Construir RAG com Milvus e Crawl4AI

O Crawl4AI oferece rastreamento da Web extremamente rápido e pronto para IA para LLMs. De código aberto e optimizado para RAG, simplifica o scraping com extração avançada e desempenho em tempo real.

Neste tutorial, mostraremos como construir um pipeline Retrieval-Augmented Generation (RAG) usando Milvus e Crawl4AI. O pipeline integra o Crawl4AI para rastreio de dados da Web, o Milvus para armazenamento de vectores e o OpenAI para gerar respostas perspicazes e sensíveis ao contexto.

Preparação

Dependências e ambiente

Para começar, instale as dependências necessárias executando o seguinte comando:

$ pip install -U crawl4ai pymilvus openai requests tqdm

Se estiver a utilizar o Google Colab, para ativar as dependências acabadas de instalar, poderá ser necessário reiniciar o tempo de execução (clique no menu "Tempo de execução" na parte superior do ecrã e selecione "Reiniciar sessão" no menu pendente).

Para configurar completamente o crawl4ai, execute os seguintes comandos:

# Run post-installation setup
$ crawl4ai-setup

# Verify installation
$ crawl4ai-doctor

[INIT].... → Running post-installation setup...
[INIT].... → Installing Playwright browsers...
[COMPLETE] ● Playwright installation completed successfully.
[INIT].... → Starting database initialization...
[COMPLETE] ● Database initialization completed successfully.
[COMPLETE] ● Post-installation setup completed!
[INIT].... → Running Crawl4AI health check...
[INIT].... → Crawl4AI 0.4.247
[TEST].... ℹ Testing crawling capabilities...
[EXPORT].. ℹ Exporting PDF and taking screenshot took 0.80s
[FETCH]... ↓ https://crawl4ai.com... | Status: True | Time: 4.22s
[SCRAPE].. ◆ Processed https://crawl4ai.com... | Time: 14ms
[COMPLETE] ● https://crawl4ai.com... | Status: True | Total: 4.23s
[COMPLETE] ● ✅ Crawling test passed!


Configurar a chave da API do OpenAI

Neste exemplo, usaremos o OpenAI como LLM. Deve preparar a OPENAI_API_KEY como uma variável de ambiente.

import os

os.environ["OPENAI_API_KEY"] = "sk-***********"

Preparar o LLM e o modelo de incorporação

Inicializamos o cliente OpenAI para preparar o modelo de incorporação.

from openai import OpenAI

openai_client = OpenAI()

Defina uma função para gerar texto incorporado utilizando o cliente OpenAI. Utilizamos o modelo text-embedding-3-small como exemplo.

def emb_text(text):
    return (
        openai_client.embeddings.create(input=text, model="text-embedding-3-small")
        .data[0]
        .embedding
    )

Gerar um embedding de teste e imprimir a sua dimensão e os primeiros elementos.

test_embedding = emb_text("This is a test")
embedding_dim = len(test_embedding)
print(embedding_dim)
print(test_embedding[:10])

1536
[0.009889289736747742, -0.005578675772994757, 0.00683477520942688, -0.03805781528353691, -0.01824733428657055, -0.04121600463986397, -0.007636285852640867, 0.03225184231996536, 0.018949154764413834, 9.352207416668534e-05]

Rastrear dados usando Crawl4AI

from crawl4ai import *


async def crawl():
    async with AsyncWebCrawler() as crawler:
        result = await crawler.arun(
            url="https://lilianweng.github.io/posts/2023-06-23-agent/",
        )
        return result.markdown


markdown_content = await crawl()

[INIT].... → Crawl4AI 0.4.247
[FETCH]... ↓ https://lilianweng.github.io/posts/2023-06-23-agen... | Status: True | Time: 0.07s
[COMPLETE] ● https://lilianweng.github.io/posts/2023-06-23-agen... | Status: True | Total: 0.08s

Processar o conteúdo rastreado

Para tornar o conteúdo recolhido gerível para inserção no Milvus, basta usar "#" para separar o conteúdo, o que pode separar aproximadamente o conteúdo de cada parte principal do ficheiro markdown recolhido.

def split_markdown_content(content):
    return [section.strip() for section in content.split("# ") if section.strip()]


# Process the crawled markdown content
sections = split_markdown_content(markdown_content)

# Print the first few sections to understand the structure
for i, section in enumerate(sections[:3]):
    print(f"Section {i+1}:")
    print(section[:300] + "...")
    print("-" * 50)

Section 1:
[Lil'Log](https://lilianweng.github.io/posts/2023-06-23-agent/<https:/lilianweng.github.io/> "Lil'Log \(Alt + H\)")
  * |


  * [ Posts ](https://lilianweng.github.io/posts/2023-06-23-agent/<https:/lilianweng.github.io/> "Posts")
  * [ Archive ](https://lilianweng.github.io/posts/2023-06-23-agent/<h...
--------------------------------------------------
Section 2:
LLM Powered Autonomous Agents 
Date: June 23, 2023 | Estimated Reading Time: 31 min | Author: Lilian Weng 
Table of Contents
  * [Agent System Overview](https://lilianweng.github.io/posts/2023-06-23-agent/<#agent-system-overview>)
  * [Component One: Planning](https://lilianweng.github.io/posts/2023...
--------------------------------------------------
Section 3:
Agent System Overview[#](https://lilianweng.github.io/posts/2023-06-23-agent/<#agent-system-overview>)
In a LLM-powered autonomous agent system, LLM functions as the agent’s brain, complemented by several key components:
  * **Planning**
    * Subgoal and decomposition: The agent breaks down large t...
--------------------------------------------------

Carregar dados no Milvus

Criar a coleção

from pymilvus import MilvusClient

milvus_client = MilvusClient(uri="./milvus_demo.db")
collection_name = "my_rag_collection"

INFO:numexpr.utils:Note: NumExpr detected 10 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 8.
INFO:numexpr.utils:NumExpr defaulting to 8 threads.

Verificar se a coleção já existe e eliminá-la se existir.

if milvus_client.has_collection(collection_name):
    milvus_client.drop_collection(collection_name)

Criar uma nova coleção com os parâmetros especificados.

Se não especificarmos qualquer informação de campo, o Milvus criará automaticamente um campo id por defeito para a chave primária e um campo vector para armazenar os dados vectoriais. Um campo JSON reservado é utilizado para armazenar campos não definidos pelo esquema e os respectivos valores.

milvus_client.create_collection(
    collection_name=collection_name,
    dimension=embedding_dim,
    metric_type="IP",  # Inner product distance
    consistency_level="Strong",  # Strong consistency level
)

Inserir dados

from tqdm import tqdm

data = []
for i, section in enumerate(tqdm(sections, desc="Processing sections")):
    embedding = emb_text(section)
    data.append({"id": i, "vector": embedding, "text": section})

# Insert data into Milvus
milvus_client.insert(collection_name=collection_name, data=data)

Processing sections:   0%|          | 0/18 [00:00<?, ?it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:   6%|▌         | 1/18 [00:00<00:12,  1.37it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  11%|█         | 2/18 [00:01<00:11,  1.39it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  17%|█▋        | 3/18 [00:02<00:10,  1.40it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  22%|██▏       | 4/18 [00:02<00:07,  1.85it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  28%|██▊       | 5/18 [00:02<00:06,  2.06it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  33%|███▎      | 6/18 [00:03<00:06,  1.94it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  39%|███▉      | 7/18 [00:03<00:05,  2.14it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  44%|████▍     | 8/18 [00:04<00:04,  2.29it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  50%|█████     | 9/18 [00:04<00:04,  2.20it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  56%|█████▌    | 10/18 [00:05<00:03,  2.09it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  61%|██████    | 11/18 [00:06<00:04,  1.68it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  67%|██████▋   | 12/18 [00:06<00:04,  1.48it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  72%|███████▏  | 13/18 [00:07<00:02,  1.75it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  78%|███████▊  | 14/18 [00:07<00:01,  2.02it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  83%|████████▎ | 15/18 [00:07<00:01,  2.12it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  89%|████████▉ | 16/18 [00:08<00:01,  1.61it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections:  94%|█████████▍| 17/18 [00:09<00:00,  1.92it/s]INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
Processing sections: 100%|██████████| 18/18 [00:09<00:00,  1.83it/s]





{'insert_count': 18, 'ids': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17], 'cost': 0}

Criar RAG

Recuperar dados para uma consulta

Vamos especificar uma pergunta de consulta sobre o sítio Web que acabámos de rastrear.

question = "What are the main components of autonomous agents?"

Pesquise a pergunta na coleção e recupere as 3 principais correspondências semânticas.

search_res = milvus_client.search(
    collection_name=collection_name,
    data=[emb_text(question)],
    limit=3,
    search_params={"metric_type": "IP", "params": {}},
    output_fields=["text"],
)

INFO:httpx:HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"

Vamos dar uma vista de olhos nos resultados da pesquisa da consulta

import json

retrieved_lines_with_distances = [
    (res["entity"]["text"], res["distance"]) for res in search_res[0]
]
print(json.dumps(retrieved_lines_with_distances, indent=4))

[
    [
        "Agent System Overview[#](https://lilianweng.github.io/posts/2023-06-23-agent/<#agent-system-overview>)\nIn a LLM-powered autonomous agent system, LLM functions as the agent\u2019s brain, complemented by several key components:\n  * **Planning**\n    * Subgoal and decomposition: The agent breaks down large tasks into smaller, manageable subgoals, enabling efficient handling of complex tasks.\n    * Reflection and refinement: The agent can do self-criticism and self-reflection over past actions, learn from mistakes and refine them for future steps, thereby improving the quality of final results.\n  * **Memory**\n    * Short-term memory: I would consider all the in-context learning (See [Prompt Engineering](https://lilianweng.github.io/posts/2023-06-23-agent/<https:/lilianweng.github.io/posts/2023-03-15-prompt-engineering/>)) as utilizing short-term memory of the model to learn.\n    * Long-term memory: This provides the agent with the capability to retain and recall (infinite) information over extended periods, often by leveraging an external vector store and fast retrieval.\n  * **Tool use**\n    * The agent learns to call external APIs for extra information that is missing from the model weights (often hard to change after pre-training), including current information, code execution capability, access to proprietary information sources and more.\n\n![](https://lilianweng.github.io/posts/2023-06-23-agent/agent-overview.png) Fig. 1. Overview of a LLM-powered autonomous agent system.",
        0.6433743238449097
    ],
    [
        "LLM Powered Autonomous Agents \nDate: June 23, 2023 | Estimated Reading Time: 31 min | Author: Lilian Weng \nTable of Contents\n  * [Agent System Overview](https://lilianweng.github.io/posts/2023-06-23-agent/<#agent-system-overview>)\n  * [Component One: Planning](https://lilianweng.github.io/posts/2023-06-23-agent/<#component-one-planning>)\n    * [Task Decomposition](https://lilianweng.github.io/posts/2023-06-23-agent/<#task-decomposition>)\n    * [Self-Reflection](https://lilianweng.github.io/posts/2023-06-23-agent/<#self-reflection>)\n  * [Component Two: Memory](https://lilianweng.github.io/posts/2023-06-23-agent/<#component-two-memory>)\n    * [Types of Memory](https://lilianweng.github.io/posts/2023-06-23-agent/<#types-of-memory>)\n    * [Maximum Inner Product Search (MIPS)](https://lilianweng.github.io/posts/2023-06-23-agent/<#maximum-inner-product-search-mips>)\n  * [Component Three: Tool Use](https://lilianweng.github.io/posts/2023-06-23-agent/<#component-three-tool-use>)\n  * [Case Studies](https://lilianweng.github.io/posts/2023-06-23-agent/<#case-studies>)\n    * [Scientific Discovery Agent](https://lilianweng.github.io/posts/2023-06-23-agent/<#scientific-discovery-agent>)\n    * [Generative Agents Simulation](https://lilianweng.github.io/posts/2023-06-23-agent/<#generative-agents-simulation>)\n    * [Proof-of-Concept Examples](https://lilianweng.github.io/posts/2023-06-23-agent/<#proof-of-concept-examples>)\n  * [Challenges](https://lilianweng.github.io/posts/2023-06-23-agent/<#challenges>)\n  * [Citation](https://lilianweng.github.io/posts/2023-06-23-agent/<#citation>)\n  * [References](https://lilianweng.github.io/posts/2023-06-23-agent/<#references>)\n\n\nBuilding agents with LLM (large language model) as its core controller is a cool concept. Several proof-of-concepts demos, such as [AutoGPT](https://lilianweng.github.io/posts/2023-06-23-agent/<https:/github.com/Significant-Gravitas/Auto-GPT>), [GPT-Engineer](https://lilianweng.github.io/posts/2023-06-23-agent/<https:/github.com/AntonOsika/gpt-engineer>) and [BabyAGI](https://lilianweng.github.io/posts/2023-06-23-agent/<https:/github.com/yoheinakajima/babyagi>), serve as inspiring examples. The potentiality of LLM extends beyond generating well-written copies, stories, essays and programs; it can be framed as a powerful general problem solver.",
        0.5462194085121155
    ],
    [
        "Component One: Planning[#](https://lilianweng.github.io/posts/2023-06-23-agent/<#component-one-planning>)\nA complicated task usually involves many steps. An agent needs to know what they are and plan ahead.\n#",
        0.5223420858383179
    ]
]

Utilizar o LLM para obter uma resposta RAG

Converter os documentos recuperados num formato de cadeia de caracteres.

context = "\n".join(
    [line_with_distance[0] for line_with_distance in retrieved_lines_with_distances]
)

Definir avisos do sistema e do utilizador para o Modelo de Linguagem. Este prompt é montado com os documentos recuperados do Milvus.

SYSTEM_PROMPT = """
Human: You are an AI assistant. You are able to find answers to the questions from the contextual passage snippets provided.
"""
USER_PROMPT = f"""
Use the following pieces of information enclosed in <context> tags to provide an answer to the question enclosed in <question> tags.
<context>
{context}
</context>
<question>
{question}
</question>
"""

Utilizar o OpenAI ChatGPT para gerar uma resposta com base nos avisos.

response = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": USER_PROMPT},
    ],
)
print(response.choices[0].message.content)

INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"


The main components of autonomous agents are:

1. **Planning**:
   - Subgoal and decomposition: Breaking down large tasks into smaller, manageable subgoals.
   - Reflection and refinement: Self-criticism and reflection to learn from past actions and improve future steps.

2. **Memory**:
   - Short-term memory: In-context learning using prompt engineering.
   - Long-term memory: Retaining and recalling information over extended periods using an external vector store and fast retrieval.

3. **Tool use**:
   - Calling external APIs for information not contained in the model weights, accessing current information, code execution capabilities, and proprietary information sources.