Microsoft Agent Framework: Workflow (Intent Detection)

on

https://www.pexels.com/photo/person-holding-compass-691637/
https://www.pexels.com/photo/person-holding-compass-691637/

We may run into use cases when we need to figure out the intention of an incoming message to our agentic solution and the determine the appropriate way to handle it. 

In this blog, we have a sample project demonstrating how to build a workflow using MicrosoftAgent Framework (MAF) to handle different types of user messages, including greetings, questions, inappropriate content, and general statements.

Source code can be found at https://github.com/dennisseah/maf-workflow

It uses Microsoft Agent Framework's Workflow API to create a workflow that processes user messages through various agents and handlers based on the intent detected in the message.

We created a simple use case to show how this works. The use case is user sends in a message, and based on the content of the message, the workflow routes the message to different handlers. The message can be one of the following types:

  1. Greeting: A friendly greeting message (e.g., "Hello", "Hi there!").
  2. Question: A message that contains a question or request for information (e.g.,  "What is 1 + 1?").
  3. Inappropriate Content: A message that contains inappropriate or offensive content (e.g., "You are stupid!").
  4. General Statement: A message that is a general statement or comment (e.g., "I  love programming.").

In this above use case, we want to make sure that we handle each type of message appropriately. And only use LLM when necessary (i.e., for questions). 

  • When it is a greeting, the first LLM agent already generates a response, so we just return that.
  • When it is inappropriate content, we return a standard message indicating that the content is not acceptable.
  • For general statements, we return a standard acknowledgment message.
  • For questions, we route the message to an assistant agent (LLM) to generate a response.

The LLM is used at most twice as explained above.

Implementation Details

Workflow diagram

In the diagram above, each node represents an agent or handler in the workflow.
  • the square brackets indicate handlers
  • the double square brackets indicate agents.
  • The arrows show the flow of messages based on the detected intent.
The workflow starts with an intent detection agent that analyzes the user's message to determine its intent. Based on the detected intent, the workflow routes the message to the appropriate handler:

1. If the message is a greeting, it is handled by the greeting handler. The greeting handler just returns the response that is already generated by the intent detection agent.
2. If the message is inappropriate, it is managed by the inappropriate content handler. The inappropriate content handler returns a standard message indicating that the content is not acceptable.
3. If the message is a general statement, it is processed by the statement handler. The statement handler returns a standard message acknowledging the statement.
4. If the message is a question, it is processed by the question response handler. The question response handler generates a response to the user's question using an assistant agent.

Essentially, the intent detection agent acts as a decision-maker, directing the flow of messages to the appropriate handlers based on their content. It generates a JSON object with the following fields:
  • is_greeting (bool): true if the message is a greeting, false otherwise.
  • is_question (bool): true if the message is a question or request for information, false otherwise.
  • is_inappropriate (bool): true if the message is inappropriate, false otherwise.
  • is_statement (bool): true if the message is a general statement, false otherwise.
  • response (string): Generate a response to the message only if it is a greeting. For all other message types, return an empty string.
Here is the pydantic model representing the intent detection result:
from pydantic import BaseModel

class IntentDetectionResult(BaseModel):
    message_content: str = Field(..., description="The original message content.")
    is_greeting: bool
    is_inappropriate: bool
    is_question: bool
    is_statement: bool
    response: str = Field(
        ...,
        description="Generated response to the message if it is a greeting; empty string otherwise.",
    )

The Microsoft Agent Framework's Workflow feature provides us a way to create a directed acyclic graph for the above implementation.

Here is demo (without audio). Please view in full screen mode. Observe the 4 different types of messages and how they are handled accordingly.


We have demonstrate how an intent detection agent can determine the message type and generate a JSON object. And the appropriate handlers are activated based on the JSON output.

Comments

Post a Comment