Agent Framework Azure DocumentDB

Store and search vector records in Azure DocumentDB with this alpha integration for Microsoft Agent Framework. The package uses PyMongo's stable asynchronous API and Azure DocumentDB's Mongo-compatible $search.cosmosSearch dialect.

• AzureDocumentDBCollection provides batch CRUD, metadata filters, index reconciliation, and vector search.
• AzureDocumentDBStore creates collection clients sharing one resolved database.
• AzureDocumentDBSettings defines the two Agent Framework-managed connection settings.

Installation

pip install agent-framework-azure-documentdb --pre

Supports Python 3.11–3.15. Python 3.10 installation remains permitted, but compatibility is not guaranteed.

Requires PyMongo 4.15+ and an Azure DocumentDB cluster. IVF indexes are intended for smaller datasets and M10/M20 tiers. HNSW and DiskANN require M30 or higher; consult the current service documentation before choosing a production tier.

Authentication and setup

Set AZURE_DOCUMENTDB_CONNECTION_STRING to the connection string from the Azure portal and AZURE_DOCUMENTDB_DATABASE_NAME to an existing or intended database. Settings precedence is explicit constructor value, selected .env file, then process environment. Connection strings are held in Agent Framework SecretString values.

Connector-created clients enforce TLS, disable retryable writes, and set an application name. They are closed by aclose() or an async context manager. An injected PyMongo AsyncMongoClient, AsyncDatabase, or AsyncCollection remains caller-owned and bypasses settings loading.

The identity running ensure_collection_exists() needs permission to create collections and indexes. Each filtered data field must declare is_indexed=True; the connector creates its ordinary ascending index alongside separate vector indexes.

Example

import asyncio
from dataclasses import dataclass
from typing import Annotated

from agent_framework import Filter, VectorStoreField, vectorstoremodel
from agent_framework_azure_documentdb import AzureDocumentDBStore


@vectorstoremodel(collection_name="articles")
@dataclass
class Article:
    id: Annotated[str, VectorStoreField("key")]
    category: Annotated[str, VectorStoreField("data", is_indexed=True)]
    embedding: Annotated[
        list[float] | None,
        VectorStoreField("vector", dimensions=3, index_kind="ivf_flat"),
    ] = None


async def main() -> None:
    async with AzureDocumentDBStore() as store:
        collection = store.get_collection(Article)
        await collection.ensure_collection_exists()
        await collection.upsert(
            [Article("one", "database", [1.0, 0.0, 0.0])],
            generate_vectors=False,
        )
        results = await collection.search(
            vector=[1.0, 0.0, 0.0],
            filter=Filter("category", "eq", "database"),
        )
        async for result in results:
            print(result["record"], result["score"])


asyncio.run(main())

Limits

The connector supports explicit string and signed 64-bit integer _id keys, multiple top-level dense vector fields, storage aliases, IVF/HNSW/DiskANN indexes, native metadata filters, server-side paging, and native searchScore values. Generated ObjectIds, ordered retrieval, hybrid/full-text search, sparse/binary vectors, nested field paths, compressed vectors, and automatic schema/index migration are not supported.

The core default index kind maps to IVF so it does not silently require an M30+ tier. Select hnsw or disk_ann explicitly when those service contracts and cluster requirements are appropriate.

Vector dimensions are conservatively limited to the service's 2,000-dimension standard-vector contract. Azure DocumentDB also offers higher limits with half-precision or product quantization; those distinct index/storage options are outside this connector.

score_threshold uses native metric units after $search selects its k candidates and before $skip/$limit. It is a minimum for cosine and inner product scores, where larger is better, and a maximum for Euclidean distance, where smaller is better. Set a larger operation_options={"k": ...} candidate window when needed. Thresholded ANN search can return fewer than top; the connector does not fetch or filter a client-side prefix. Algorithm tuning uses n_probes, ef_search, or l_search for IVF, HNSW, or DiskANN respectively.

See the Azure DocumentDB vector search guide, Azure DocumentDB limits, PyMongo async API, and Agent Framework documentation.