Azure AI Search indexers, skillsets and the knowledge store

A skillset is the chain of AI skills an indexer runs on each document. Skills chunk and vectorise text, read images, translate and extract entities, and custom skills call your own code.

Chunking and vectorisation

• Integrated vectorisation: the Text Split skill chunks, then the Azure OpenAI Embedding skill vectorises each chunk.
• Language detection, entities and key phrases add metadata, not vectors.
• The Merge skill folds OCR text back into the content but doesn’t chunk.

Images and OCR

• Set imageAction: generateNormalizedImages (with dataToExtract: contentAndMetadata). The images land in /document/normalized_images, and the OCR skill uses /document/normalized_images/*.
• OCR reads printed and handwritten text. Image analysis gives tags and captions, and Document Extraction cracks files but doesn’t read pixels.
• Blob metadata supports citations.

Other built-in skills

• Text Translation skill: leave out defaultFromLanguageCode to auto-detect the source language.
• Document Extraction pulls text from file content in the record.
• EntityRecognitionSkill (V3) returns Location and other categories. EntityLinkingSkill links entities to Wikipedia. The Azure Language entity linking feature behind it retires from Azure Language on 1 September 2028 (use named entity recognition instead).

Custom skills

• A Custom Web API skill (Microsoft.Skills.Custom.WebApiSkill with a uri, such as an Azure Function) can wrap Document Intelligence’s prebuilt invoice model to extract invoice fields.
• Its output lands at <context>/<output>, and a context ending /* runs once per item.

Billing for skills

• Built-in skills are free for only 20 documents per indexer per day.
• Beyond that, attach a billable multi-service Foundry resource, which covers both Vision OCR and Language skills with one key.

Common trap: attaching an Azure Vision (Computer Vision) resource to bill the skillset - the Language skills aren’t covered; attach a multi-service Foundry resource.

Enrichment pipeline for multilingual documents

A pipeline that processes multilingual PDFs and keeps the enriched output uses these stages:

Common trap: choosing Azure Files as the enrichment destination - knowledge stores write only to Blob and Table Storage.

Get the whole book

This note is one section of Ultra Transcenders AI-103: Developing AI Apps and Agents on Azure, an independent study guide that explains every topic the exam covers by technology, with comparison tables, diagrams and the common traps, plus a glossary linked to Microsoft Learn.

About the book · Free AI-103 glossary · All AI-103 study notes