llamaindex

zechenzhangAGI

更新于 Today

73 次查看

元wordaidesigndata

关于

LlamaIndex is a data framework for building RAG-powered LLM applications, specializing in document ingestion, indexing, and querying. It provides key features like vector indices, query engines, and agents, and supports over 300 data connectors. Use it for document Q&A, chatbots, and knowledge retrieval when building data-centric applications.

技能文档

LlamaIndex - Data Framework for LLM Applications

The leading framework for connecting LLMs with your data.

When to use LlamaIndex

Use LlamaIndex when:

Building RAG (retrieval-augmented generation) applications
Need document question-answering over private data
Ingesting data from multiple sources (300+ connectors)
Creating knowledge bases for LLMs
Building chatbots with enterprise data
Need structured data extraction from documents

Metrics:

45,100+ GitHub stars
23,000+ repositories use LlamaIndex
300+ data connectors (LlamaHub)
1,715+ contributors
v0.14.7 (stable)

Use alternatives instead:

LangChain: More general-purpose, better for agents
Haystack: Production search pipelines
txtai: Lightweight semantic search
Chroma: Just need vector storage

Quick start

Installation

# Starter package (recommended)
pip install llama-index

# Or minimal core + specific integrations
pip install llama-index-core
pip install llama-index-llms-openai
pip install llama-index-embeddings-openai

5-line RAG example

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader

# Load documents
documents = SimpleDirectoryReader("data").load_data()

# Create index
index = VectorStoreIndex.from_documents(documents)

# Query
query_engine = index.as_query_engine()
response = query_engine.query("What did the author do growing up?")
print(response)

Core concepts

1. Data connectors - Load documents

from llama_index.core import SimpleDirectoryReader, Document
from llama_index.readers.web import SimpleWebPageReader
from llama_index.readers.github import GithubRepositoryReader

# Directory of files
documents = SimpleDirectoryReader("./data").load_data()

# Web pages
reader = SimpleWebPageReader()
documents = reader.load_data(["https://example.com"])

# GitHub repository
reader = GithubRepositoryReader(owner="user", repo="repo")
documents = reader.load_data(branch="main")

# Manual document creation
doc = Document(
    text="This is the document content",
    metadata={"source": "manual", "date": "2025-01-01"}
)

2. Indices - Structure data

from llama_index.core import VectorStoreIndex, ListIndex, TreeIndex

# Vector index (most common - semantic search)
vector_index = VectorStoreIndex.from_documents(documents)

# List index (sequential scan)
list_index = ListIndex.from_documents(documents)

# Tree index (hierarchical summary)
tree_index = TreeIndex.from_documents(documents)

# Save index
index.storage_context.persist(persist_dir="./storage")

# Load index
from llama_index.core import load_index_from_storage, StorageContext
storage_context = StorageContext.from_defaults(persist_dir="./storage")
index = load_index_from_storage(storage_context)

3. Query engines - Ask questions

# Basic query
query_engine = index.as_query_engine()
response = query_engine.query("What is the main topic?")
print(response)

# Streaming response
query_engine = index.as_query_engine(streaming=True)
response = query_engine.query("Explain quantum computing")
for text in response.response_gen:
    print(text, end="", flush=True)

# Custom configuration
query_engine = index.as_query_engine(
    similarity_top_k=3,          # Return top 3 chunks
    response_mode="compact",     # Or "tree_summarize", "simple_summarize"
    verbose=True
)

4. Retrievers - Find relevant chunks

# Vector retriever
retriever = index.as_retriever(similarity_top_k=5)
nodes = retriever.retrieve("machine learning")

# With filtering
retriever = index.as_retriever(
    similarity_top_k=3,
    filters={"metadata.category": "tutorial"}
)

# Custom retriever
from llama_index.core.retrievers import BaseRetriever

class CustomRetriever(BaseRetriever):
    def _retrieve(self, query_bundle):
        # Your custom retrieval logic
        return nodes

Agents with tools

Basic agent

from llama_index.core.agent import FunctionAgent
from llama_index.llms.openai import OpenAI

# Define tools
def multiply(a: int, b: int) -> int:
    """Multiply two numbers."""
    return a * b

def add(a: int, b: int) -> int:
    """Add two numbers."""
    return a + b

# Create agent
llm = OpenAI(model="gpt-4o")
agent = FunctionAgent.from_tools(
    tools=[multiply, add],
    llm=llm,
    verbose=True
)

# Use agent
response = agent.chat("What is 25 * 17 + 142?")
print(response)

RAG agent (document search + tools)

from llama_index.core.tools import QueryEngineTool

# Create index as before
index = VectorStoreIndex.from_documents(documents)

# Wrap query engine as tool
query_tool = QueryEngineTool.from_defaults(
    query_engine=index.as_query_engine(),
    name="python_docs",
    description="Useful for answering questions about Python programming"
)

# Agent with document search + calculator
agent = FunctionAgent.from_tools(
    tools=[query_tool, multiply, add],
    llm=llm
)

# Agent decides when to search docs vs calculate
response = agent.chat("According to the docs, what is Python used for?")

Advanced RAG patterns

Chat engine (conversational)

from llama_index.core.chat_engine import CondensePlusContextChatEngine

# Chat with memory
chat_engine = index.as_chat_engine(
    chat_mode="condense_plus_context",  # Or "context", "react"
    verbose=True
)

# Multi-turn conversation
response1 = chat_engine.chat("What is Python?")
response2 = chat_engine.chat("Can you give examples?")  # Remembers context
response3 = chat_engine.chat("What about web frameworks?")

Metadata filtering

from llama_index.core.vector_stores import MetadataFilters, ExactMatchFilter

# Filter by metadata
filters = MetadataFilters(
    filters=[
        ExactMatchFilter(key="category", value="tutorial"),
        ExactMatchFilter(key="difficulty", value="beginner")
    ]
)

retriever = index.as_retriever(
    similarity_top_k=3,
    filters=filters
)

query_engine = index.as_query_engine(filters=filters)

Structured output

from pydantic import BaseModel
from llama_index.core.output_parsers import PydanticOutputParser

class Summary(BaseModel):
    title: str
    main_points: list[str]
    conclusion: str

# Get structured response
output_parser = PydanticOutputParser(output_cls=Summary)
query_engine = index.as_query_engine(output_parser=output_parser)

response = query_engine.query("Summarize the document")
summary = response  # Pydantic model
print(summary.title, summary.main_points)

Data ingestion patterns

Multiple file types

# Load all supported formats
documents = SimpleDirectoryReader(
    "./data",
    recursive=True,
    required_exts=[".pdf", ".docx", ".txt", ".md"]
).load_data()

Web scraping

from llama_index.readers.web import BeautifulSoupWebReader

reader = BeautifulSoupWebReader()
documents = reader.load_data(urls=[
    "https://docs.python.org/3/tutorial/",
    "https://docs.python.org/3/library/"
])

Database

from llama_index.readers.database import DatabaseReader

reader = DatabaseReader(
    sql_database_uri="postgresql://user:pass@localhost/db"
)
documents = reader.load_data(query="SELECT * FROM articles")

API endpoints

from llama_index.readers.json import JSONReader

reader = JSONReader()
documents = reader.load_data("https://api.example.com/data.json")

Vector store integrations

Chroma (local)

from llama_index.vector_stores.chroma import ChromaVectorStore
import chromadb

# Initialize Chroma
db = chromadb.PersistentClient(path="./chroma_db")
collection = db.get_or_create_collection("my_collection")

# Create vector store
vector_store = ChromaVectorStore(chroma_collection=collection)

# Use in index
from llama_index.core import StorageContext
storage_context = StorageContext.from_defaults(vector_store=vector_store)
index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)

Pinecone (cloud)

from llama_index.vector_stores.pinecone import PineconeVectorStore
import pinecone

# Initialize Pinecone
pinecone.init(api_key="your-key", environment="us-west1-gcp")
pinecone_index = pinecone.Index("my-index")

# Create vector store
vector_store = PineconeVectorStore(pinecone_index=pinecone_index)
storage_context = StorageContext.from_defaults(vector_store=vector_store)

index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)

FAISS (fast)

from llama_index.vector_stores.faiss import FaissVectorStore
import faiss

# Create FAISS index
d = 1536  # Dimension of embeddings
faiss_index = faiss.IndexFlatL2(d)

vector_store = FaissVectorStore(faiss_index=faiss_index)
storage_context = StorageContext.from_defaults(vector_store=vector_store)

index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)

Customization

Custom LLM

from llama_index.llms.anthropic import Anthropic
from llama_index.core import Settings

# Set global LLM
Settings.llm = Anthropic(model="claude-sonnet-4-5-20250929")

# Now all queries use Anthropic
query_engine = index.as_query_engine()

Custom embeddings

from llama_index.embeddings.huggingface import HuggingFaceEmbedding

# Use HuggingFace embeddings
Settings.embed_model = HuggingFaceEmbedding(
    model_name="sentence-transformers/all-mpnet-base-v2"
)

index = VectorStoreIndex.from_documents(documents)

Custom prompt templates

from llama_index.core import PromptTemplate

qa_prompt = PromptTemplate(
    "Context: {context_str}\n"
    "Question: {query_str}\n"
    "Answer the question based only on the context. "
    "If the answer is not in the context, say 'I don't know'.\n"
    "Answer: "
)

query_engine = index.as_query_engine(text_qa_template=qa_prompt)

Multi-modal RAG

Image + text

from llama_index.core import SimpleDirectoryReader
from llama_index.multi_modal_llms.openai import OpenAIMultiModal

# Load images and documents
documents = SimpleDirectoryReader(
    "./data",
    required_exts=[".jpg", ".png", ".pdf"]
).load_data()

# Multi-modal index
index = VectorStoreIndex.from_documents(documents)

# Query with multi-modal LLM
multi_modal_llm = OpenAIMultiModal(model="gpt-4o")
query_engine = index.as_query_engine(llm=multi_modal_llm)

response = query_engine.query("What is in the diagram on page 3?")

Evaluation

Response quality

from llama_index.core.evaluation import RelevancyEvaluator, FaithfulnessEvaluator

# Evaluate relevance
relevancy = RelevancyEvaluator()
result = relevancy.evaluate_response(
    query="What is Python?",
    response=response
)
print(f"Relevancy: {result.passing}")

# Evaluate faithfulness (no hallucination)
faithfulness = FaithfulnessEvaluator()
result = faithfulness.evaluate_response(
    query="What is Python?",
    response=response
)
print(f"Faithfulness: {result.passing}")

Best practices

Use vector indices for most cases - Best performance
Save indices to disk - Avoid re-indexing
Chunk documents properly - 512-1024 tokens optimal
Add metadata - Enables filtering and tracking
Use streaming - Better UX for long responses
Enable verbose during dev - See retrieval process
Evaluate responses - Check relevance and faithfulness
Use chat engine for conversations - Built-in memory
Persist storage - Don't lose your index
Monitor costs - Track embedding and LLM usage

Common patterns

Document Q&A system

# Complete RAG pipeline
documents = SimpleDirectoryReader("docs").load_data()
index = VectorStoreIndex.from_documents(documents)
index.storage_context.persist(persist_dir="./storage")

# Query
query_engine = index.as_query_engine(
    similarity_top_k=3,
    response_mode="compact",
    verbose=True
)
response = query_engine.query("What is the main topic?")
print(response)
print(f"Sources: {[node.metadata['file_name'] for node in response.source_nodes]}")

Chatbot with memory

# Conversational interface
chat_engine = index.as_chat_engine(
    chat_mode="condense_plus_context",
    verbose=True
)

# Multi-turn chat
while True:
    user_input = input("You: ")
    if user_input.lower() == "quit":
        break
    response = chat_engine.chat(user_input)
    print(f"Bot: {response}")

Performance benchmarks

Operation	Latency	Notes
Index 100 docs	~10-30s	One-time, can persist
Query (vector)	~0.5-2s	Retrieval + LLM
Streaming query	~0.5s first token	Better UX
Agent with tools	~3-8s	Multiple tool calls

LlamaIndex vs LangChain

Feature	LlamaIndex	LangChain
Best for	RAG, document Q&A	Agents, general LLM apps
Data connectors	300+ (LlamaHub)	100+
RAG focus	Core feature	One of many
Learning curve	Easier for RAG	Steeper
Customization	High	Very high
Documentation	Excellent	Good