Integrate Azure AI Search with Python

This skill builds full-text, vector, and hybrid search applications with AI enrichment using azure-search-documents (Python), authenticated via AzureKeyCredential or DefaultAzureCredential (preferred). It distinguishes four client types: SearchClient (query and document CRUD), SearchIndexClient (index management, synonym maps, knowledge sources/bases), SearchIndexerClient (indexers, data sources, skillsets), and KnowledgeBaseRetrievalClient (agentic, LLM-powered Q&A retrieval). Index creation defines fields (SimpleField, SearchableField, a vector SearchField with a declared dimension count and profile name) plus a VectorSearch configuration wiring an HNSW algorithm to a named profile - extended in a more advanced pattern with an AzureOpenAIVectorizer (resource URL, deployment, and model name) so the service can generate embeddings automatically, and a SemanticSearch configuration prioritizing specific fields for semantic ranking. Document operations cover upload_documents/merge_documents/merge_or_upload_documents/delete_documents directly via SearchClient, or SearchIndexingBufferedSender as a context manager that handles batching and retries automatically for bulk uploads. Search patterns cover basic keyword search with field selection; vector search via VectorizedQuery (a query embedding, k-nearest-neighbors count, and target vector field); hybrid search combining search_text with vector_queries in one call; semantic ranking (query_type=SEMANTIC with a named semantic configuration, surfacing @search.captions); OData filters and faceted counts (top=0 to fetch only facets, no documents); and autocomplete/suggest against a suggester defined at index-creation time (suggesters cannot be added later). An AI-enrichment indexer pipeline is shown end to end: creating a blob data source connection, a skillset with an EntityRecognitionSkill mapping document content to extracted organization entities, and an indexer tying the data source, target index, and skillset together. Agentic retrieval introduces Knowledge Sources (pointing to a search index) and Knowledge Bases (wrapping sources plus an LLM for query planning and answer synthesis), with two output modes - EXTRACTIVE_DATA (raw chunks) or ANSWER_SYNTHESIS (LLM-generated answers). An async variant (azure.search.documents.aio) mirrors the sync API with async with/async for. A field-type reference table maps EDM types (Edm.String, Edm.Int32/Int64, Edm.Double, Edm.Boolean, Edm.DateTimeOffset, Collection(Edm.Single) for vectors, Collection(Edm.String)) to their Python equivalents. Error handling distinguishes ResourceNotFoundError from a general HttpResponseError. Best practices: use hybrid search for best relevance, enable semantic ranking for natural-language queries, index in batches of 100-1000 documents (or use the buffered sender), configure vector dimensions to match the embedding model, use HNSW for large-scale vector search, create suggesters at index-creation time, always define a semantic configuration for agentic-retrieval indexes, and close clients via context managers.

Use this skill when building search functionality in Python on Azure AI Search - full-text, vector/embedding, semantic, or hybrid search, AI-enriched content extraction via skillsets, or LLM-powered agentic retrieval over a knowledge base. It is not a deep dive into indexer/skillset orchestration beyond the entity-recognition example shown, nor a full agentic-retrieval implementation guide - both point to companion reference files for detailed patterns.

Input is documents (with optional precomputed embeddings) to index, or search queries (keyword text, vector embeddings, filters, or natural-language questions for agentic retrieval). Output is ranked search results with scores, facet counts, autocomplete/suggestion candidates, semantic captions, or for agentic retrieval, either raw extracted chunks or synthesized LLM answers.

Built on Azure AI Search via azure-search-documents, authenticated with Azure Identity's DefaultAzureCredential, with optional AzureOpenAIVectorizer integration for automatic embedding generation and KnowledgeBaseRetrievalClient for LLM-powered agentic retrieval.

pip install azure-search-documents

Python developers building search experiences - full-text, vector, semantic, hybrid, AI-enriched, or agentic retrieval - on Azure AI Search, from basic keyword queries to LLM-synthesized answers over a knowledge base.