IO Mapper Agent

Contributor-Covenant

About the IO Mapper Agent

When connecting agents in an application, the output of an agent needs to be compatible with the input of the agent that is connected to it. This compatibility needs to be guaranteed at three different levels:

  1. Transport level: the two agents need to use the same transport protocol.

  2. Format level: the two agents need to carry information using the same format (for example, the same JSON data structures).

  3. Semantic level: the two agents need to “talk about the same thing”.

Communication between agents is not possible if there are discrepancies between the agents at any of these levels.

Ensuring that agents are semantically compatible, that is, the output of the one agent contains the information needed by later agents, is an problem of composition or planning in the application. The IO Mapper Agent addresses level 2 and 3 compatibility. It is a component, implemented as an agent, that can make use of an LLM to transform the output of one agent to become compatible to the input of another agent. This can mean many different things:

  • JSON structure transcoding: A JSON dictionary needs to be remapped into another JSON dictionary.

  • Text summarisation: A text needs to be summarised or some information needs to be removed.

  • Text translation: A text needs to be translated from one language to another.

  • Text manipulation: Part of the information of one text needs to be reformulated into another text.

  • A combination of the above.

The IO mapper Agent can be fed the schema definitions of inputs and outputs as defined by the Agent Connect Protocol.

Getting Started

To get a local copy up and running, follow the steps below.

Prerequisites

Installation

  1. Clone the repository

    git clone https://github.com/agntcy/iomapper-agnt.git

Usage

There are several different ways to leverage the IO Mapper functions in Python. There is an agentic interface using models that can be invoked on different AI platforms and an imperative interface that does deterministic JSON remapping without using any AI models.

Key Features

The IO Mapper Agent uses an LLM to transform the inputs (typically the output of an agent) to match the desired output (typically the input of another agent). As such, it additionally supports specifying the model prompts for the translation. The configuration object provides a specification for the system and default user prompts:

This project supports specifying model interactions using LangGraph.

How to Use the IO Mapper Agent

Note: For each example, the detailed process of creating agents and configuring the respective multi-agent software is omitted. Instead, only the essential steps for configuring and integrating the IO Mapper Agent are presented.

How to use the Agent IO mapping

:warning: For each example, the detailed process of creating agents and configuring the respective multi-agent software is omitted. Instead, only the essential steps for configuring and integrating the IO mapper agent are presented.

LangGraph

We support usages with both LangGraph state defined with TypedDict or as a Pydantic object

Entities

Expand to better understand the IOMappingAgentMetadata Interface

IOMappingAgentMetadata model Interface

Field Description Required Example
input_fields An array of json paths and or instances of FieldMetadata. :white_check_mark:

["state.fiedl1", "state.field2", FieldMetadata(json_path="state", description="this is a list of items")]
output_fields An array of json paths and or instances of FieldMetadata. :white_check_mark:

["state.output_fiedl1"]
input_schema Defines the schema of the input data. :heavy_minus_sign: 
{ 
    "type": "object",
    "properties": {
        "title": {"type": "string"},
        "ingredients": {"type": "array", "items": {"type": "string"}},
        "instructions": {"type": "string"},
    },
    "required": ["title", "ingredients, instructions"],
}
OR 
from pydantic import TypeAdapter
TypeAdapter(GraphState).json_schema()
output_schema Defines the schema for the output data. :heavy_minus_sign: same as input_schema

IOMappingAgent

IOMappingAgent model

Field Description Required Example
metadata Instance of IOMappingAgentMetadata. :white_check_mark: 
IOMappingAgentMetadata(
    input_fields=["documents.0.page_content"],
    output_fields=["recipe"],
    input_schema=TypeAdapter(GraphState).json_schema(),
    output_schema={
        "type": "object",
        "properties": {
            "title": {"type": "string"},
            "ingredients": {"type": "array", "items": {"type": "string"}},
            "instructions": {"type": "string"},
        },
        "required": ["title", "ingredients, instructions"],
    },
)
llm An instance of the large language model to be used. :white_check_mark: 
        AzureChatOpenAI(
            model=model_version,
            api_version=api_version,
            seed=42,
            temperature=0,
        )

LangGraph Example 1

This example involves a multi-agent software system designed to process a create engagement campaign and share within an organization. It interacts with an agent specialized in creating campaigns, another agent specialized in identifying suitable users. The information is then relayed to an IO mapper, which converts the list of users and the campaign details to present statistics about the campaign.

Define an agent io mapper metadata

metadata = IOMappingAgentMetadata(
    input_fields=["selected_users", "campaign_details.name"],
    output_fields=["stats.status"],
)

The above instruction directs the IO mapper agent to utilize the selected_users and name from the campaign_details field and map them to the stats.status. No further information is needed since the type information can be derived from the input data which is a pydantic model.

:information_source: Both input_fields and output_fields can also be sourced with a list composed of str and/or instances of FieldMetadata as the bellow example shows:

metadata = IOMappingAgentMetadata(
    input_fields=[
        FieldMetadata(
            json_path="selected_users", description="A list of users to be targeted"
        ),
        FieldMetadata(
            json_path="campaign_details.name",
            description="The name that can be used by the campaign",
            examples=["Campaign A"]
        ),
    ],
    output_fields=["stats"],
)

Define an Instance of the Agent

mapping_agent = IOMappingAgent(metadata=metadata, llm=llm)

Add the node to the LangGraph graph

workflow.add_node(
    "io_mapping",
    mapping_agent.langgraph_node,
)

Add the Edge

With the edge added, you can run the your LangGraph graph.

workflow.add_edge("create_communication", "io_mapping")
workflow.add_edge("io_mapping", "send_communication")

A flow chart of Io Mapper in a LangGraph graph of the discussed multi agent software discussed above

        flowchart TD
    A[create_communication] -->|input in specific format| B(IO Mapper Agent)
    B -->|output expected format| D[send_communication]

LangGraph Example 2

This example involves a multi-agent software system designed to process a list of ingredients. It interacts with an agent specialized in recipe books to identify feasible recipes based on the provided ingredients. The information is then relayed to an IO mapper, which converts it into a format suitable for display to the user.

Define an Agent IO Mapper Metadata

metadata = IOMappingAgentMetadata(
    input_fields=["documents.0.page_content"],
    output_fields=["recipe"],
    input_schema=TypeAdapter(GraphState).json_schema(),
    output_schema={
        "type": "object",
        "properties": {
            "title": {"type": "string"},
            "ingredients": {"type": "array", "items": {"type": "string"}},
            "instructions": {"type": "string"},
        },
        "required": ["title", "ingredients, instructions"],
    },
)

Define an Instance of the Agent

mapping_agent = IOMappingAgent(metadata=metadata, llm=llm)

Add the node to the LangGraph graph

graph.add_node(
    "recipe_io_mapper",
    mapping_agent.langgraph_node,
)

Add the Edge

With the edge added, you can run the your LangGraph graph.

graph.add_edge("recipe_expert", "recipe_io_mapper")

LlamaIndex

We support both LlamaIndex Workflow and the new AgentWorkflow multi agent software

Entities

IOMappingInputEvent

IOMappingInputEvent event received by io mapper step

Property Description Required Value Example
metadata Object used to describe the input fields, output fields schema and any relevant information to be used in the mapping :white_check_mark: 
IOMappingAgentMetadata(
    input_fields=["selected_users", "campaign_details.name"],
    output_fields=["stats"],
)
config Object containing information such as the llm instance that will be used to perform the translation :white_check_mark: 
LLamaIndexIOMapperConfig(llm=llm)
data Represents the input data to be used in the translation :white_check_mark: 
OverallState(campaign_details=campaign_details, selected_users=ev.list_users),
IOMappingOutputEvent

IOMappingOutputEvent event received by io mapper step

Property Description Required Value Example
mapping_result A dictionary containing the result of the mapping :white_check_mark: N/A

Example of usage in a LlamaIndex workflow

In this example we recreate the campaign workflow using LlamaIndex workflow

Begin by importing the neccessary object

from agntcy_iomapper import IOMappingAgent, IOMappingAgentMetadata

Define the workflow

class CampaignWorkflow(Workflow):
    @step
    async def prompt_step(self, ctx: Context, ev: StartEvent) -> PickUsersEvent:
        await ctx.set("llm", ev.get("llm"))
        return PickUsersEvent(prompt=ev.get("prompt"))

    @step
    async def pick_users_step(
        self, ctx: Context, ev: PickUsersEvent
    ) -> CreateCampaignEvent:
        return CreateCampaignEvent(list_users=users)

    # The step that will trigger IO mapping 
    @step
    async def create_campaign(
        self, ctx: Context, ev: CreateCampaignEvent
    ) -> IOMappingInputEvent:
        prompt = f"""
        You are a campaign builder for company XYZ. Given a list of selected users and a user prompt, create an engaging campaign. 
        Return the campaign details as a JSON object with the following structure:
        {{
            "name": "Campaign Name",
            "content": "Campaign Content",
            "is_urgent": yes/no
        }}
        Selected Users: {ev.list_users}
        User Prompt: Create a campaign for all users
        """
        parser = PydanticOutputParser(output_cls=Campaign)
        llm = await ctx.get("llm", default=None)

        llm_response = llm.complete(prompt)
        try:
            campaign_details = parser.parse(str(llm_response))
            metadata = IOMappingAgentMetadata(
                input_fields=["selected_users", "campaign_details.name"],
                output_fields=["stats"],
            )
            config = LLamaIndexIOMapperConfig(llm=llm)

            io_mapping_input_event = IOMappingInputEvent(
                metadata=metadata,
                config=config,
                data=OverallState(
                    campaign_details=campaign_details,
                    selected_users=ev.list_users,
                ),
            )
            return io_mapping_input_event
        except Exception as e:
            print(f"Error parsing campaign details: {e}")
            return StopEvent(result=f"{e}")

    @step
    async def after_translation(self, evt: IOMappingOutputEvent) -> StopEvent:
        return StopEvent(result="Done")

It is important to notice: The step create_campaign will trigger the IO mapper. Why? Well, because:

  1. It declares that it returns an instance of IOMappingInputEvent

async def create_campaign(self, ctx: Context, ev: CreateCampaignEvent) -> IOMappingInputEvent:

  1. And finally it creates and returns a valid instance of the IOMappingInputEvent

# define an instance metadata
metadata = IOMappingAgentMetadata(
   input_fields=["selected_users", "campaign_details.name"],
   output_fields=["stats"]
)

#define an instance of config with must have an llm instance
config = LLamaIndexIOMapperConfig(llm=llm)

# Finally return define and return the IOMappingInputEvent
io_mapping_input_event = IOMappingInputEvent(
    metadata=metadata,
    config=config,
    data=OverallState(
        campaign_details=campaign_details,
        selected_users=ev.list_users,
    ),
)
return io_mapping_input_event

Add The IO mapper step

    w = CampaignWorkflow()

    IOMappingAgent.as_worfklow_step(workflow=w)

Example of usage in a LlamaIndex AgentWorkflow

In this example we recreate the recipe workflow using LlamaIndex workflow

Import the necessary objects

from agntcy_iomapper import FieldMetadata, IOMappingAgent, IOMappingAgentMetadata

Define an instance of the IOMappingAgentMetadata

mapping_metadata = IOMappingAgentMetadata(
    input_fields=["documents.0.text"],
    output_fields=[
        FieldMetadata(
            json_path="recipe",
            description="this is a recipe for the ingredients you've provided",
        )
    ],
    input_schema=TypeAdapter(GraphState).json_schema(),
    output_schema={
        "type": "object",
        "properties": {
            "title": {"type": "string"},
            "ingredients": {"type": "array", "items": {"type": "string"}},
            "instructions": {"type": "string"},
        },
        "required": ["title", "ingredients, instructions"],
    },
)

Finally define the IOMappingAgent and add it to the AgentWorkflow.

Important to note that a tool is passed, to instruct the io mapper where to go next in the flow.

io_mapping_agent = IOMappingAgent.as_workflow_agent(
    mapping_metadata=mapping_metadata,
    llm=llm,
    name="IOMapperAgent",
    description="Useful for mapping a recipe document into recipe object",
    can_handoff_to=["Formatter_Agent"],
    tools=[got_to_format],
)


io_mapping_agent = IOMappingAgent.as_workflow_agent(
    mapping_metadata=mapping_metadata,
    llm=llm,
    name="IOMapperAgent",
    description="Useful for mapping a recipe document into recipe object",
    can_handoff_to=["Formatter_Agent"],
    tools=[got_to_format],
)

Use Examples

  1. Install:

  2. From the examples folder run the desired make command, for example:

make make run_lg_eg_py

Contributing

Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated. For detailed contributing guidelines, please see CONTRIBUTING.md