Content Understanding custom analyzer builds stuck in "Creating" indefinitely

Question

Content Understanding custom analyzer builds stuck in "Creating" indefinitely

Tyler Broudou 0

Custom Content Understanding analyzers that use a labeled-data knowledge source (in-context learning) never finish building. They remain in status: creating indefinitely (multiple analyzers stuck for 1.5+ hours and counting). Content Understanding Studio surfaces repeated "Build failed — Network Error" notifications during these builds. Analyzers without a labeled-data knowledge source build successfully and reach Ready normally.

Model configuration — Analyzer models.completion = gpt-5.2, models.embedding = text-embedding-3-large. Both are in the analyzer's own supportedModels list. Both deployments exist on the resource and show Succeeded.

Studio/session — Hard refresh, re-auth, and incognito/clean-profile retries do not change the stuck status.
Storage networking — The storage account holding the labeled training data has Public network access: Enabled from all networks (no firewall/NSP restriction).
Storage RBAC — The Foundry resource's managed identity has Storage Blob Data Contributor on the storage account (read/write/list).
Training data presence — The labeling project's train prefix contains the expected, well-formed labeled dataset (source blobs + .labels.json + .result.json for each sample).
Build operation status — Querying the analyzer via GET returns only status: creating with an empty warnings array and no error object. The build operation does not surface any error message.

0 comments

1 answer

Your answer

Answer 1

Hi Tyler Broudou,

Welcome to Microsoft Q&A,

In-context learning with a labeledData knowledge source is only supported when the base analyzer uses the Document Analysis template. An accepted community thread confirms that the Knowledge tab and labeled-data pipeline are not wired up for other templates. If your analyzer uses a different base (e.g., a custom or invoice template), the API will accept the knowledgeSources config without error, but the build pipeline silently fails, which matches exactly what you're seeing.

Verify your analyzer's baseAnalyzerId is prebuilt-document. If not, recreate the analyzer using Document Analysis as the base template and re-attach the labeled-data knowledge source.

Try substituting gpt-4.1 or gpt-4o as models.completion and rebuild. If the labeled-data build succeeds with an older model, that confirms a gpt-5.2 compatibility gap with the in-context learning pipeline.

Please upvote and accept the answer if it helps!!