Azure Mistral OCR API does not support confidence_scores_granularity parameter

Question

I am currently integrating the Azure-hosted Mistral OCR API

I would like to retrieve OCR confidence scores (especially word-level confidence), similar to the native Mistral OCR API functionality.

However, when sending the following parameter:

{
  "confidence_scores_granularity": "word"
}

the API returns HTTP 422 with this error:

{
  "error": {
    "code": "Invalid input",
    "message": "{\"detail\":[{\"type\":\"extra_forbidden\",\"loc\":[\"body\",\"confidence_scores_granularity\"],\"msg\":\"Extra inputs are not permitted\",\"input\":\"word\"}]}",
    "status": 422
  }
}

Environment details:

• Region: Japan East
• Endpoint type: Azure-hosted Mistral OCR

Questions:

1. Does Azure Mistral OCR currently support OCR confidence scores?
2. Is there another Azure-compatible parameter or API version for retrieving word/page confidence values in Mistral?

Our use case requires OCR confidence values to validate scanned PDF quality and detect low-confidence OCR regions.

Thank you for your support.

Answer 1

Vu Thi Mai Linh(TMI)

Azure-hosted Mistral OCR currently does not support the confidence_scores_granularity parameter.

The HTTP 422 (extra_forbidden) error confirms the parameter is not part of the Azure-exposed API schema.

There is no alternative Azure parameter today to retrieve word/page confidence scores from Mistral OCR.

🔍 What’s might be happening

1. Native Mistral vs Azure-hosted Mistral (key difference)

The native Mistral OCR API supports:

JSON"confidence_scores_granularity": "word" | "page"Show more lines

→ returns confidence scores per word/page [docs.mistral.ai]

However, in Azure AI Foundry (Azure-hosted Mistral):

The request schema is restricted

Unsupported fields are rejected with:

extra_forbidden → Extra inputs are not permitted

Azure exposes a subset of the upstream Mistral API, and confidence_scores_granularity is currently not included.

Does Azure Mistral OCR return confidence scores?

Based on available documentation and behavior:

Azure Mistral OCR:

✅ Returns extracted content (markdown/text/structured output)

❌ Does NOT return confidence scores (word/page)

There is no documented field in Azure Mistral responses for confidence values.

⚠️ Important implication for your use case

“validate scanned PDF quality and detect low-confidence regions”

This cannot currently be implemented using Azure Mistral OCR alone.

✅ Recommended alternatives (Azure-supported)

Since you’re doing quality validation / low-confidence detection, here are practical workarounds used in real customer scenarios:

Option 1 — Use Azure Document Intelligence

Azure Document Intelligence provides:

✅ Word-level confidence scores

✅ Page-level / field-level confidence

✅ Production-grade OCR + structured extraction

Microsoft explicitly documents:

“Document Intelligence returns confidence for predicted words… between 0 and 1” [learn.microsoft.com]

👉 Best fit for:

OCR quality validation

Threshold-based filtering (e.g., reject < 0.8 confidence)

Compliance / human-in-the-loop workflows

Option 2 — Dual-pass pipeline (common workaround)

If you must use Mistral OCR for layout/quality:

Pattern

Run Mistral OCR

Get high-quality markdown + structure

Run Document Intelligence (Read OCR)

Extract confidence scores

Align results

Map words/regions between outputs

Use Document Intelligence confidence as proxy

Option 3 — Native Mistral (non-Azure endpoint)

Native Mistral API supports the parameter, You can call them via Custom functions

Hope it helps.

Thank you.