feat: mistral ocr converter

[Hansehart]

Related Issues

• part of Build MistralOCRDocumentConverter similar to AzureOCRDocumentConverter #2341

Proposed Changes:

This PR adds a new OCR document converter component for the Mistral integration in Haystack.
The MistralOCRDocumentConverter uses Mistral’s Document AI / OCR API to extract text and structured annotations from documents and images.

Key features:

• Automatically converts PDFs output into a single Haystack Document object using OCR for hard parsable formats.
• Supports multiple document input types (DocumentURLChunk, FileChunk (La Plateforme), ImageURLChunk).
• Optionally enriches image regions with structured annotations via Pydantic schemas.
• Supports document-level annotations (e.g. language detection, extracted URLs, any other definied metadata).

How did you test it?

• Manual verification using real-world PDFs and images (e.g. public arXiv PDFs).
• Confirmed Mistral OCR API call returns structured responses with both bbox and document-level annotations.
• Verified Document creation and metadata enrichment in Haystack.

Notes for the reviewer

There are definetly some things missing. Its works great on my structure, however needs adjustments for haystack. Including:

• Tests
• Docs
• Package Integration e.g in the example i use haystack_integrations.components.converters.mistral.ocr_document_converter which can not be resolved
• Due to using pipe union syntax (| instead Optional[]) its for python>=3.10 but can be adjusted
• Mistral API key required
• The other components are not using mistral sdk. I am unsure if its required to rely only on openai or can I use mistral SDK?
• I developed this component for my own purpose but want to share it. When you provide me some feedback on whats needs to be done, I will tackle that. Functionality is working, for the integrations steps i am going proactively into the pr nowing it need adjustments.

Checklist

• I have read the contributors guidelines and the code of conduct
• I have updated the related issue with new insights and changes
  - I added unit tests and updated the docstrings
• I've used one of the conventional commit types for my PR title: fix:, feat:, build:, chore:, ci:, docs:, style:, refactor:, perf:, test:.

[anakin87]

Thanks for your contribution. I have left some initial comments about the implementation. In general, I would make this component more similar to the AzureOCRDocumentConverter.

Other points you mentioned:

• It is OK to use Mistral SDK
• We want to keep Python 3.9 compatibility for the moment, so let's use Union.
• Once we agree on the main aspects of the implementation, let's add unit and integration tests.

[Hansehart]

I guess I have currently the last issue that I cant run hatch run fmt, because then my

  "type": (
      "haystack_integrations.components.converters.mistral."
      "ocr_document_converter.MistralOCRDocumentConverter"
  ),

Will become: "haystack_integrations.components.converters.mistral.ocr_document_converter.MistralOCRDocumentConverter"

And then three lines exceeding line limit 125>120

[anakin87]

Hey @Hansehart, I've been a bit busy... I'll review it again soon

[Hansehart]

Take your time. I have no deadline on this feature. Instead I realy appreciate your valulable feedback. It makes me really fun to contribute.

[anakin87]

Please update https://github.com/deepset-ai/haystack-core-integrations/blob/main/README.md
"Embedder, Generator" -> "Converter, Embedder, Generator"

---

In general, we are close to merging this PR. 👍

[anakin87]

I ran integration tests locally (they don't run on PRs from forks) and all works well.

Thank you!