only use the first 5 pages anyway for this endpoint, it's recommended you only submit 5 pages of your documents to avoid failures.

Key Limitations

Page Processing Limit:

• The API will only analyze the first 5 pages of any PDF
• If your document is longer, content beyond page 5 will be ignored, which may affect classification accuracy

File Size Limit:

• The maximum allowed file size is 10MB
• Requests with larger documents will be rejected immediately with a DOCUMENT_TOO_LARGE error

Unsupported Content:

• The API cannot process password-protected or encrypted PDFs
• Corrupted or non-standard PDF files may fail during conversion

Error Handling

The API returns structured error responses with specific error codes. Here's a comprehensive guide:

Error Code	HTTP Status	Common Cause	Recommended Action for User
UNAUTHORIZED	401	The Authorization header is missing, malformed, or contains an invalid key	Ensure you are sending a Bearer <YOUR_INFERENCENET_API_KEY> header with a valid key
DOCUMENT_TOO_LARGE	413	The submitted document exceeds the 10MB file size limit	Reduce the file size of your document before encoding it to base64. Consider compressing images within the PDF if possible
INVALID_DOCUMENT_FORMAT	415	The base64 string does not represent a valid PDF file	Verify that the document is a valid PDF and that the base64 encoding is correct. The API does not support encrypted or password-protected PDFs
EMPTY_PDF	422	The PDF was successfully processed but found to contain zero pages	Check your source document to ensure it is not empty or corrupted
PDF_CONVERSION_FAILED	502	The external PDF conversion service (PDF.co) could not process the file	This usually indicates a corrupted or non-standard PDF. Try re-saving the PDF from a trusted source and submit it again
CLASSIFICATION_FAILED	502	The AI model failed to analyze the images and return a classification	This may be a temporary issue with the AI service. Please retry your request. If the problem persists, the document may be unclassifiable (e.g., blank pages)
INTERNAL_ERROR	500	An unexpected error occurred on the server	This is a server-side issue. Please wait a moment and retry your request. Use the correlationId from the response if you need to report the issue

Error Response Format

All errors return a consistent JSON structure:

{
  "error": "Human-readable error message",
  "code": "ERROR_CODE"
}

Notes & best practices

• Keep label lists concise and mutually exclusive when possible
• Merge defaults + additionalLabels to extend coverage without losing core categories
• Monitor the processingTimeMs field to optimize your usage patterns