llama-index-agent-introspective

LlamaIndex Agent Integration: Introspective Agent

Introduction

This agent integration package includes three main agent classes:

1. IntrospectiveAgentWorker
2. ToolInteractiveReflectionAgentWorker
3. SelfReflectionAgentWorker

These classes are used together in order to build an "Introspective" Agent which performs tasks while applying the reflection agentic pattern. In other words, an introspective agent produces an initial response to a task and then performs reflection and subsequently correction to produce an improved response to the task.

The IntrospectiveAgentWorker

This is the agent that is responsible for performing the task while utilizing the reflection agentic pattern. It does so by merely delegating the work to two other agents in a purely deterministic fashion.

Specifically, when given a task, this agent delegates the task to first a MainAgentWorker that generates the initial response to the query. This initial response is then passed to the ReflectiveAgentWorker to perform the reflection and subsequent correction of the initial response. Optionally, the MainAgentWorker can be skipped if none is provided. In this case, the users input query will be assumed to contain the original response that needs to go thru reflection and correction.

The Reflection Agent Workers

These subclasses of the BaseAgentWorker are responsible for performing the reflection and correction iterations of responses (starting with the initial response from the MainAgentWorker). This package contains two reflection agent workers: ToolInteractiveReflectionAgentWorker and SelfReflectionAgentWorker.

The ToolInteractiveReflectionAgentWorker

This agent worker implements the CRITIC reflection framework introduced by Gou, Zhibin, et al. (2024) ICLR. (source: https://arxiv.org/pdf/2305.11738)

CRITIC stands for Correcting with tool-interactive critiquing. It works by performing a reflection on a response to a task/query using external tools (e.g., fact checking using a Google search tool) and subsequently using the critique to generate a corrected response. It cycles thru tool-interactive reflection and correction until a specific stopping criteria has been met or a max number of iterations has been reached.

The SelfReflectionAgentWorker

This agent performs a reflection without any tools on a given response and subsequently performs correction. Cycles of reflection and correction are executed until a satisfactory correction has been generated or a max number of cycles has been reached. To perform reflection, this agent utilizes a user-specified LLM along with a PydanticProgram to generate a structured output that contains an LLM generated reflection of the current response. After reflection, the same user-specified LLM is used again but this time with another PydanticProgram to generate a structured output that contains an LLM generated corrected version of the current response against the priorly generated reflection.

Usage

To build an introspective agent, we make use of the typical agent usage pattern, where we construct an IntrospectiveAgentWorker and wrap it with an AgentRunner. (Note this can be done convienently with the .as_agent() method of any AgentWorker class.)

IntrospectiveAgent using SelfReflectionAgentWorker

from llama_index.agent.introspective import IntrospectiveAgentWorker
from llama_index.agent.introspective import SelfReflectionAgentWorker
from llama_index.llms.openai import OpenAI
from llama_index.agent.openai import OpenAIAgentWorker

verbose = True
self_reflection_agent_worker = SelfReflectionAgentWorker.from_defaults(
    llm=OpenAI("gpt-4-turbo-preview"),
    verbose=verbose,
)
main_agent_worker = OpenAIAgentWorker.from_tools(
    tools=[], llm=OpenAI("gpt-4-turbo-preview"), verbose=verbose
)

introspective_worker_agent = IntrospectiveAgentWorker.from_defaults(
    reflective_agent_worker=self_reflection_agent_worker,
    main_agent_worker=main_agent_worker,
    verbose=True,
)

introspective_agent = introspective_worker_agent.as_agent(verbose=verbose)
introspective_agent.chat("...")

IntrospectiveAgent using ToolInteractiveReflectionAgentWorker

Unlike with self reflection, here we need to define another agent worker, namely the CritiqueAgentWorker that performs the reflection (or critique) using a specified set of tools.

from llama_index.llms.openai import OpenAI
from llama_index.agent.openai import OpenAIAgentWorker
from llama_index.agent.introspective import (
    ToolInteractiveReflectionAgentWorker,
)
from llama_index.core.agent import FunctionCallingAgentWorker

verbose = True
critique_tools = []
critique_agent_worker = FunctionCallingAgentWorker.from_tools(
    tools=[critique_tools], llm=OpenAI("gpt-3.5-turbo"), verbose=verbose
)

correction_llm = OpenAI("gpt-4-turbo-preview")
tool_interactive_reflection_agent_worker = (
    ToolInteractiveReflectionAgentWorker.from_defaults(
        critique_agent_worker=critique_agent_worker,
        critique_template=(
            "..."
        ),  # template containing instructions for performing critique
        correction_llm=correction_llm,
        verbose=verbose,
    )
)


introspective_worker_agent = IntrospectiveAgentWorker.from_defaults(
    reflective_agent_worker=tool_interactive_reflection_agent_worker,
    main_agent_worker=None,  # if None, then its assumed user input is initial response
    verbose=verbose,
)
introspective_agent = introspective_worker_agent.as_agent(verbose=verbose)
introspective_agent.chat("...")

Examples

• Toxicity Reduction Example (WARNING: this notebook contains content that may be considered offensive or sensitive to some.)