Define an agent’s input and output schema

MLflow Model Signatures define the input and output schema requirements for your AI agent. The Model Signature tells internal and external components how to interact with your agent. The Model Signature acts as a validation check to ensure that inputs adhere to schema requirements.

For example, to use the Agent Evaluation review app, your agent must adhere to the Agent Evaluation input schema.

Supported input schemas

Mosaic AI Agent Framework supports the following input schemas.

OpenAI chat completion schema

Note

Databricks recommends the OpenAI chat completion schema because it’s widely used and is interoperable with many agent frameworks and applications. If the OpenAI chat completion schema does not meet your needs, you can define your own schema. See Custom agent schemas.

• (Recommended) Databricks recommends using the OpenAI chat completion schema. The OpenAI chat completion schema should have an array of objects as a messages parameter. This format is best for RAG applications.

  question = {
      "messages": [
          {
              "role": "user",
              "content": "What is Retrieval-Augmented Generation?",
          },
          {
              "role": "assistant",
              "content": "RAG, or Retrieval Augmented Generation, is a generative AI design pattern that combines a large language model (LLM) with external knowledge retrieval. This approach allows for real-time data connection to generative AI applications, improving their accuracy and quality by providing context from your data to the LLM during inference. Databricks offers integrated tools that support various RAG scenarios, such as unstructured data, structured data, tools & function calling, and agents.",
          },
          {
              "role": "user",
              "content": "How to build RAG for unstructured data",
          },
      ]
  }
  

SplitChatMessageRequest

SplitChatMessagesRequest is recommended for multi-turn chat applications, especially when you want to manage the current query and history separately.

  question = {
      "query": "What is MLflow",
      "history": [
          {
              "role": "user",
              "content": "What is Retrieval-augmented Generation?"
          },
          {
              "role": "assistant",
              "content": "RAG is"
          }
      ]
  }

Langchain Expression Language

If your agent uses LangChain, you can write your chain in LangChain Expression Language. In your chain definition code, you can use an itemgetter to get the messages or query or history objects depending on your input format.

Supported output schemas

Mosaic AI Agent Framework supports the following output schemas.

ChatCompletionResponse

(Recommended) ChatCompletionResponse is recommended for customers with OpenAI response format interoperability.

LangChain - ChatCompletionsOutputParser

If your agent uses LangChain, use ChatCompletionsOutputParser() from MLflow as your final chain step. Doing so formats the LangChain AI message into an agent-compatible format.

  from mlflow.langchain.output_parsers import ChatCompletionsOutputParser

  chain = (
      {
          "user_query": itemgetter("messages")
          | RunnableLambda(extract_user_query_string),
          "chat_history": itemgetter("messages") | RunnableLambda(extract_chat_history),
      }
      | RunnableLambda(DatabricksChat)
      | ChatCompletionsOutputParser()
  )

PyFunc - annotate input and output classes

If you are using PyFunc, Databricks recommends using type hints to annotate the predict() function with input and output data classes that are subclasses of classes defined in mlflow.models.rag_signatures.

You can construct an output object from the data class inside predict() to ensure the format is followed. The returned object must be transformed into a dictionary representation to ensure it can be serialized.

  from mlflow.models.rag_signatures import ChatCompletionRequest, ChatCompletionResponse, ChainCompletionChoice, Message

  class RAGModel(PythonModel):
    ...
      def predict(self, context, model_input: ChatCompletionRequest) -> ChatCompletionResponse:
        ...
        return asdict(ChatCompletionResponse(
            choices=[ChainCompletionChoice(message=Message(content=text))]
        ))

Explicit and inferred signatures

MLflow can infer the input and output schema of your agent at runtime and create a signature automatically. If you use supported input and output schemas, the inferred signatures are compatible with the Agent Framework. For information about supported schemas, see Supported input schemas.

However, if you use a custom agent schema, you must explicitly define your Model Signature according to the instructions in Custom agent schemas.

Custom agent schemas

You can customize an agent’s schema to pass and return additional fields to and from the agent by creating a subclass of a supported input/output schema. Then, add the extra keys custom_inputs and custom_outputs to contain the additional fields. See code examples for Pyfunc and Langchain below.

To use the databricks-agents SDK, Databricks client UIs such as the AI Playground and the Review App, and other Mosaic AI Agent Framework features, your agent’s schema must fulfill the following requirements:

• The agent must use mlflow version 2.17.1 or above.

• In the agent notebook, mark additional fields added in your subclass as Optional and assign default values.

• In the driver notebook, construct a ModelSignature using infer_signature with instances of your subclasses.

• In the driver notebook, construct an input example by calling asdict on your subclass.

PyFunc custom schemas

In addition to the requirements above, PyFunc-based agents must also meet the following requirements to interact with Mosaic AI agent features.

PyFunc custom schema requirements

In the agent notebook, the predict and predict stream functions must meet the following requirements:

• Have type hints for your input subclass.

• Use dot notation to access dataclass fields (for example, use model_input.custom_input.id instead of model_input["custom_inputs"]).

• Return a dictionary. You can call asdict on an instance of your subclass to format the return as a dictionary.

The following notebooks show a custom schema example using PyFunc.

PyFunc custom schema agent notebook

Open notebook in new tab Copy link for import

PyFunc custom schema driver notebook

Open notebook in new tab Copy link for import

Langchain custom schemas

The following notebooks show a custom schema example using LangChain. You can modify the wrap_output function in the the notebooks to parse and extract information from the message stream.

Langchain custom schema agent notebook

Open notebook in new tab Copy link for import

Langchain custom schema driver notebook

Open notebook in new tab Copy link for import