Azure AI Search: Why is OCR Reprocessing All Pages on Incremental Update?

Hello,
I'm sorry but the provided insights are not helping. It really looks like an AI generated answer.

I'd like to have somebody looking into the technical details, instead of vague answers pointing to the documentation that I already looked into a couple of times.

Kind regards

Hi @mathias Herbaux ,

Sorry for the late Response.

• Go to Azure Cognitive Search in the Azure Portal.
• Under Indexers, select your indexer and ensure Incremental Enrichment is enabled.
• Check if your skills are configured correctly.
• Ensure that the OCR skill is applied only when necessary.
• Make sure that Document Cracking or other skills aren’t triggering reprocessing of all pages unnecessarily.
• In the Indexers section, after running an incremental update, check "Logs" to see what skills are being triggered.
• If OCR is running on all pages, it will appear in the logs.
• Update the metadata of the blob and trigger an incremental update.
• If OCR processes all pages despite minimal changes, check if metadata changes are causing a full reindex. Incremental Enrichment in Azure Cognitive Search Create an indexer in Azure Cognitive Search This approach should help you that why OCR is being re-triggered on all pages.

Don't know what to say. I understand that Azure Cognitive Search was the previous name of Azure AI Search, but event the links provided are wrong, pointing to 404 web pages.

FYI: I read those links

• https://learn.microsoft.com/en-us/azure/search/search-howto-incremental-index?tabs=portal
• https://learn.microsoft.com/en-us/azure/search/cognitive-search-incremental-indexing-conceptual

The enableReprocessing is set to True.

The incremental enrichment cache is set, as I mentioned in my original question:

I checked in the cached data, and I've found that for this document I have 262 images in the binary folder.
I might not have be clear enough:
I'm referring to the container created by the incremental enrichment cache, where we can find the documents processed. I noticed that a document after being indexed for the first time had 131 images in the binary folder, corresponding to all images present in the PDF. After updating one single metadata of this PDF and running the indexer again, the binary folder had 262 images, meaning that even the PDF hadn't changed, it had been fully reprocessed.

Hi @mathias Herbaux ,
Thank you for clarifying.
It seems like the root of the issue is related to how the incremental enrichment cache handles changes. Even with metadata updates, the behavior indicates that the cache might be invalidating or not functioning as expected for your scenario.
Validate Incremental Enrichment Cache:

Ensure the cache is enabled and correctly configured in your skillset JSON

{
  "enrichmentCache": {
    "storageConnectionString": "<YourBlobStorageConnectionString>",
    "containerName": "<YourCacheContainerName>"
  }
}

Disable Reprocessing:

Set enableReprocessing to false for all skills except those requiring reprocessing (like OCR).

Add a Custom ShouldProcess Skill:

Implement a custom skill to evaluate whether OCR should run based on metadata changes.

{
  "@odata.type": "#Microsoft.Skills.Text.ConditionalSkill",
  "inputs": [
    {
      "name": "condition",
      "source": "/document/metadata/lastModified"
    }
  ],
  "outputs": [
    {
      "name": "output",
      "targetName": "shouldProcess"
    }
  ]
}

Inspect Indexer Logs:

Check Indexer Run Details for cache misses or triggers causing full reprocessing.
Go to Azure Portal > Indexer > Run Details > Logs.

Modify Metadata Handling:

Ensure metadata updates do not affect the fields mapped to skills unnecessarily, avoiding cache invalidation.

Test with Minimal Skills:

Temporarily simplify the skillset to isolate the issue.

Humm,
In my understanding, enableReprocessing is a property only available on the cache, not a per skill property.
I've already tried to conditionally activate OCR, but it does not seem to work because:
It takes only one Input that comes from context: image (https://learn.microsoft.com/en-us/azure/search/cognitive-search-skill-ocr), moreover as mentioned in the documentation : Currently only works with "/document/normalized_images" field. Hence the fact we can't add a condition to perform OCR.
What are your thoughts about that?

Hi @mathias Herbaux ,
Thankyou for your response.
Yes you are correct. The enableReprocessing property is available only at the cache level, not on a per-skill basis.

Regarding conditionally activating OCR:

• The OCR skill only accepts /document/normalized_images as input, meaning you cannot add a condition directly within the OCR skill to control its execution.
• Conditional activation of OCR isn’t feasible due to this input constraint. The pipeline will process all images if /document/normalized_images exists, regardless of metadata changes.
• If metadata changes trigger the indexer, the OCR skill will reprocess all images because it relies on normalized images from the pipeline, and no inherent mechanism exists to skip this step based on metadata alone.

Does it means that there is no possible solution? Sounds weird since this is the exact purpose of the Incremental Enrichment caching feature, to cache skills output so that they are only reexecuted if the input(s) has changed.

Is there somebody technical on your side that could look into it?