Integrate MongoDB with the Agno Framework

You can integrate MongoDB with Agno, an open-source Python framework for building multi-modal AI agents. Agno provides two MongoDB integrations:

Vector database that stores your agent's knowledge base and retrieves it with MongoDB Vector Search.
Session and memory database that persists agent sessions, user memories, metrics, and evaluation runs in your MongoDB collections.

You can use either integration on its own, or you can use both in the same agent.

Get Started

To complete a tutorial that configures Agno to use MongoDB, see Get Started with the MongoDB Agno Integration.

Installation

Agno requires Python v3.9 or later. To install the latest Agno and PyMongo versions, run the following command:

pip install agno pymongo

The examples on this page use OpenAI models, which Agno uses by default. To install the OpenAI client library, run the following command:

pip install openai

The Custom configuration examples on this page use Voyage AI embeddings. To install the Voyage AI client library, run the following command:

pip install voyageai

Environment Variables

Before running the examples on this page, set the following environment variables:

Connection string: Set MONGODB_URI to the connection string for your MongoDB deployment.
Model API key: Set OPENAI_API_KEY or VOYAGE_API_KEY to the API key for your model provider.

Vector Database

The MongoVectorDb class implements the Agno vector database interface with MongoDB Vector Search. To give an agent semantic search over your documents, pass an instance of this class to the vector_db parameter of an Agno Knowledge object, then pass that object to the knowledge parameter of your agent.

Important

MongoDB Class Names

Agno names both MongoDB classes MongoDb, but imports them from different modules. Import the vector database class from agno.vectordb.mongodb and the database class from agno.db.mongo. To avoid a name collision, Agno also exports the vector database class as MongoVectorDb. The examples on this page use that alias.

Select the Basic initialization tab to view a minimal configuration example, or select the Custom configuration tab to see how to configure the integration with custom settings.

To initialize the vector database, specify the collection that stores your embeddings and your MongoDB connection string. The following example creates a knowledge base, adds a document, and gives an agent access to the data:

import os
from agno.agent import Agent
from agno.knowledge.knowledge import Knowledge
from agno.vectordb.mongodb import MongoVectorDb
vector_db = MongoVectorDb(
    collection_name="agent_knowledge",
    db_url=os.environ["MONGODB_URI"],
    database="agno",
)
knowledge = Knowledge(vector_db=vector_db)
knowledge.insert(
    name="Recipes",
    url="https://agno-public.s3.amazonaws.com/recipes/ThaiRecipes.pdf",
)
agent = Agent(knowledge=knowledge, search_knowledge=True)

The following example enables hybrid search, uses Voyage AI to generate embeddings instead of the OpenAI default, and specifies a MongoDB Vector Search index name:

import os
from agno.agent import Agent
from agno.knowledge.embedder.voyageai import VoyageAIEmbedder
from agno.knowledge.knowledge import Knowledge
from agno.vectordb.mongodb import MongoVectorDb
from agno.vectordb.search import SearchType
embedder = VoyageAIEmbedder(
    id="voyage-4-large",
    dimensions=1024,
    api_key=os.environ["VOYAGE_API_KEY"],
    request_params={"output_dimension": 1024},
)
vector_db = MongoVectorDb(
    collection_name="agent_knowledge",
    db_url=os.environ["MONGODB_URI"],
    database="agno",
    embedder=embedder,
    search_type=SearchType.hybrid,
    hybrid_vector_weight=0.7,
    hybrid_keyword_weight=0.3,
    search_index_name="agno_vector_index",
)
knowledge = Knowledge(vector_db=vector_db)
agent = Agent(knowledge=knowledge, search_knowledge=True)

If you specify an embedder, Agno uses the number of dimensions that the embedder produces when it creates the MongoDB Vector Search index. The request_params dictionary passes provider-specific keyword arguments to the embedding call. For Voyage AI, output_dimension controls the size of the returned vector for models that support variable dimensions.

When you create a Knowledge object, Agno verifies whether the collection already exists. If it doesn't, Agno creates the collection and the MongoDB Vector Search index. If the collection exists, Agno doesn't check whether the index exists, even if you deleted it separately.

After creating the index, Agno waits until the index appears on the collection. However, it doesn't verify the index status, so a query that runs immediately afterward can return no results while the index finishes building.

The MongoVectorDb class accepts the following parameters:

Parameter	Necessity	Description
`collection_name`	Required	The name of the collection that stores your documents and their embeddings.
`db_url`	Optional	The connection string for your MongoDB deployment. Defaults to `mongodb://localhost:27017/`. To learn more about finding your connection string, see Connect to a Cluster via Client Libraries.
`client`	Optional	An existing `MongoClient` instance. If you specify this parameter, Agno ignores `db_url`.
`database`	Optional	The name of the database that holds the collection. Defaults to `agno`.
`embedder`	Optional	The embedding model that Agno uses to embed your documents and queries. Defaults to `OpenAIEmbedder()`.
`distance_metric`	Optional	The similarity function that the MongoDB Vector Search index uses. Only the default value, `Distance.cosine`, is currently supported.
`search_type`	Optional	The retrieval strategy. Accepts `SearchType.vector` or `SearchType.hybrid`. Defaults to `SearchType.vector`.
`hybrid_vector_weight`	Optional	The weight that hybrid search applies to semantic results. Defaults to `0.5`.
`hybrid_keyword_weight`	Optional	The weight that hybrid search applies to keyword results. Defaults to `0.5`.
`hybrid_rank_constant`	Optional	The constant that hybrid search adds to each rank before it computes the reciprocal rank fusion score. Defaults to `60`.
`search_index_name`	Optional	The name of the MongoDB Vector Search index on your collection. Defaults to `vector_index_1`.
`wait_until_index_ready_in_seconds`	Optional	Whether Agno waits, after creating the MongoDB Vector Search index, for the index name to appear in the collection's search indexes. Set this parameter to a positive number to wait, or to `None` to skip the wait. Asynchronous methods use the value as a timeout, in seconds. Defaults to `3`.
`wait_after_insert_in_seconds`	Optional	The time, in seconds, that Agno waits after it inserts documents. Defaults to `3`.
`max_pool_size`	Optional	The maximum number of connections in the driver connection pool. Defaults to `100`.
`retry_writes`	Optional	Whether the driver retries supported write operations. Defaults to `True`.
`cosmos_compatibility`	Optional	Whether Agno creates indexes in the Azure Cosmos DB for MongoDB vCore format instead of the MongoDB Vector Search format. Defaults to `False`.
`name`, `description`, `id`	Optional	The identifying metadata for the vector database instance. If you don't specify `id`, Agno derives it from the connection string, database name, and collection name.

Vector Database Methods

The Knowledge class calls the following methods when you add content to a knowledge base or when an agent searches it. Call these methods directly only if you manage documents outside of a Knowledge object.

To limit a write or search operation to one user's documents, pass a user_id value to the operation method.

Method	Description
`search(query, limit, filters, min_score, user_id)`	Embeds `query` and returns up to `limit` matching documents. If you set `search_type` to `SearchType.hybrid`, Agno also runs a keyword search and merges the two result sets. The `filters` parameter accepts a dictionary of metadata filters. If you pass a list of Agno filter expressions instead, Agno logs a warning and applies no filters because the MongoDB integration doesn't support them.
`insert(content_hash, documents, filters, user_id)`	Embeds each document in `documents` and inserts the documents into your collection. Agno derives each document's `_id` value from a hash of the document content.
`upsert(content_hash, documents, filters, user_id)`	Replaces the documents that share `content_hash`, then inserts the documents in `documents`. Use this method to update a document whose content changed.
`delete()`	Deletes every document from your collection but keeps the collection and its MongoDB Vector Search index.
`delete_by_id(id)`	Deletes the document with the specified `_id` value.
`delete_by_name(name)`	Deletes the documents with the specified `name` value.
`delete_by_metadata(metadata)`	Deletes the documents whose `meta_data` field matches the specified key-value pairs.
`drop()`	Drops your collection and its MongoDB Vector Search index.
`exists()`	Returns `True` if your collection exists.
`get_count()`	Returns the number of documents in your collection.

Tip

Asynchronous Methods

The following methods described in the preceding table each have a corresponding asynchronous method:

search(): Use async_search()
insert(): Use async_insert()
upsert(): Use async_upsert()
drop(): Use async_drop()
exists(): Use async_exists()

Session and Memory Database

The MongoDb class in the agno.db.mongo module implements the Agno database interface. To persist an agent's sessions, user memories, metrics, and evaluation runs in MongoDB, pass an instance of this class to the db parameter of an Agent, Team, or Workflow object.

Select the Basic initialization tab to view a minimal configuration example, or select the Custom configuration tab to see how to configure the integration with custom settings.

To initialize the database, specify your MongoDB connection string. The following example creates an agent that stores its session history in MongoDB and includes that history in the model context:

import os
from agno.agent import Agent
from agno.db.mongo import MongoDb
db = MongoDb(db_url=os.environ["MONGODB_URI"])
agent = Agent(db=db, add_history_to_context=True)

The following example specifies the database name and the names of the collections that store sessions, memories, and metrics:

import os
from agno.agent import Agent
from agno.db.mongo import MongoDb
db = MongoDb(
    db_url=os.environ["MONGODB_URI"],
    db_name="agno_prod",
    session_collection="agent_sessions",
    memory_collection="agent_memories",
    metrics_collection="agent_metrics",
)
agent = Agent(db=db, add_history_to_context=True)

Agno creates each collection that you configure the first time it writes to that collection.

The MongoDb class accepts the following parameters:

Parameter	Necessity	Description
`db_url`	Conditional	The connection string for your MongoDB deployment. Required if you don't specify `db_client`.
`db_client`	Conditional	An existing `MongoClient` instance. Required if you don't specify `db_url`. If you specify both, Agno ignores `db_url`.
`db_name`	Optional	The name of the database that holds the collections. Defaults to `agno`.
`session_collection`	Optional	The name of the collection that stores agent, team, and workflow sessions. Defaults to `agno_sessions`.
`runs_collection`	Optional	The name of the collection that stores runs, with one document per run. Defaults to `agno_runs`. If you specify `session_collection` but not this parameter, Agno creates the name from your session collection name, such as `agent_sessions_runs`.
`memory_collection`	Optional	The name of the collection that stores user memories. Defaults to `agno_memories`.
`metrics_collection`	Optional	The name of the collection that stores aggregated usage metrics. Defaults to `agno_metrics`.
`eval_collection`	Optional	The name of the collection that stores evaluation runs. Defaults to `agno_eval_runs`.
`knowledge_collection`	Optional	The name of the collection that stores knowledge content metadata. Defaults to `agno_knowledge`.
`traces_collection`, `spans_collection`	Optional	The names of the collections that store traces and spans. Default to `agno_traces` and `agno_spans`.
`schedules_collection`, `schedule_runs_collection`	Optional	The names of the collections that store schedules and schedule runs. Default to `agno_schedules` and `agno_schedule_runs`.
`learnings_collection`	Optional	The name of the collection that stores learnings. Defaults to `agno_learnings`.
`id`	Optional	The identifier for the database instance. If you don't specify this parameter, Agno derives it from the connection string and database name.

As your agent runs, Agno automatically reads from and writes to the session, memory, and metrics collections described in this section. Agno creates each collection the first time it's needed. You can also call methods directly on your MongoDb instance, such as the following:

upsert_session() and get_sessions() read and write sessions.
upsert_user_memory() writes a single user memory.
get_metrics() reads aggregated usage metrics.
upsert_sessions() and upsert_memories() write many records in a single operation, such as when you migrate existing session history into MongoDB.
close() releases the client connection. Call this method when your application shuts down.

Search Index

The first time Agno creates your vector collection, it also creates an MongoDB Vector Search index on the collection. The index name is the value of the search_index_name parameter, which defaults to vector_index_1. You can also create or inspect this index manually. To learn more, see Create a MongoDB Vector Search Index.

The index uses the following definition, where numDimensions matches the number of dimensions that your embedder produces and similarity matches your distance_metric value:

{
  "fields": [
    {
      "type": "vector",
      "path": "embedding",
      "numDimensions": 1536,
      "similarity": "cosine"
    },
    {
      "type": "filter",
      "path": "user_id"
    }
  ]
}

Agno indexes the user_id field as a filter field so that it can scope searches to a single end user's documents.

Hybrid search doesn't require a MongoDB Search index. Agno runs its keyword search as a regular expression query on the content field, then combines those results with the MongoDB Vector Search results by using reciprocal rank fusion.

Additional Resources

To learn more about using Agno with MongoDB, see the following resources:

Agno MongoDB Vector Store reference
Agno MongoDB Database reference
Agno Knowledge documentation
Agno GitHub repository

Back

Mastra

Get Started

Integrate MongoDB with the Agno Framework

Get Started

Installation

Environment Variables

Vector Database

Vector Database ParametersUse these parameters to configure the vector database.

Vector Database Methods

Session and Memory Database

Database ParametersUse these parameters to configure the database.

Search Index

Additional Resources