Implement information extraction and knowledge mining solutions
The enrichment pipeline
Azure AI Search mining follows a fixed chain: data source (Blob, ADLS, SQL, Cosmos DB) → indexer (scheduling, change detection, field mappings) → skillset (AI enrichment) → index (searchable, filterable, sortable fields). Skills build an enrichment tree rooted at /document; built-in skills include OCR, text merge, text split (chunking), language detection, entity recognition, key phrase extraction, entity linking, Document Intelligence layout, embedding, and an Azure OpenAI chat completion skill. Custom skills are Azure Functions you host.
Wiring, the usual failure point
Every skill writes to a context path, and every value you want to search must be mapped into the index through output field mappings. If a field is empty, check the skill's context and the output mapping before suspecting the model. Attach a knowledge store with projections (objects, tables, files) only when you need to export enriched data, not to make it searchable.
Retrieval quality
Enable integrated vectorization with a vectorizer so queries are embedded automatically, use hybrid search plus the semantic ranker for natural-language queries, and apply scoring profiles or filters for business rules. Indexers are incremental by default using a high-water mark; full reindexing is explicit.
Traps: confusing indexer field mappings (source → index) with skillset output mappings (enrichment → index); expecting chunking to happen automatically without a split skill; assuming the semantic ranker replaces keyword search rather than reranking results.
Worked rule: when a search result is missing content, trace the value through the tree - data source → skill context → output mapping → index field.