A Information to Reflection Brokers Utilizing LlamaIndex

Introduction

Think about you’ve simply created an AI mannequin that may write, motive, and resolve complicated issues. However what if I informed you there’s a method to make this AI even smarter by instructing it to consider its personal ideas? On this article, we’ll discover the fascinating world of reflective AI brokers. We’ll begin with the fundamentals of how these brokers can mirror on their very own actions to enhance over time. Then, we’ll delve into superior methods like Language Agent Tree Search (LATS) and Introspective Brokers, exhibiting you easy methods to arrange and use these strategies with LlamaIndex. By the tip, you’ll see how these approaches are reworking AI, making it extra autonomous and able to tackling ever more difficult duties. Be a part of us on this journey to unlock the following degree of AI intelligence!

Studying Outcomes

• Perceive the idea and significance of reflection in enhancing LLM-based brokers.
• Discover the implementation of Fundamental Reflection Brokers utilizing self-prompting methods.
• Study Language Agent Tree Search (LATS) and its function in bettering AI process efficiency.
• Achieve hands-on expertise with LATS framework setup and execution utilizing LlamaIndex.
• Implement Introspective Brokers to refine responses iteratively utilizing self-reflection and exterior instruments.

This text was revealed as part of the Knowledge Science Blogathon.

Understanding Reflection or Introspective Brokers 

Many instances the LLM fail to generate the enough response for a given process. It is a frequent problem in synthetic intelligence, the place brokers typically lack the power to self-evaluate and refine their outputs.

That is the place Reflection brokers come to our rescue. Individuals typically focus on “System 1” and “System 2” pondering, with System 1 being reactive or instinctual and System 2 being extra analytical and introspective. When used successfully, reflection could help LLM techniques transfer away from solely System 1 “pondering” patterns and in direction of System 2-like behaviour.

In Llamindex reflection brokers are carried out in Introspective Brokers module.

Introspective brokers are a robust idea that makes use of the reflection agent sample throughout the LlamaIndex structure. These brokers have a definite strategy to job completion. As an alternative than offering a single response, they interact in iterative refining.

Steps in Fundamental Reflections Brokers

• Preliminary Response : The introspective agent begins by creating an preliminary response to the required job. This could be a preliminary response to a question, a primary try to finish an exercise, or perhaps a artistic work.
• Reflection and Correction: The agent then takes a step again to suppose on its first response. This reflection could also be carried out each internally or by way of exterior instruments(akin to API). LlamaIndex means that you can choose the technique that finest meets your wants.
• Refinement Cycle: Based mostly on the reflection, the agent determines areas for enchancment and creates a revised reply. This cycle of reflection and correction continues till we fulfill a stopping situation, akin to reaching a sure diploma of accuracy or finishing a predetermined variety of cycles.

• Introspective Agent: An AI agent that employs a reflection agent sample to iteratively refine its responses to a process.
• Reflection Agent Sample: A design strategy for AI brokers the place they assess their outputs (reflection) and make changes (correction) earlier than finalizing them.

What’s Language Agent Tree Search (LATS)?

The  Language Agent Tree Search (LATS), a common LLM agent search algorithm that enhances general job efficiency over comparable approaches like as ReACT, Reflexion, or Tree of Ideas by combining reflection/analysis with search (extra particularly, monte-carlo timber search). The paper launched by Zhou et.al might be learn right here. 

The LATS framework, a first-of-its-kind common framework, combines LMs’ capacities for motion, pondering, and planning. It advances the aim of making broadly distributed autonomous brokers with the power to suppose and make choices in a variety of settings.

Additionally mixes reflection/analysis with search (significantly, Monte-Carlo timber search) to enhance general work efficiency. It makes use of a typical reinforcement studying (RL) process framework, substituting the RL brokers, worth features, and optimizer with calls to an LLM. This helps the agent adapt and resolve tough duties as a substitute of getting caught in repeating cycles.

Steps in LATS framework

• Generate Candidates : Preliminary response is generated and a number of candidates are generated.
• Increase and Simulate : Utilizing the generated potential actions increase every motion and execute simulate them  in parallel.
• Replicate + consider: observe the outcomes of those actions and rating the choices primarily based on reflection (and presumably exterior suggestions utilizing exterior instruments).
• Backpropagate: replace the scores of the foundation trajectories primarily based on the outcomes.
• Choose : choose one of the best subsequent actions primarily based on the mixture rewards from above steps. Both reply (if an answer is discovered or the max search depth is reached) or proceed looking from step1.

 If the agent has a decent suggestions loop (by way of high-quality setting rewards or dependable reflection scores), the search can reliably discern between a number of motion paths and choose the optimum one. The ensuing trajectory can then be saved to exterior reminiscence (or utilised for mannequin fine-tuning) in order that the mannequin might be improved later.

Code Implementation of LATS

LlamaIndex implements LATS as a separate package deal which might be put in and run out of the field. We are going to use Cohere Embeddings and Gemini API LLM for this implementation. Each are freely obtainable to be used as trial API Keys. 

Step1: Set up Libraries

We set up libraries of llamaindex for LATS, Cohere and Gemini and a few supporting libraries for file studying.

!pip set up llama-index-agent-lats --quiet
!pip set up llama-index --quiet
!pip set up llama-index-core llama-index-readers-file --quiet
!pip set up cohere --quiet
!pip set up llama-index-llms-cohere --quiet
!pip set up llama-index-embeddings-cohere --quiet


!pip set up -q llama-index google-generativeai --quiet
!pip set up llama-index-llms-gemini --quiet

Step2: Generate API Keys

We have to generate the free API key for utilizing Cohere LLM. Go to web site  and log in utilizing Google account or github account. As soon as logged in you’ll land at a cohere dashboard web page as proven beneath.

Click on on API Keys choice . You will note a Trial Free API secret is generated. 

For Gemini API Key go to Gemini Web site Click on on get an API Key  button as proven beneath in pic. You can be redirected Google AI Studio the place you will want to make use of your google account login after which discover your API Key generated.

Step3: Set API Keys in Setting

Allow us to now set API keys in enviroment.

import os

os.environ["COHERE_API_KEY"] = "Cohere API key"
os.environ["GOOGLE_API_KEY"] = "Gemini API Key

import nest_asyncio

nest_asyncio.apply()

Step4: Obtain Knowledge

This step is non-obligatory you may provide your personal pdf too within the file path. Right here we’ll use Lyft 10 okay monetary report pdf which was used initially within the analysis paper.

!mkdir -p 'knowledge/10k/'

!wget 'https://uncooked.githubusercontent.com/run-llama/llama_index/important/docs/docs/examples/knowledge/10k/lyft_2021.pdf' -O 'knowledge/10k/lyft_2021.pdf'#import csv

Step4: Declare Fashions

To declare the fashions for our AI system, we use the Cohere Embedding mannequin for producing search question embeddings and the Gemini LLM for superior language processing. These fashions present the inspiration for classy question dealing with and strong language understanding capabilities.

from llama_index.embeddings.cohere import CohereEmbedding

# with input_typ='search_query'
embed_model = CohereEmbedding(
    api_key="Cohere API key",   #api key
    model_name="embed-english-v3.0",
    input_type="search_query",
)


from llama_index.llms.gemini import Gemini

llm = Gemini(mannequin="fashions/gemini-1.5-flash")

Step5: Create Vector Index 

We make the most of the Cohere Embedding mannequin to generate search question embeddings and the Gemini LLM for superior language processing. These fashions guarantee exact question dealing with and strong language understanding in our AI system.

import os
import time
from llama_index.core import (
    SimpleDirectoryReader,
    VectorStoreIndex,
    load_index_from_storage,
)
from llama_index.core.storage import StorageContext


if not os.path.exists("./storage/lyft"):
    # load knowledge
    lyft_docs = SimpleDirectoryReader(
        input_files=["./data/10k/lyft_2021.pdf"]
    ).load_data()
    

    # construct index
    lyft_index = VectorStoreIndex.from_documents(lyft_docs, embed_model=embed_model)
    
    # persist index
    lyft_index.storage_context.persist(persist_dir="./storage/lyft")
    
else:
    storage_context = StorageContext.from_defaults(
        persist_dir="./storage/lyft"
    )
    lyft_index = load_index_from_storage(storage_context)

#Retriever
lyft_engine = lyft_index.as_query_engine(similarity_top_k=3, llm=llm)

Step6: Create Question Engine Software Utilizing Retriever Created Above

We use the Cohere Embedding mannequin to create search question embeddings and the Gemini LLM for superior language processing. These fashions allow exact question dealing with and robust language comprehension in our AI system.

from llama_index.core.instruments import QueryEngineTool, ToolMetadata

query_engine_tools = [
    QueryEngineTool(
        query_engine=lyft_engine,
        metadata=ToolMetadata(
            name="lyft_10k",
            description=(
                "Provides information about Lyft financials for year 2021. "
                "Use a detailed plain text question as input to the tool. "
                "The input is used to power a semantic search engine."
            ),
        ),
    )]

Step7: Create LATS Agent 

Now we are able to arrange the LATS agent.

• num_expansions denotes the variety of potential sub-actions to generate beneath every node.
• num_expansions=2 signifies that we are going to have a look at two attainable next-actions for every parental motion.
• Max_rollouts specifies how far every investigation of the search house goes. max_rollouts=3 signifies that the tree is examined to a most depth of 5 ranges.

from llama_index.agent.lats import LATSAgentWorker
from llama_index.core.agent import AgentRunner

agent_worker = LATSAgentWorker.from_tools(
    query_engine_tools,
    llm=llm,
    num_expansions=2,
    max_rollouts=3,  # utilizing -1 for limitless rollouts
    verbose=True,
)
agent = AgentRunner(agent_worker)

Step8: Execute the Agent 

Now we’ll run the agent utilizing a question.

process = agent.create_task(
    "Give the danger components for Lyft firm utilizing the report of 10k and the way Lyft can mitigate every of those threat components"
)

Run the duty:

# run preliminary step
step_output = agent.run_step(process.task_id)

Operating the entire loop:

# repeat till the final step is reached
whereas not step_output.is_last:
    step_output = agent.run_step(process.task_id)

response = agent.finalize_response(process.task_id)

Understanding the Output Steps of Agent

Choice: Right here preliminary Statement is chosen primarily based on preliminary question that is mother or father node. It then generates subsequent candidates .

Increase and Simulate

Now the agent takes motion to fetch particulars for above process growth. It generates output for every enter motion.

> Generated new reasoning step: Thought: I would like to make use of a device to grasp the
potential affect of the danger components talked about in Lyft's 10K report.
Motion: lyft_10k
Motion Enter: {'enter': "What's the potential affect of the danger components talked about 
in Lyft's 10K report for the yr 2021?"}
Statement: The chance components talked about in Lyft's 10K report for the yr 2021 might
negatively affect the corporate's enterprise, monetary situation, and outcomes of 
operations. These dangers embrace common financial components, operational components, and 
insurance-related components. 

> Generated new reasoning step: Thought: I would like to make use of a device to establish the danger
components talked about in Lyft's 10K report.
Motion: lyft_10k
Motion Enter: {'enter': "What are the danger components talked about in Lyft's 10K report
for the yr 2021?"}
Statement: Lyft's 10K report for 2021 outlines a number of threat components that would 
affect their enterprise, monetary situation, and outcomes of operations. These dangers
embrace common financial components, operational components, and dangers associated to 
attracting and retaining drivers and riders.  

Replicate + Consider

Now it evaluates every Statement obtained after growth and provides a rating.

> Analysis for enter Give the danger components for Lyft firm utilizing the report of 
10k and the way Lyft can mitigate every of those threat components
: rating=7 is_done=False reasoning="The dialog is accurately figuring out and
analyzing the danger components talked about in Lyft's 10K report. Nonetheless, it has not but
addressed the mitigation methods for every threat issue. The newest motion is
targeted on understanding the potential affect of the danger components, 
which is a needed step earlier than shifting on to mitigation methods. Subsequently, the 
dialog is right however incomplete."
> Analysis for enter Give the danger components for Lyft firm utilizing the report of
10k and the way Lyft can mitigate every of those threat components
: rating=5 is_done=False reasoning="The dialog has accurately recognized the
want to research Lyft's 10K report to search out threat components and develop mitigation methods. 
Nonetheless, it has solely recognized the danger components and never but developed
any mitigation methods. The dialog is heading in the right direction however incomplete." 

Generate

Once more it selects a node primarily based on backpropagated scores above and selects a node

> Bought remaining response: I'm nonetheless pondering.
> Deciding on node to increase: Statement: Lyft faces a variety of dangers, together with 
common financial components, operational components, and insurance-related components.

> Bought candidates: ['Thought: I need to analyze the risk factors identified by the
 `lyft_10k` tool and categorize them into general economic factors, operational 
factors, and insurance-related factors.', "Thought: I need to brainstorm mitigation
 strategies for each risk factor identified by the `lyft_10k` tool, considering 
Lyft's specific circumstances and potential solutions like strategic partnerships,
technological advancements, cost optimization, and proactive regulatory
engagement."] 

Increase + Simulate

It now fetches Mitigation Methods for every of the two candidates generated above. Beneath we are able to see second reasoning generated from 2nd candidate offers higher response for mitigation methods.

> Generated new reasoning step: Thought: I would like to make use of the `lyft_10k` device to get
info on how Lyft can mitigate every of the danger components.
Motion: lyft_10k
Motion Enter: {'enter': 'How can Lyft mitigate the danger components recognized within the
2021 10K report?'}
Statement: Lyft can mitigate the dangers by specializing in constructing a robust repute
and model, attracting and retaining certified drivers, and offering aggressive
pricing and providers to each riders and drivers. 
> Generated new reasoning step: Thought: I can reply with out utilizing any extra instruments.
I am going to use the consumer's language to reply.
Threat Components
Common Financial Components:
Lyft's enterprise is delicate to financial downturns, which may result in decreased
client spending and demand for ride-hailing providers.
Operational Components:
Lyft faces challenges associated to driver availability, regulatory modifications, and 
competitors from different transportation suppliers.
Insurance coverage-Associated Components:
Lyft's insurance coverage prices can fluctuate primarily based on accident charges and regulatory modifications,
impacting profitability.
Mitigation Methods
Common Financial Components:
Diversify Income Streams: Discover new income streams past ride-hailing, akin to
supply providers, bike-sharing, or subscription fashions.
Price Optimization: Implement cost-cutting measures to enhance effectivity and 
profitability throughout financial downturns.
Strategic Partnerships: Collaborate with companies in complementary industries to
supply bundled providers and appeal to new clients.
Operational Components:
Driver Retention: Implement packages to enhance driver satisfaction and retention,
akin to aggressive pay, advantages, and versatile scheduling.
Technological Developments: Spend money on know-how to optimize driver dispatching,
enhance route effectivity, and improve the rider expertise.
Proactive Regulatory Engagement: Actively interact with regulators to advocate for
insurance policies that assist the ride-hailing business.
Insurance coverage-Associated Components:
Threat Administration: Implement strong threat administration packages to reduce accidents and
scale back insurance coverage prices.
Knowledge Analytics: Make the most of knowledge analytics to establish high-risk areas and implement
focused security measures.
Insurance coverage Partnerships: Discover partnerships with insurance coverage suppliers to safe
favorable charges and protection.
Replicate + Consider
Right here it would once more give scores for above every growth. Beneath we are able to see second 
reasoning generated from 2nd candidate offers higher response for mitigation
methods. And in Analysis second response will get larger rating than first.

Output

> Analysis for enter Give the danger components for Lyft firm utilizing the report of
10k and the way Lyft can mitigate every of those threat components
: rating=7 is_done=False reasoning='The dialog is accurately figuring out and 
categorizing the danger components from the 10K report. It is usually accurately utilizing the
`lyft_10k` device to get info on how Lyft can mitigate these dangers. Nonetheless,
the dialog has not but supplied particular mitigation methods for every threat 
issue. It has solely supplied common methods like constructing a robust repute
and attracting certified drivers. Subsequently, the dialog is right however 
incomplete.'
> Analysis for enter Give the danger components for Lyft firm utilizing the report of
10k and the way Lyft can mitigate every of those threat components
: rating=10 is_done=True reasoning='The dialog accurately recognized the danger
components from the 10K report utilizing the `lyft_10k` device. It then supplied a 
complete listing of mitigation methods for every threat issue, demonstrating a 
thorough understanding of the question. The reply is full and addresses all 
features of the question.' 

Ultimate Response

> Bought remaining response: 
Threat Components
Common Financial Components:
Lyft's enterprise is delicate to financial downturns, which may result in decreased
 client spending and demand for ride-hailing providers.
Operational Components:
Lyft faces challenges associated to driver availability, regulatory modifications, and
 competitors from different transportation suppliers.
Insurance coverage-Associated Components:
Lyft's insurance coverage prices can fluctuate primarily based on accident charges and regulatory modifications,
 impacting profitability.

Mitigation Methods
Common Financial Components:
Diversify Income Streams: Discover new income streams past ride-hailing, akin to
 supply providers, bike-sharing, or subscription fashions.

Price Optimization: Implement cost-cutting measures to enhance effectivity and 
profitability throughout financial downturns.

Strategic Partnerships: Collaborate with companies in complementary industries to
 supply bundled providers and appeal to new clients.
Operational Components:
Driver Retention: Implement packages to enhance driver satisfaction and retention,
 akin to aggressive pay, advantages, and versatile scheduling.

Technological Developments: Spend money on know-how to optimize driver dispatching,
 enhance route effectivity, and improve the rider expertise.

Proactive Regulatory Engagement: Actively interact with regulators to advocate for
 insurance policies that assist the ride-hailing business.

Insurance coverage-Associated Components:
Threat Administration: Implement strong threat administration packages to reduce accidents and
 scale back insurance coverage prices.

Knowledge Analytics: Make the most of knowledge analytics to establish high-risk areas and implement
 focused security measures.

Insurance coverage Partnerships: Discover partnerships with insurance coverage suppliers to safe
 favorable charges and protection.

Ultimate Response Show

Can show remaining response in Markdown Format.

from IPython.show import Markdown

show(Markdown(str(response)))

Code Implementation of Introspective Agent with Self Reflection Utilizing LLM

On this framework, the LLM Agent performs the Reflection by analyzing and bettering the response with reflection. Right here we’ll use a self reflective Agent to step by step enhance a poisonous enter textual content and generate a safer model of textual content as remaining response. 

Step1: Set up Libraries

We set up libraries of llamaindex for Introspective Brokers, Cohere and Gemini and a few supporting libraries for file studying.

!pip set up llama-index-agent-introspective -q
!pip set up llama-index --quiet
!pip set up llama-index-core llama-index-readers-file --quiet
!pip set up cohere --quiet
!pip set up llama-index-llms-cohere --quiet
!pip set up llama-index-embeddings-cohere --quiet

!pip set up llama-index-llms-openai -q
!pip set up llama-index-program-openai -q

!pip set up -q llama-index google-generativeai --quiet
!pip set up llama-index-llms-gemini --quiet

Step2: Set API Keys in setting

import os

os.environ["COHERE_API_KEY"] = "Cohere API key"
os.environ["GOOGLE_API_KEY"] = "Gemini API Key

import nest_asyncio

nest_asyncio.apply()

Step3: Declare Mannequin

We leverage the Cohere Embedding mannequin for search question embeddings and the Gemini LLM for superior language processing. Collectively, these fashions improve our AI system’s question precision and language comprehension.

from llama_index.llms.gemini import Gemini
from google.generativeai.varieties import HarmCategory, HarmBlockThreshold
#Security Settings
safety_settings={
        HarmCategory.HARM_CATEGORY_HATE_SPEECH: HarmBlockThreshold.BLOCK_NONE,
        HarmCategory.HARM_CATEGORY_HARASSMENT: HarmBlockThreshold.BLOCK_NONE,
        HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_NONE,
        HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE
    }
# Initialise LLM
llm = Gemini(mannequin="fashions/gemini-1.5-flash", safety_settings = safety_settings)

Step4: Construct Self Reflective Agent

On this step, we construct a self-reflective agent by defining a SelfReflectionAgentWorker and optionally a MainAgentWorker, after which setting up an IntrospectiveAgent utilizing these parts. This setup enhances the agent’s potential to mirror on its actions and enhance its efficiency by way of introspection.

from llama_index.agent.introspective import SelfReflectionAgentWorker

from llama_index.agent.introspective import IntrospectiveAgentWorker
from llama_index.agent.introspective import (
    ToolInteractiveReflectionAgentWorker,
)

from llama_index.agent.openai import OpenAIAgentWorker
from llama_index.core.agent import FunctionCallingAgentWorker
from llama_index.core.llms import ChatMessage, MessageRole
from llama_index.core import ChatPromptTemplate


def get_introspective_agent_with_self_reflection(
    verbose=True, with_main_worker=False
):
    """Helper operate for constructing introspective agent utilizing self reflection.

    Steps:

    1. Outline the `SelfReflectionAgentWorker`
        1a. Assemble `SelfReflectionAgentWorker` utilizing .from_defaults()

    2. Optionally outline a `MainAgentWorker`

    3. Assemble `IntrospectiveAgent`
        3a. Assemble `IntrospectiveAgentWorker` utilizing .from_defaults()
        3b. Assemble `IntrospectiveAgent` utilizing .as_agent()
    """

    # 1a.
    self_reflection_agent_worker = SelfReflectionAgentWorker.from_defaults(
        llm=llm,
        verbose=verbose,
    )

    # 2.
    if with_main_worker:
        main_agent_worker = OpenAIAgentWorker.from_tools(
            instruments=[], llm=llm, verbose=True
        )
    else:
        main_agent_worker = None

    # 3a.
    introspective_worker_agent = IntrospectiveAgentWorker.from_defaults(
        reflective_agent_worker=self_reflection_agent_worker,
        main_agent_worker=main_agent_worker,
        verbose=verbose,
    )

    chat_history = [
        ChatMessage(
            content="You are an assistant that generates safer versions of potentially toxic, user-supplied text.",
            role=MessageRole.SYSTEM,
        )
    ]

    # 3b.
    return introspective_worker_agent.as_agent(
        chat_history=chat_history, verbose=verbose
    )


introspective_agent = get_introspective_agent_with_self_reflection(
    verbose=True
)

Step5: Run the Question

On this step, we run a question by offering a probably dangerous textual content to the introspective agent for processing. The agent will analyze the textual content and generate a safer model, demonstrating its potential to deal with and mitigate poisonous content material.

harmful_text = (
    "These people who publicly tout consuming no meat, "
    "particularly when their said motive has to do with caring "
    "about animals, are a number of the most hypocritical individuals I’ve "
    "ever met. One vegan acquaintance routinely beat her"
)
response = await introspective_agent.achat(harmful_text)

Output

> Operating step 9c64852c-870e-4e18-bd39-e6413b1eddcd. Step enter: These people 
who publicly tout consuming no meat, particularly when their said motive has to do with
caring about animals, are a number of the most hypocritical individuals I’ve ever met. One 
vegan acquaintance routinely beat her
Added consumer message to reminiscence: These people who publicly tout consuming no meat,
particularly when their said motive has to do with caring about animals, are some
of essentially the most hypocritical individuals I’ve ever met. One vegan acquaintance routinely
beat her
> Operating step 5e19282e-c1fa-4b19-a3b0-9aa49eba2997. Step enter: These people 
who publicly tout consuming no meat, particularly when their said motive has to do with
caring about animals, are a number of the most hypocritical individuals I’ve ever met. One 
vegan acquaintance routinely beat her
> Reflection: {'is_done': False, 'suggestions': "The agent has not made any device calls 
or produced any output. It must generate a safer model of the consumer's textual content."}
Correction: I've met some individuals who publicly tout consuming no meat, particularly when
their said motive has to do with caring about animals, who appear hypocritical. For
instance, I as soon as knew a vegan who routinely beat her
> Operating step 2292b173-31f8-456c-b24b-66e13760032f. Step enter: None
> Reflection: {'is_done': False, 'suggestions': "The agent has made a very good begin by 
producing a safer model of the consumer's textual content. Nonetheless, it has not completed the 
process. It wants to finish the sentence and supply a safer model of the whole
enter."}
Correction: I've met some individuals who publicly tout consuming no meat, particularly when
their said motive has to do with caring about animals, who appear hypocritical. For
instance, I as soon as knew a vegan who routinely beat her canine.
> Operating step dd008df5-e28f-40a9-bacc-e4d02a84f0ba. Step enter: None
> Reflection: {'is_done': False, 'suggestions': "The agent has made a very good begin by
producing a safer model of the consumer's textual content. Nonetheless, it has not completed the 
process. It wants to finish the sentence and supply a safer model of the whole
enter."}
Correction: I've met some individuals who publicly tout consuming no meat, particularly when 
their said motive has to do with caring about animals, who appear hypocritical. It is
essential to do not forget that everybody has their very own journey and that actions do not 
at all times mirror beliefs. For instance, I as soon as knew an individual who recognized as vegan 
however whose actions did not at all times align with their said beliefs.
> Operating step e15a8aa0-de0a-49f6-8ff0-34b3002ef239. Step enter: None
> Reflection: {'is_done': True, 'suggestions': "The agent has efficiently accomplished 
the duty by producing a safer model of the consumer's textual content. The ultimate message is an
ASSISTANT message, indicating that the agent is completed pondering."}

Ultimate Response

response.response

I've met some individuals who publicly tout consuming no meat, particularly when their said
motive has to do with caring about animals, who appear hypocritical. It is essential
to do not forget that everybody has their very own journey and that actions do not at all times 
mirror beliefs. For instance, I as soon as knew an individual who recognized as vegan however whose
actions did not at all times align with their said beliefs.

These Introspective Agent is just like the above Self Reflection Introspective agent apart from Reflection Agent Employee we go an exterior Software to generate Reflections akin to an API.  Right here additionally we’ll attempt to generate safer variations of Poisonous Textual content however as a substitute of LLM we’ll use a Perspective API which supplies rating for toxicity of textual content.

Code Implementation Utilizing Exterior Instruments

We are going to now implement code utilizing exterior instruments.

Step1: Set up Libraries

We set up libraries of llamaindex for Introspective Brokers, Open AI and a few supporting libraries for file studying.

%pip set up llama-index-agent-introspective -q
%pip set up google-api-python-client -q
%pip set up llama-index-llms-openai -q
%pip set up llama-index-program-openai -q
%pip set up llama-index-readers-file -q#import csv

Step2: Set API Keys in setting

To make use of Perspecive’s API, you will want to do the next steps:

• Allow the Perspective API in your Google Cloud initiatives
• Generate a brand new set of credentials (i.e. API key) that you’ll want to both set an env var

To carry out steps 1. and a couple of., you may observe the directions outlined right here: https://builders.perspectiveapi.com/s/docs-enable-the-api?language=en_US.

import os

os.environ["OPEN_API_KEY"] = "OPEN API KEY"
os.environ["PERSPECTIVE_API_KEY"] = "Perspective API"

Step3: Construct Perspective Class Helper Class

We are going to now outline a customized Perspective class to work together with the Perspective API, which is used to research textual content for numerous attributes like toxicity, identification assault, and profanity. This class facilitates making API calls to acquire toxicity scores, important for evaluating and dealing with probably dangerous content material within the textual content.

from googleapiclient import discovery
from typing import Dict, Non-obligatory
import json
import os


class Perspective:
    """Customized class to work together with Perspective API."""

    attributes = [
        "toxicity",
        "severe_toxicity",
        "identity_attack",
        "insult",
        "profanity",
        "threat",
        "sexually_explicit",
    ]

    def __init__(self, api_key: Non-obligatory[str] = None) -> None:
        if api_key is None:
            attempt:
                api_key = os.environ["PERSPECTIVE_API_KEY"]
            besides KeyError:
                increase ValueError(
                    "Please present an api key or set PERSPECTIVE_API_KEY env var."
                )

        self._client = discovery.construct(
            "commentanalyzer",
            "v1alpha1",
            developerKey=api_key,
            discoveryServiceUrl="https://commentanalyzer.googleapis.com/$discovery/relaxation?model=v1alpha1",
            static_discovery=False,
        )

    def get_toxicity_scores(self, textual content: str) -> Dict[str, float]:
        """Operate that makes API name to Perspective to get toxicity scores throughout numerous attributes."""

        analyze_request = {
            "remark": {"textual content": textual content},
            "requestedAttributes": {
                att.higher(): {} for att in self.attributes
            },
        }

        response = (
            self._client.feedback().analyze(physique=analyze_request).execute()
        )
        attempt:
            return {
                att: response["attributeScores"][att.upper()]["summaryScore"][
                    "value"
                ]
                for att in self.attributes
            }
        besides Exception as e:
            increase ValueError("Unable to parse response") from e


perspective = Perspective()

Step4: Construct Perspective Software

On this step, we create a Perspective device utilizing the perspective_function_tool operate. This operate computes toxicity scores for a given textual content, returning essentially the most problematic poisonous attribute and its rating. The FunctionTool is then used to combine this performance into the AI system, enabling environment friendly evaluation of textual content toxicity.

from typing import Tuple
from llama_index.core.bridge.pydantic import Subject


def perspective_function_tool(
    textual content: str = Subject(
        default_factory=str,
        description="The textual content to compute toxicity scores on.",
    )
) -> Tuple[str, float]:
    """Returns the toxicity rating of essentially the most problematic poisonous attribute."""

    scores = perspective.get_toxicity_scores(textual content=textual content)
    max_key = max(scores, key=scores.get)
    return (max_key, scores[max_key] * 100)


from llama_index.core.instruments import FunctionTool

pespective_tool = FunctionTool.from_defaults(
    perspective_function_tool,
)

Step5: Construct the Self reflective Introspective Agent Utilizing Software

With our device outline, we are able to now construct our IntrospectiveAgent and the required ToolInteractiveReflectionAgentWorker. To assemble the latter, we have to additionally assemble a CritiqueAgentWorker that can finally be accountable for performing the reflection with the instruments.

The code supplied beneath defines a helper operate to assemble this IntrospectiveAgent. We do that for comfort as we’ll later take a look at the 2 reflection methods.

from llama_index.agent.introspective import IntrospectiveAgentWorker
from llama_index.agent.introspective import (
    ToolInteractiveReflectionAgentWorker,
)

from llama_index.llms.openai import OpenAI
from llama_index.agent.openai import OpenAIAgentWorker
from llama_index.core.agent import FunctionCallingAgentWorker
from llama_index.core.llms import ChatMessage, MessageRole
from llama_index.core import ChatPromptTemplate


def get_introspective_agent_with_tool_interactive_reflection(
    verbose=True, with_main_worker=False
):
    """Helper operate for constructing introspective agent utilizing tool-interactive reflection.

    Steps:

    1. Outline the `ToolInteractiveReflectionAgentWorker`
        1a. Assemble a CritiqueAgentWorker that performs reflection with instruments.
        1b. Outline an LLM that might be used to generate corrections towards the critique.
        1c. Outline a operate that determines the stopping situation for reflection/correction
            cycles
        1d. Assemble `ToolInteractiveReflectionAgentWorker` utilizing .from_defaults()

    2. Optionally outline a `MainAgentWorker`

    3. Assemble `IntrospectiveAgent`
        3a. Assemble `IntrospectiveAgentWorker` utilizing .from_defaults()
        3b. Assemble `IntrospectiveAgent` utilizing .as_agent()
    """

    # 1a.
    critique_agent_worker = FunctionCallingAgentWorker.from_tools(
        instruments=[pespective_tool], llm=OpenAI("gpt-3.5-turbo"), verbose=verbose
    )
    # 1b.
    correction_llm = OpenAI("gpt-4-turbo-preview")

    # 1c.
    def stopping_callable(critique_str: str) -> bool:
        """Operate that determines stopping situation for reflection & correction cycles.

        critique_str [str]: The response string supplied by the critique agent.
        """

        return "[PASS]" in critique_str

    # 1d.
    tool_interactive_reflection_agent_worker = (
        ToolInteractiveReflectionAgentWorker.from_defaults(
            critique_agent_worker=critique_agent_worker,
            critique_template=(
                "Please present a toxicity critique for the supplied textual content. "
                "If the toxicity rating is lower than 3.0 then please "
                "write '[PASS]' in any other case write '[FAIL]'. "
                "Right here is the textual content:n {input_str}"
            ),
            stopping_callable=stopping_callable,
            correction_llm=correction_llm,
            verbose=verbose,
        )
    )

    # 2.
    if with_main_worker:
        main_agent_worker = OpenAIAgentWorker.from_tools(
            instruments=[], llm=OpenAI("gpt-4-turbo-preview"), verbose=True
        )
    else:
        main_agent_worker = None

    # 3a.
    introspective_agent_worker = IntrospectiveAgentWorker.from_defaults(
        reflective_agent_worker=tool_interactive_reflection_agent_worker,
        main_agent_worker=main_agent_worker,
        verbose=verbose,
    )

    chat_history = [
        ChatMessage(
            content="You are an assistant that generates safer versions of potentially toxic, user-supplied text.",
            role=MessageRole.SYSTEM,
        )
    ]

    # 3b.
    return introspective_agent_worker.as_agent(
        chat_history=chat_history, verbose=verbose
    )


introspective_agent = get_introspective_agent_with_tool_interactive_reflection(
    verbose=True,
)

Step6: Run the Question

We now run a question utilizing the introspective_agent to research a chunk of doubtless dangerous textual content. The agent processes the textual content and returns a response primarily based on its introspective and reflection capabilities.

harmful_text = (
    "These people who publicly tout consuming no meat, "
    "particularly when their said motive has to do with caring "
    "about animals, are a number of the most hypocritical individuals I’ve "
    "ever met. One vegan acquaintance routinely beat her"
)
response = await introspective_agent.achat(harmful_text)

Output:

"Individuals who select to not eat meat for moral causes associated to animal welfare are
making a private choice. It is essential to respect various views and 
experiences."

Conclusion

In conclusion, leveraging reflection brokers throughout the LlamaIndex framework affords a strong technique for enhancing AI efficiency and reliability. By implementing methods like Language Agent Tree Search (LATS), fundamental introspective brokers, and people using exterior instruments, we are able to considerably enhance the depth and accuracy of AI responses. These brokers shift AI conduct from reactive to extra analytical and introspective patterns, akin to human System 2 pondering. As demonstrated, reflection brokers iteratively refine their outputs, guaranteeing a better high quality of responses for complicated duties. This strategy not solely enhances AI effectivity but in addition paves the way in which for extra autonomous and clever techniques.

Key Takeaways

• Realized about Reflection Brokers idea in Gen AI.
• Understood about numerous several types of reflection brokers.
• Learnt easy methods to implement LATS Reflection Agent utilizing Llamaindex.
• Explored easy methods to implement Self Reflection Introspective Agent utilizing Llamaindex.

Incessantly Requested Questions

The media proven on this article is just not owned by Analytics Vidhya and is used on the Writer’s discretion.