AIFondry-ContentUnderstanding

Content Understanding - Model Explanation

Content Understanding requires three AI model deployments to power its document processing pipeline. Each model serves a distinct role.

gpt-4.1 (Completion Model)

This is the brain of Content Understanding. When you define fields with generate or classify methods, the service sends the extracted document content to this LLM to:

• Generate summaries, insights, and inferred values
• Classify documents into categories (invoice, receipt, contract, etc.)
• Interpret complex field extraction where literal text matching isn’t enough

It’s the most capable model in the pipeline and handles nuanced reasoning about document content.

Note: Microsoft now recommends gpt-5.2 as the completion model. The GPT-4.1 family retires October 2026, but gpt-4.1 works fine for demos today.

gpt-4.1-mini (Lightweight Completion)

Used for faster, cheaper processing on simpler tasks within the pipeline:

• Field confidence estimation
• Source grounding (mapping extracted values back to their location in the document)
• Simpler extraction tasks that don’t need the full gpt-4.1

Think of it as the “fast path.” Content Understanding routes simpler subtasks here to reduce latency and cost.

text-embedding-3-large (Embedding Model)

This converts text into vector representations (embeddings). Used for:

• Semantic matching - finding where in the document a field value came from
• Field-to-content alignment - matching your field schema descriptions to relevant sections of the document
• Contextualization - the service tokenizes each page into approximately 1,000 tokens and uses embeddings to identify which parts are relevant to each field

How They Work Together

Document --> OCR/Layout --> Markdown
                              |
                    text-embedding-3-large (find relevant sections)
                              |
                    gpt-4.1-mini (simple extractions, confidence scores)
                              |
                    gpt-4.1 (generate summaries, classify, complex fields)
                              |
                    Structured Output + Confidence Scores

Cost Implications

• You pay for the Content Understanding service (per-page extraction) plus the Azure OpenAI model usage for field extraction
• Using gpt-4.1-mini for simpler tasks keeps costs down compared to routing everything through gpt-4.1
• Contextualization is billed at approximately 1,000 tokens per page/image, 100,000 per hour of audio, and 1,000,000 per hour of video
• There is no monthly subscription. Pricing is pure pay-as-you-go

Resources

• Content Understanding Overview
• Analyzer Reference (Field Methods)
• Service Limits
• Pricing

This site is open source. Improve this page.