Document Library

Document Library is a built-in tool tool that enables agents to access documents from Mistral Cloud.

You can manage your libraries and upload your documents to them, enabling a seemless integration with your agents, letting them access external data of your choice.

It is a built-in RAG capability that enhances your agents' knowledge with the data you have uploaded.

To use the Document Library tool, you can create an agent or use the Conversations API directly.

However, you will need to create a library first, we provide a complete API allowing full control over your libraries.

library_agent = client.beta.agents.create(
    model="mistral-medium-2505",
    name="Document Library Agent",
    description="Agent used to access documents from the document library.",
    instructions="Use the  library tool to access external documents.",
    tools=[{"type": "document_library", "library_ids": [new_library.id]}],
    completion_args={
        "temperature": 0.3,
        "top_p": 0.95,
    }
)

library_agent = client.beta.agents.create(
    model="mistral-medium-2505",
    name="Document Library Agent",
    description="Agent used to access documents from the document library.",
    instructions="Use the  library tool to access external documents.",
    tools=[{"type": "document_library", "library_ids": [new_library.id]}],
    completion_args={
        "temperature": 0.3,
        "top_p": 0.95,
    }
)

Now that we have our document library agent ready, we can search them on demand at any point.

To start a conversation with our document library agent, we can use the following code:

response = client.beta.conversations.start(
    agent_id=library_agent.id,
    inputs="How does the vision encoder for pixtral 12b work"
)

response = client.beta.conversations.start(
    agent_id=library_agent.id,
    inputs="How does the vision encoder for pixtral 12b work"
)

Below we will explain the different outputs of the response of the previous snippet example:

• tool.execution: This entry corresponds to the execution of the document library tool. It includes metadata about the execution, such as:

  • name: The name of the tool, which in this case is document_library.
  • object: The type of object, which is entry.
  • type: The type of entry, which is tool.execution.
  • created_at and completed_at: Timestamps indicating when the tool execution started and finished.
  • id: A unique identifier for the tool execution.

• message.output: This entry corresponds to the generated answer from our agent. It includes metadata about the message, such as:

  • content: The actual content of the message, which in this case is a list of chunks. These chunks correspond to the text chunks, the actual message response of the model, sometimes interleaved with reference chunks. These reference chunks are used for citations during Retrieval-Augmented Generation (RAG) related tool usages. In this case, it provides the source of the information it just answered with, which is extremely useful for web search. This allows for transparent feedback on where the model got its response from for each section and fact answered with. The content section includes:
    • type: The type of chunk, which can be text or tool_reference.
    • text: The actual text content of the message.
  • object: The type of object, which is entry.
  • type: The type of entry, which is message.output.
  • created_at and completed_at: Timestamps indicating when the message was created and completed.
  • id: A unique identifier for the message.
  • agent_id: A unique identifier for the agent that generated the message.
  • model: The model used to generate the message, which in this case is mistral-medium-2505.
  • role: The role of the message, which is assistant.

Another tool that pro-actively uses references is the websearch tool, feel free to take a look here.
For more information regarding the use of citations, you can find more here.