Microsoft Agent Framework adds experimental vector-store APIs for Python agents
The brief
Microsoft Agent Framework’s Python support carries forward Semantic Kernel’s vector-store model for typed records, collections, embeddings, batch operations, and search, while adding streamlined registration and agent-focused context providers. The APIs support searchable data and bounded conversation history across multiple database integrations, but remain experimental, with individual integration packages classified as alpha or beta.
Key points
- The Python APIs support vector-store models based on dataclasses, Pydantic, msgspec, plain classes, dictionaries, and custom codecs.
- Agents can use vector search through standalone tools or the VectorCollectionContextProvider, which can expose upsert, get, delete, and search operations with optional approvals and scoping.
- VectorStoreHistoryProvider stores full conversations in a searchable collection while loading a compact working set into the model context.
- Supported integrations include Azure AI Search, Azure Cosmos DB for NoSQL, Azure DocumentDB, MongoDB, Oracle AI Database 26ai, Redis, SQL Server and Azure SQL, PostgreSQL with pgvector, DuckDB, Qdrant, and an in-memory store.
- The vector-store APIs are experimental, and integration packages are alpha or beta with capabilities varying by implementation.
Read the source report
Microsoft Agent Framework Blog ↗Full summary
Microsoft Agent Framework has ported Python vector-store work from Semantic Kernel and retains its core model of typed records, collections, batch storage and retrieval, embedding generation, and vector search. The framework starts with application-owned data models, using VectorStoreField and the @vectorstoremodel decorator to define keys, data, and vector fields. It supports several model styles, including dataclasses, Pydantic models, msgspec structs, plain classes, dictionaries, and externally owned types through custom codecs.
The framework can expose vector search as an agent tool through create_vector_search_tool, with configurable names, descriptions, result formats, filters, and approval behavior. Its Param helper supports typed model-supplied filter arguments, including enum-like Literal values and optional parameters. VectorCollectionContextProvider offers a broader interface that can provide upsert, get, delete, and search tools, along with usage instructions, while allowing read-only access, approvals, data scoping, and specialized search tools.
VectorStoreHistoryProvider addresses growing conversation context by persisting the full transcript in a searchable vector collection and loading only a compacted working set into the model context. With include_search_tool=True, an agent can search the complete scoped history when it needs older decisions, requirements, or preferences. Embeddings may be supplied in advance or generated during upsert and search, with separate document and query options; vector dimensions are defined by the model and can be validated before storage.
The same collection API is available across supported integrations, with portable filters and keyword-hybrid search where implemented, although indexes, metrics, filters, and search modes vary by backend. Current Python integrations include Azure AI Search, Azure Cosmos DB for NoSQL, Azure DocumentDB, MongoDB, Oracle AI Database 26ai, Redis, SQL Server and Azure SQL, PostgreSQL with pgvector, DuckDB, Qdrant, and an in-memory store for tests and local development. The vector-store APIs are experimental, and integration packages are alpha or beta; users are directed to each implementation’s documentation for setup and supported capabilities.