Azure Content Understanding OCR is less accurate that ADI

Hello Chris Moseley,

Greetings! Thanks for raising this question in Q&A forum.

This is a great observation and you are not alone in noticing this. The glyph/symbol accuracy difference you are seeing between Azure Document Intelligence (ADI) and Azure Content Understanding (ACU) comes down to how each service's underlying OCR engine is built and optimized.

Azure Document Intelligence uses a specialized, document-optimized OCR model that has been trained and fine-tuned specifically for precise character-level recognition, including special and mathematical symbols like "≤". Azure Content Understanding, on the other hand, is powered by a generative AI and LLM-based approach, which excels at layout understanding, multimodal content, and semantic reasoning but is currently less precise when it comes to visually ambiguous glyphs and special Unicode characters. This is a known tradeoff between the two approaches, and the "≤" being read as "2" is a classic example of a visually similar glyph being misidentified by a generative model.

Here are a few options you can consider:

1. Use Azure Document Intelligence for symbol-heavy content. Since both ADI and ACU now exist under the same Content Understanding umbrella in Azure Foundry Tools, you do not have to fully downgrade. You can continue using the Document Intelligence prebuilt or layout models specifically for documents that contain mathematical symbols, operators, or special glyphs, and use ACU analyzers for other unstructured content where its generative AI strengths shine. Microsoft's own guidance recommends ADI for scenarios where "consistency, low latency, and proven accuracy are the priority."

Try the high-resolution feature in Document Intelligence. If you are processing scanned or image-based PDFs, enabling the highResolution feature in the ADI Read or Layout model can improve accuracy on small or ambiguous characters. You can do this by adding the features parameter to your API call:

POST {endpoint}/documentintelligence/documentModels/prebuilt-layout:analyze?features=ocrHighResolution&api-version=2024-11-30

Check confidence scores. Both services return per-word confidence scores in their JSON output. You can use these to flag low-confidence extractions (for example, those below 0.80) for post-processing review or substitution logic in your pipeline.

Add a post-processing normalization step. For known symbols like "≤", "≥", "±", and others, you can maintain a small lookup or regex substitution layer that corrects commonly misrecognized glyphs in ACU output. This is a lightweight and effective workaround while the LLM-based OCR engine matures.

Submit feedback to Microsoft. Since ACU is a relatively new service (GA since November 2025), the product team is actively improving it. You can share this specific glyph recognition issue through the Azure Feedback portal at: https://feedback.azure.com/d365community/forum/79b1327d-d925-ec11-b6e6-000d3a4f06a4. Concrete examples like "≤ 2 years" being read as "2 2 years" are exactly the kind of feedback that helps the team prioritize OCR model improvements.

If symbol accuracy is critical and you cannot tolerate any misrecognition, stay on Azure Document Intelligence for now. There is no pressure to migrate since Microsoft has confirmed that ADI APIs, endpoints, SDKs, and billing remain fully unchanged and no migration is required for existing workloads.

For more reading on choosing between the two tools, this Microsoft Learn guide is very helpful: https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/choosing-right-ai-tool

And for OCR accuracy characteristics and limitations: https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/computer-vision/ocr-characteristics-and-limitations

If this answer helps you kindly accept the answer which will help others who have similar questions.

Best Regards,

Jerald Felix.