Geração Aumentada por Recuperação (RAG) com Milvus e Camel

Este guia demonstra como construir um sistema RAG (Retrieval-Augmented Generation) utilizando o CAMEL e o Milvus.

O sistema RAG combina um sistema de recuperação com um modelo generativo para gerar novo texto com base num determinado pedido. O sistema começa por recuperar documentos relevantes de um corpus utilizando o Milvus e, em seguida, utiliza um modelo generativo para gerar novo texto com base nos documentos recuperados.

CAMEL é uma estrutura multi-agente. O Milvus é a base de dados vetorial de código aberto mais avançada do mundo, criada para alimentar a pesquisa de semelhanças de incorporação e as aplicações de IA.

Neste caderno, mostramos a utilização do módulo CAMEL Retrieve tanto de forma personalizada como de forma automática. Também mostraremos como combinar AutoRetriever com ChatAgent, e ainda combinar AutoRetriever com RolePlaying usando Function Calling.

4 partes principais incluídas:

• RAG personalizado
• RAG automático
• Agente único com RAG automático
• Interpretação de papéis com o Auto RAG

Carregar dados

Vamos primeiro carregar o documento CAMEL de https://arxiv.org/pdf/2303.17760.pdf. Estes serão os nossos dados de exemplo locais.

$ pip install -U "camel-ai[all]" pymilvus

import os
import requests

os.makedirs("local_data", exist_ok=True)

url = "https://arxiv.org/pdf/2303.17760.pdf"
response = requests.get(url)
with open("local_data/camel paper.pdf", "wb") as file:
    file.write(response.content)

1. RAG personalizado

Nesta secção, vamos definir o nosso pipeline RAG personalizado, vamos utilizar VectorRetriever como exemplo. Vamos definir OpenAIEmbedding como o modelo de incorporação e MilvusStorage como o armazenamento para ele.

Para definir a incorporação do OpenAI, precisamos definir o OPENAI_API_KEY abaixo.

os.environ["OPENAI_API_KEY"] = "Your Key"

Importar e definir a instância de incorporação:

from camel.embeddings import OpenAIEmbedding

embedding_instance = OpenAIEmbedding()

Importar e definir a instância de armazenamento do vetor:

from camel.storages import MilvusStorage

storage_instance = MilvusStorage(
    vector_dim=embedding_instance.get_output_dim(),
    url_and_api_key=(
        "./milvus_demo.db",  # Your Milvus connection URI
        "",  # Your Milvus token
    ),
    collection_name="camel_paper",
)

Importar e definir a instância do retriever:

Por defeito, o similarity_threshold está definido para 0,75. Pode alterá-lo.

from camel.retrievers import VectorRetriever

vector_retriever = VectorRetriever(
    embedding_model=embedding_instance, storage=storage_instance
)

Utilizamos o Unstructured Module integrado para dividir o conteúdo em pequenos pedaços, o conteúdo será dividido automaticamente com a sua função chunk_by_title, o carácter máximo para cada pedaço é de 500 caracteres, que é um comprimento adequado para OpenAIEmbedding. Todo o texto nos pedaços será incorporado e armazenado na instância de armazenamento de vectores, o que levará algum tempo, por favor aguarde.

vector_retriever.process(content_input_path="local_data/camel paper.pdf")

[nltk_data] Downloading package punkt to /root/nltk_data...
[nltk_data]   Unzipping tokenizers/punkt.zip.
[nltk_data] Downloading package averaged_perceptron_tagger to
[nltk_data]     /root/nltk_data...
[nltk_data]   Unzipping taggers/averaged_perceptron_tagger.zip.

Agora podemos obter informações do armazenamento vetorial através de uma consulta. Por predefinição, será devolvido o conteúdo de texto do primeiro bloco com a pontuação de semelhança de cosseno mais elevada, e a pontuação de semelhança deve ser superior a 0,75 para garantir que o conteúdo recuperado é relevante para a consulta. Também é possível alterar o valor de top_k.

A lista de cadeias de caracteres devolvida inclui:

• pontuação de similaridade
• caminho do conteúdo
• metadados
• texto

retrieved_info = vector_retriever.query(query="What is CAMEL?", top_k=1)
print(retrieved_info)

[{'similarity score': '0.8321675658226013', 'content path': 'local_data/camel paper.pdf', 'metadata': {'last_modified': '2024-04-19T14:40:00', 'filetype': 'application/pdf', 'page_number': 45}, 'text': 'CAMEL Data and Code License The intended purpose and licensing of CAMEL is solely for research use. The source code is licensed under Apache 2.0. The datasets are licensed under CC BY NC 4.0, which permits only non-commercial usage. It is advised that any models trained using the dataset should not be utilized for anything other than research purposes.\n\n45'}]

Vamos tentar uma consulta irrelevante:

retrieved_info_irrelevant = vector_retriever.query(
    query="Compared with dumpling and rice, which should I take for dinner?", top_k=1
)

print(retrieved_info_irrelevant)

[{'text': 'No suitable information retrieved from local_data/camel paper.pdf                 with similarity_threshold = 0.75.'}]

2. RAG automático

Nesta secção, vamos executar o AutoRetriever com as definições predefinidas. Utiliza OpenAIEmbedding como modelo de incorporação predefinido e Milvus como armazenamento vetorial predefinido.

O que você precisa fazer é:

• Definir caminhos de entrada de conteúdo, que podem ser caminhos locais ou URLs remotos
• Definir o URL remoto e a chave da API para o Milvus
• Fornecer uma consulta

O pipeline Auto RAG criará colecções para os caminhos de entrada de conteúdos indicados. O nome da coleção será definido automaticamente com base no nome do caminho de entrada de conteúdos e, se a coleção existir, será feita a recuperação diretamente.

from camel.retrievers import AutoRetriever
from camel.types import StorageType

auto_retriever = AutoRetriever(
    url_and_api_key=(
        "./milvus_demo.db",  # Your Milvus connection URI
        "",  # Your Milvus token
    ),
    storage_type=StorageType.MILVUS,
    embedding_model=embedding_instance,
)

retrieved_info = auto_retriever.run_vector_retriever(
    query="What is CAMEL-AI",
    content_input_paths=[
        "local_data/camel paper.pdf",  # example local path
        "https://www.camel-ai.org/",  # example remote url
    ],
    top_k=1,
    return_detailed_info=True,
)

print(retrieved_info)

Original Query:
{What is CAMEL-AI}
Retrieved Context:
{'similarity score': '0.8252888321876526', 'content path': 'local_data/camel paper.pdf', 'metadata': {'last_modified': '2024-04-19T14:40:00', 'filetype': 'application/pdf', 'page_number': 7}, 'text': ' Section 3.2, to simulate assistant-user cooperation. For our analysis, we set our attention on AI Society setting. We also gathered conversational data, named CAMEL AI Society and CAMEL Code datasets and problem-solution pairs data named CAMEL Math and CAMEL Science and analyzed and evaluated their quality. Moreover, we will discuss potential extensions of our framework and highlight both the risks and opportunities that future AI society might present.'}
{'similarity score': '0.8378663659095764', 'content path': 'https://www.camel-ai.org/', 'metadata': {'filetype': 'text/html', 'languages': ['eng'], 'page_number': 1, 'url': 'https://www.camel-ai.org/', 'link_urls': ['#h.3f4tphhd9pn8', 'https://join.slack.com/t/camel-ai/shared_invite/zt-2g7xc41gy-_7rcrNNAArIP6sLQqldkqQ', 'https://discord.gg/CNcNpquyDc'], 'link_texts': [None, None, None], 'emphasized_text_contents': ['Mission', 'CAMEL-AI.org', 'is an open-source community dedicated to the study of autonomous and communicative agents. We believe that studying these agents on a large scale offers valuable insights into their behaviors, capabilities, and potential risks. To facilitate research in this field, we provide, implement, and support various types of agents, tasks, prompts, models, datasets, and simulated environments.', 'Join us via', 'Slack', 'Discord', 'or'], 'emphasized_text_tags': ['span', 'span', 'span', 'span', 'span', 'span', 'span']}, 'text': 'Mission\n\nCAMEL-AI.org is an open-source community dedicated to the study of autonomous and communicative agents. We believe that studying these agents on a large scale offers valuable insights into their behaviors, capabilities, and potential risks. To facilitate research in this field, we provide, implement, and support various types of agents, tasks, prompts, models, datasets, and simulated environments.\n\nJoin us via\n\nSlack\n\nDiscord\n\nor'}

3. Agente único com RAG automático

Nesta secção, mostraremos como combinar o AutoRetriever com um ChatAgent.

Vamos definir uma função de agente, nesta função podemos obter a resposta fornecendo uma consulta a este agente.

from camel.agents import ChatAgent
from camel.messages import BaseMessage
from camel.types import RoleType
from camel.retrievers import AutoRetriever
from camel.types import StorageType


def single_agent(query: str) -> str:
    # Set agent role
    assistant_sys_msg = BaseMessage(
        role_name="Assistant",
        role_type=RoleType.ASSISTANT,
        meta_dict=None,
        content="""You are a helpful assistant to answer question,
         I will give you the Original Query and Retrieved Context,
        answer the Original Query based on the Retrieved Context,
        if you can't answer the question just say I don't know.""",
    )

    # Add auto retriever
    auto_retriever = AutoRetriever(
        url_and_api_key=(
            "./milvus_demo.db",  # Your Milvus connection URI
            "",  # Your Milvus token
        ),
        storage_type=StorageType.MILVUS,
        embedding_model=embedding_instance,
    )

    retrieved_info = auto_retriever.run_vector_retriever(
        query=query,
        content_input_paths=[
            "local_data/camel paper.pdf",  # example local path
            "https://www.camel-ai.org/",  # example remote url
        ],
        # vector_storage_local_path="storage_default_run",
        top_k=1,
        return_detailed_info=True,
    )

    # Pass the retrieved infomation to agent
    user_msg = BaseMessage.make_user_message(role_name="User", content=retrieved_info)
    agent = ChatAgent(assistant_sys_msg)

    # Get response
    assistant_response = agent.step(user_msg)
    return assistant_response.msg.content


print(single_agent("What is CAMEL-AI"))

CAMEL-AI is an open-source community dedicated to the study of autonomous and communicative agents. It provides, implements, and supports various types of agents, tasks, prompts, models, datasets, and simulated environments to facilitate research in this field.

4. Interpretação de papéis com o Auto RAG

Nesta secção, mostraremos como combinar o RETRIEVAL_FUNCS com o RolePlaying aplicando o Function Calling.

from typing import List
from colorama import Fore

from camel.agents.chat_agent import FunctionCallingRecord
from camel.configs import ChatGPTConfig
from camel.functions import (
    MATH_FUNCS,
    RETRIEVAL_FUNCS,
)
from camel.societies import RolePlaying
from camel.types import ModelType
from camel.utils import print_text_animated


def role_playing_with_rag(
    task_prompt, model_type=ModelType.GPT_4O, chat_turn_limit=10
) -> None:
    task_prompt = task_prompt

    user_model_config = ChatGPTConfig(temperature=0.0)

    function_list = [
        *MATH_FUNCS,
        *RETRIEVAL_FUNCS,
    ]
    assistant_model_config = ChatGPTConfig(
        tools=function_list,
        temperature=0.0,
    )

    role_play_session = RolePlaying(
        assistant_role_name="Searcher",
        user_role_name="Professor",
        assistant_agent_kwargs=dict(
            model_type=model_type,
            model_config=assistant_model_config,
            tools=function_list,
        ),
        user_agent_kwargs=dict(
            model_type=model_type,
            model_config=user_model_config,
        ),
        task_prompt=task_prompt,
        with_task_specify=False,
    )

    print(
        Fore.GREEN
        + f"AI Assistant sys message:\n{role_play_session.assistant_sys_msg}\n"
    )
    print(Fore.BLUE + f"AI User sys message:\n{role_play_session.user_sys_msg}\n")

    print(Fore.YELLOW + f"Original task prompt:\n{task_prompt}\n")
    print(
        Fore.CYAN
        + f"Specified task prompt:\n{role_play_session.specified_task_prompt}\n"
    )
    print(Fore.RED + f"Final task prompt:\n{role_play_session.task_prompt}\n")

    n = 0
    input_msg = role_play_session.init_chat()
    while n < chat_turn_limit:
        n += 1
        assistant_response, user_response = role_play_session.step(input_msg)

        if assistant_response.terminated:
            print(
                Fore.GREEN
                + (
                    "AI Assistant terminated. Reason: "
                    f"{assistant_response.info['termination_reasons']}."
                )
            )
            break
        if user_response.terminated:
            print(
                Fore.GREEN
                + (
                    "AI User terminated. "
                    f"Reason: {user_response.info['termination_reasons']}."
                )
            )
            break

        # Print output from the user
        print_text_animated(Fore.BLUE + f"AI User:\n\n{user_response.msg.content}\n")

        # Print output from the assistant, including any function
        # execution information
        print_text_animated(Fore.GREEN + "AI Assistant:")
        tool_calls: List[FunctionCallingRecord] = assistant_response.info["tool_calls"]
        for func_record in tool_calls:
            print_text_animated(f"{func_record}")
        print_text_animated(f"{assistant_response.msg.content}\n")

        if "CAMEL_TASK_DONE" in user_response.msg.content:
            break

        input_msg = assistant_response.msg

Execute a representação de papéis com a função de recuperação definida:

role_playing_with_rag(
    task_prompt="""What is the main termination reasons for AI Society
                   dataset, how many number of messages did camel decided to
                   limit, what's the value plus 100? You should refer to the
                   content in path camel/local_data/camel paper.pdf"""
)

AI Assistant sys message:
BaseMessage(role_name='Searcher', role_type=<RoleType.ASSISTANT: 'assistant'>, meta_dict={'task': "What is the main termination reasons for AI Society dataset, how many number of messages did camel decided to limit, what's the value plus 100?", 'assistant_role': 'Searcher', 'user_role': 'Professor'}, content="===== RULES OF ASSISTANT =====\nNever forget you are a Searcher and I am a Professor. Never flip roles! Never instruct me!\nWe share a common interest in collaborating to successfully complete a task.\nYou must help me to complete the task.\nHere is the task: What is the main termination reasons for AI Society dataset, how many number of messages did camel decided to limit, what's the value plus 100?. Never forget our task!\nI must instruct you based on your expertise and my needs to complete the task.\n\nI must give you one instruction at a time.\nYou must write a specific solution that appropriately solves the requested instruction and explain your solutions.\nYou must decline my instruction honestly if you cannot perform the instruction due to physical, moral, legal reasons or your capability and explain the reasons.\nUnless I say the task is completed, you should always start with:\n\nSolution: <YOUR_SOLUTION>\n\n<YOUR_SOLUTION> should be very specific, include detailed explanations and provide preferable detailed implementations and examples and lists for task-solving.\nAlways end <YOUR_SOLUTION> with: Next request.")

AI User sys message:
BaseMessage(role_name='Professor', role_type=<RoleType.USER: 'user'>, meta_dict={'task': "What is the main termination reasons for AI Society dataset, how many number of messages did camel decided to limit, what's the value plus 100?", 'assistant_role': 'Searcher', 'user_role': 'Professor'}, content='===== RULES OF USER =====\nNever forget you are a Professor and I am a Searcher. Never flip roles! You will always instruct me.\nWe share a common interest in collaborating to successfully complete a task.\nI must help you to complete the task.\nHere is the task: What is the main termination reasons for AI Society dataset, how many number of messages did camel decided to limit, what\'s the value plus 100?. Never forget our task!\nYou must instruct me based on my expertise and your needs to solve the task ONLY in the following two ways:\n\n1. Instruct with a necessary input:\nInstruction: <YOUR_INSTRUCTION>\nInput: <YOUR_INPUT>\n\n2. Instruct without any input:\nInstruction: <YOUR_INSTRUCTION>\nInput: None\n\nThe "Instruction" describes a task or question. The paired "Input" provides further context or information for the requested "Instruction".\n\nYou must give me one instruction at a time.\nI must write a response that appropriately solves the requested instruction.\nI must decline your instruction honestly if I cannot perform the instruction due to physical, moral, legal reasons or my capability and explain the reasons.\nYou should instruct me not ask me questions.\nNow you must start to instruct me using the two ways described above.\nDo not add anything else other than your instruction and the optional corresponding input!\nKeep giving me instructions and necessary inputs until you think the task is completed.\nWhen the task is completed, you must only reply with a single word <CAMEL_TASK_DONE>.\nNever say <CAMEL_TASK_DONE> unless my responses have solved your task.')

Original task prompt:
What is the main termination reasons for AI Society dataset, how many number of messages did camel decided to limit, what's the value plus 100?

Specified task prompt:
None

Final task prompt:
What is the main termination reasons for AI Society dataset, how many number of messages did camel decided to limit, what's the value plus 100?



AI User:

Instruction: Provide a summary of the main termination reasons in the AI Society dataset.
Input: None


AI Assistant:

Function Execution: local_retriever
    Args: {'query': 'main termination reasons for AI Society dataset'}
    Result: Original Query:
{main termination reasons for AI Society dataset}
Retrieved Context:
Next we examine the conversation termination reasons for both AI Society and Code datasets. As can be seen in Figure 8, the main termination reasons for AI Society dataset is Assistant Instruct whereas for Code it is Token Limit. The latter is expected as the since responses that contain code tend to be long. It is also interesting to note that in both datasets, the termination due to Maximum Number of Messages is low indicating that the limit of 40 maximum messages is reasonable. Our decision t

Solution: The main termination reason for the AI Society dataset is "Assistant Instruct." This indicates that conversations in the AI Society dataset typically end when the assistant is instructed to terminate the conversation.

Next request.


AI User:

Instruction: Identify the number of messages that the camel decided to limit in the context provided.
Input: None


AI Assistant:

Solution: The context provided from the local retriever indicates that the camel decided to limit the number of messages to a maximum of 40 messages. This is mentioned in the retrieved context where it states that the termination due to Maximum Number of Messages is low, indicating that the limit of 40 maximum messages is reasonable.

Next request.


AI User:

Instruction: Calculate the value of the message limit plus 100.
Input: None


AI Assistant:

Function Execution: add
    Args: {'a': 40, 'b': 100}
    Result: 140

Solution: The value of the message limit plus 100 is 140.

Next request.


AI User:

CAMEL_TASK_DONE


AI Assistant:

Solution: Understood, the task is completed.

Next request.