Add MongoDB Vector Search Tool

[blink1073]

I am a maintainer of langchain-mongodb, which I wrapped to create a crewAI tool.

[lorenzejay]

@blink1073 nice work. Can we try an approach where we are using the direct apis from mongo over wrapping another package just to call this? Don't want the dependency overhead of langchain here.

[blink1073]

Not a problem, I'll just end up duplicating some code from langchain-mongodb.

[blink1073]

@lorenzejay I'm made the requested changes while preserving the API.

[lorenzejay]

@blink1073 made a push for some fixes. use crewai==0.126.0

using this:

if __name__ == "__main__":
    from crewai import Agent, Task, Crew
    from crewai_tools import MongoDBVectorSearchConfig, MongoDBVectorSearchTool

    # Setup custom embedding model and customize the parameters.
    query_config = MongoDBVectorSearchConfig(limit=10)
    tool = MongoDBVectorSearchTool(
        database_name="sample_mflix",
        collection_name="embedded_movies",
        connection_string="<>",
        query_config=query_config,
        vector_index_name="_id_",
        generative_model="gpt-4o",
    )

    # Adding the tool to an agent
    rag_agent = Agent(
        name="rag_agent",
        role="You are a helpful assistant that can answer questions with the help of the MongoDBVectorSearchTool.",
        goal="You are a helpful assistant that can answer questions with the help of the MongoDBVectorSearchTool.",
        backstory="You are a helpful assistant that can answer questions with the help of the MongoDBVectorSearchTool.",
        llm="gpt-4o-mini",
        tools=[tool],
    )
    task = Task(
        name="rag_task",
        description="You are a helpful assistant that can answer questions with the help of the MongoDBVectorSearchTool. the query: {query}",
        expected_output="The answer to the question",
        agent=rag_agent,
    )
    crew = Crew(agents=[rag_agent], tasks=[task], verbose=True)
    res = crew.kickoff(inputs={"query": "tell me about the movie: From Hand to Mouth"})
    print("res", res)

i'm getting pretty poor results on the default db. any help?

[blink1073]

I'll take a look. Here's the integ test I had written that will run using our creds nightly: mongodb-labs/ai-ml-pipeline-testing#71.

[blink1073]

Ah, I see the difference, what you're using it as is actually a follow-up capability, searching within a database itself. This initial PR is for vector search only, which is what my example does. It creates embeddings for each page of the PDF and then runs the query against those embeddings.

[blink1073]

I updated the crewai dep

[blink1073]

To clarify, as part of INTPYTHON-332 we would add a mongodb_search_tool directory which could be used to perform your query.

[lucasgomide]

Good work here!

I dropped a few comments let mw know what you think

[blink1073]

Just a heads up, I'm still working on updating our integration tests in mongodb-labs/ai-ml-pipeline-testing#71, I'll let y'all know when I get it working

[lorenzejay]

@blink1073 can you try running this:

i'm getting poor results:

if __name__ == "__main__":
    from crewai import Agent, Task, Crew

    tool = MongoDBVectorSearchTool(
        database_name="sample_mflix",
        collection_name="embedded_movies",
        connection_string="<>",
        embedding_key="plot_embedding",
    )
   

    agent = Agent(
        role="MongoDBVectorSearchTool",
        goal="You are a helpful assistant that can answer questions about the MongoDB database.",
        backstory="You are a helpful assistant that can answer questions about the MongoDB database.",
        tools=[tool],
        llm="gpt-4.1",
    )

    task = Task(
        description="get the movies with the director Alfred J. Goulding, use no filters",
        expected_output="The movies with the director Alfred J. Goulding",
        agent=agent,
    )

    crew = Crew(
        agents=[agent],
        tasks=[task],
        verbose=True,
    )
    result = crew.kickoff()
    print("result", result)

[lorenzejay]

im using the default collection when you create a new mongo instance

[blink1073]

We had that same conversation last month. ;)

#319 (comment)

[lorenzejay]

Do you have an example of it working then? Something I can run to confirm ? Let’s bring this home today

[blink1073]

Yes, the integration test is now passing: https://github.com/mongodb-labs/ai-ml-pipeline-testing/pull/71/files#diff-5c01b996bf644e0a14a5aa2a00ec357d24dbe961c3157919a979bc762f1344c4

[lorenzejay]

LGTM.

[blink1073]

Excellent, thank you both!