基于 LlamaIndex、Claude-3.5 Sonnet 和 MongoDB，构建具有超级检索能力的智能体

节前，我们组织了一场算法岗技术&面试讨论会，邀请了一些互联网大厂朋友、今年参加社招和校招面试的同学。

针对大模型技术趋势、算法项目落地经验分享、新手如何入门算法岗、该如何准备面试攻略、面试常考点等热门话题进行了深入的讨论。

总结链接如下：

《大模型面试宝典》(2024版) 正式发布

喜欢本文记得收藏、关注、点赞。更多实战和面试交流，文末加入我们

想象一下，AI 助手可以无缝地与你互动，动态地根据你的需求检索信息并完成任务。随着智能体检索增强生成（RAG）的兴起，这一愿景正逐渐成为现实。

在本文中，我们将深入探讨这个令人兴奋的领域，探索如何利用强大的工具组合：LlamaIndex、Claude-3.5 Sonnet 和 MongoDB 来创建具有检索超级能力的 AI 智能体。

工具集成

让我们看看如何将这些强大的工具结合在一起：

LlamaIndex：这个先进的搜索引擎擅长基于意义而不是关键词查找相似信息。它充当 AI 智能体的“眼睛”，在海量信息中定位最相关的数据。
Claude-3.5 Sonnet：它允许智能体处理 LlamaIndex 检索到的信息，生成响应。
MongoDB：一个强大的 NoSQL 数据库，MongoDB 在存储和管理支持 AI 智能体的知识库中起着关键作用。其灵活性允许存储各种数据类型，使其成为复杂信息检索任务的理想选择。

集成优势

这种协同作用带来了许多好处：

增强的信息检索：LlamaIndex 的向量搜索功能确保 AI 智能体检索到最相关的信息，即使是细微的查询也不例外。
动态任务完成：Claude-3.5 Sonnet 使 AI 智能体能够分析检索到的数据并采取适当的行动，使其真正成为能够独立行动的智能体。
可扩展性和灵活性：MongoDB 处理大型数据集的能力允许系统随着信息需求的增加而增长。

代码实现

让我们深入研究使用 LlamaIndex、Claude-3.5 Sonnet 和 MongoDB 的智能体 RAG。

第一步：安装库

!pip install --quiet llama-index  # main llamaindex library
!pip install --quiet llama-index-vector-stores-MongoDB # mongodb vector database
!pip install --quiet llama-index-llms-anthropic # anthropic LLM provider
!pip install --quiet llama-index-embeddings-openai # openai embedding provider
!pip install --quiet pymongo pandas datasets # others

第二步：设置环境变量

import os
os.environ["ANTHROPIC_API_KEY"] = ""
os.environ["HF_TOKEN"] = ""
os.environ["OPENAI_API_KEY"] = ""
# WARNING: Never commit API keys or sensitive information to public repositories

LLM 和嵌入模型配置

from llama_index.embeddings.openai import OpenAIEmbedding
from llama_index.llms.anthropic import Anthropic
from llama_index.core import Settings

llm = Anthropic(model="claude-3-5-sonnet-20240620")

embed_model = OpenAIEmbedding(
    model="text-embedding-3-small", 
    dimensions=256,
    embed_batch_size=10, 
    openai_api_key=os.environ["OPENAI_API_KEY"]
)

Settings.embed_model = embed_model
Settings.llm = llm

第三步：数据加载和处理

from datasets import load_dataset
import pandas as pd

# https://huggingface.co/datasets/MongoDB/airbnb_embeddings
dataset = load_dataset("MongoDB/airbnb_embeddings", split="train", streaming=True)
dataset = dataset.take(4000)

# Convert the dataset to a pandas dataframe
dataset_df = pd.DataFrame(dataset)

# Dataset comes with embeddings created with OpenAI, but we will recreate new ones
dataset_df = dataset_df.drop(columns=['text_embeddings'])
dataset_df.head(5)

第四步：生成嵌入

import json
from llama_index.core import Document
from llama_index.core.schema import MetadataMode

documents_json = dataset_df.to_json(orient='records')
documents_list = json.loads(documents_json)

llama_documents = []

for document in documents_list:
    # Convert complex objects to JSON strings
    for field in ["amenities", "images", "host", "address", "availability", "review_scores", "reviews", "image_embeddings"]:
        document[field] = json.dumps(document[field])

    # Create a Document object
    llama_document = Document(
        text=document["description"],
        metadata=document,
        excluded_llm_metadata_keys=["_id", "transit", "minimum_nights", "maximum_nights", "cancellation_policy", "last_scraped", "calendar_last_scraped", "first_review", "last_review", "security_deposit", "cleaning_fee", "guests_included", "host", "availability", "reviews", "image_embeddings"],
        excluded_embed_metadata_keys=["_id", "transit", "minimum_nights", "maximum_nights", "cancellation_policy", "last_scraped", "calendar_last_scraped", "first_review", "last_review", "security_deposit", "cleaning_fee", "guests_included", "host", "availability", "reviews", "image_embeddings"],
        metadata_template="{key}=>{value}",
        text_template="Metadata: {metadata_str}\n-----\nContent: {content}",
    )
    llama_documents.append(llama_document)

# Observing input examples
print("\nThe LLM sees this: \n", llama_documents[0].get_content(metadata_mode=MetadataMode.LLM))
print("\nThe Embedding model sees this: \n", llama_documents[0].get_content(metadata_mode=MetadataMode.EMBED))
from llama_index.core.node_parser import SentenceSplitter, SemanticSplitterNodeParser
from llama_index.core.schema import MetadataMode
from tqdm import tqdm

# semantic_splitter = SemanticSplitterNodeParser(
#     buffer_size=10, breakpoint_percentile_threshold=95, embed_model=embed_model
# )

base_splitter = SentenceSplitter(chunk_size=5000, chunk_overlap=200)

nodes = base_splitter.get_nodes_from_documents(llama_documents)

# Progress bar
pbar = tqdm(total=len(nodes), desc="Embedding Progress", unit="node")

for node in nodes:
    node_embedding = embed_model.get_text_embedding(
        node.get_content(metadata_mode=MetadataMode.EMBED)
    )
    node.embedding = node_embedding
    
    # Update the progress bar
    pbar.update(1)

# Close the progress bar
pbar.close()

print("Embedding process completed!")

第五步：MongoDB 设置

import pymongo

os.environ["MONGO_URI"] = ""

def get_mongo_client(mongo_uri):
    """Establish and validate connection to the MongoDB."""
    
    client = pymongo.MongoClient(mongo_uri, appname="devrel.showcase.python")

    # Validate the connection
    ping_result = client.admin.command('ping')
    if ping_result.get('ok') == 1.0:
        # Connection successful
        print("Connection to MongoDB successful")
        return client
    else:
        print("Connection to MongoDB failed")
    return None


mongo_client = get_mongo_client(mongo_uri)

DB_NAME = "airbnb"
COLLECTION_NAME = "listings_reviews"

db = mongo_client.get_database(DB_NAME)
collection = db.get_collection(COLLECTION_NAME)

第六步：向量数据库集成

from llama_index.vector_stores.mongodb import MongoDBAtlasVectorSearch

vector_store = MongoDBAtlasVectorSearch(
    mongo_client, 
    db_name=DB_NAME, 
    collection_name=COLLECTION_NAME, 
    index_name="vector_index"
)

vector_store.add(nodes)

第七步：创建检索工具和智能体

from llama_index.core import VectorStoreIndex
from llama_index.core.tools import QueryEngineTool, ToolMetadata
from llama_index.core.agent import FunctionCallingAgentWorker

index = VectorStoreIndex.from_vector_store(vector_store)
query_engine = index.as_query_engine(similarity_top_k=5, llm=llm)

query_engine_tool = QueryEngineTool(
    query_engine=query_engine,
    metadata=ToolMetadata(
        name="knowledge_base",
        description=(
            "Provides information about Airbnb listings and reviews."
            "Use a detailed plain text question as input to the tool."
        ),
    ),
)

agent_worker = FunctionCallingAgentWorker.from_tools(
    [query_engine_tool], llm=llm, verbose=True
)
agent = agent_worker.as_agent()

response = agent.chat("Tell me the best listing for a place in New York")
print(str(response))