Agent Skills: LlamaIndex Development

LlamaIndex Development

Complete guide for building RAG applications with LlamaIndex framework.

Overview

LlamaIndex is a data framework for LLM applications, providing tools for data ingestion, indexing, and retrieval.

Key Characteristics:

• RAG (Retrieval Augmented Generation) support
• Multiple data connectors (300+)
• Various index types
• Query and chat engines
• Vector store integrations
• Agent framework

Installation

Setup

# Basic installation (v0.14+)
pip install llama-index

# With OpenAI
pip install llama-index-llms-openai
pip install llama-index-embeddings-openai

# With workflow engine support
pip install llama-index-core

# With vector stores
pip install llama-index-vector-stores-chroma
pip install llama-index-vector-stores-pinecone
pip install llama-index-vector-stores-qdrant

# With evaluation
pip install llama-index-llms-openai ragas

Version Notes

• Current stable: v0.14.x (rapid monthly releases)
• Import paths unified under llama_index.core
• Workflow engine added for complex multi-step operations

Basic Configuration

import os
from llama_index.core import Settings
from llama_index.llms.openai import OpenAI
from llama_index.embeddings.openai import OpenAIEmbedding

# Set API key
os.environ["OPENAI_API_KEY"] = "YOUR_API_KEY"

# Configure global settings
Settings.llm = OpenAI(model="gpt-4o", temperature=0.0)
Settings.embed_model = OpenAIEmbedding(model="text-embedding-3-small")
Settings.chunk_size = 512
Settings.chunk_overlap = 50

Quick Start

Basic RAG Pipeline (v0.14+)

# Unified imports in v0.14+
from llama_index.core import VectorStoreIndex, SimpleDirectoryReader, StorageContext, load_index_from_storage

# Load documents
documents = SimpleDirectoryReader("./docs").load_data()

# Create index
index = VectorStoreIndex.from_documents(documents)

# Create query engine
query_engine = index.as_query_engine()

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

With Storage

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader, StorageContext
from llama_index.embeddings.openai import OpenAIEmbedding
import chromadb
from llama_index.vector_stores.chroma import ChromaVectorStore

# Setup ChromaDB
db = chromadb.PersistentClient(path="./chroma_db")
chroma_collection = db.get_or_create_collection("my-index")
vector_store = ChromaVectorStore(chroma_collection=chroma_collection)
storage_context = StorageContext.from_defaults(vector_store=vector_store)

# Load and index
documents = SimpleDirectoryReader("./docs").load_data()
index = VectorStoreIndex.from_documents(
    documents,
    storage_context=storage_context,
    embed_model=OpenAIEmbedding(),
)

# Persist
storage_context.persist()

# Load from disk
index = VectorStoreIndex.from_vector_store(
    vector_store,
    embed_model=OpenAIEmbedding(),
)

Data Loading

Document Loaders

from llama_index.core import SimpleDirectoryReader, Document

# Load from directory
documents = SimpleDirectoryReader(
    input_dir="./docs",
    required_exts=[".pdf", ".txt", ".md"],
    exclude=["*.tmp"],
    recursive=True,
).load_data()

# Load specific files
documents = SimpleDirectoryReader(
    input_files=["./file1.pdf", "./file2.txt"]
).load_data()

# Create documents manually
documents = [
    Document(text="Content here", metadata={"source": "manual"}),
]

# With metadata extraction
def custom_metadata_func(file_path: str) -> dict:
    return {
        "file_path": file_path,
        "file_name": os.path.basename(file_path),
    }

documents = SimpleDirectoryReader(
    input_dir="./docs",
    file_metadata=custom_metadata_func,
).load_data()

Custom Data Connectors

from llama_index.core import Document
from typing import List

class CustomDataReader:
    """Custom data loader."""

    def load_data(self, source: str) -> List[Document]:
        documents = []

        # Load from custom source
        # Example: API, database, etc.
        data = self._fetch_from_source(source)

        for item in data:
            doc = Document(
                text=item["content"],
                metadata={
                    "source": source,
                    "id": item["id"],
                    "timestamp": item["timestamp"],
                },
            )
            documents.append(doc)

        return documents

    def _fetch_from_source(self, source: str):
        # Implement data fetching
        pass

# Usage
reader = CustomDataReader()
documents = reader.load_data("api://endpoint")

Chunking Decision Guide

Chunk boundaries directly determine citation quality and retrieval accuracy. Choose based on your data structure and query patterns.

When Sentence Splitting Beats Semantic Chunking

Use sentence splitting when:

• Your documents have clear structural divisions (headings, sections, pages)
• You need predictable chunk sizes for cost/latency estimation
• Your queries are fact-based rather than concept-based
• You're working with technical documentation or legal texts

Use semantic chunking when:

• Your documents lack clear structure (emails, chat logs, unstructured notes)
• Query intent varies significantly within paragraphs
• You need to preserve thematic boundaries over structural ones
• You have budget for embedding-based splitting overhead

Fixed-Size-with-Overlap Pitfalls

# ❌ PITFALL: Ignoring semantic boundaries
splitter = SentenceSplitter(
    chunk_size=512,
    chunk_overlap=50,
    # No paragraph or section awareness
)

# ✅ BETTER: Respect document structure
splitter = SentenceSplitter(
    chunk_size=512,
    chunk_overlap=50,
    paragraph_separator="\n\n",  # Split on paragraphs first
    secondary_chunking_regex=r"\n#{1,6} ",  # Don't break across headings
)

Common overlap mistakes:

• Overlap too small (<10%): Context boundaries get cut mid-sentence
• Overlap too large (>30%): Redundant embeddings, higher cost, diluted relevance
• Fixed overlap ignores structure: Section headers appear in wrong chunks

Metadata to Attach at Ingest Time

Metadata enables filtering and improves retrieval precision. Attach these at ingest:

| Metadata Field | Why It Matters | |----------------|----------------| | source (file path) | Cite exact document when answering | | section_heading | Preserve document hierarchy in context | | page_number | Critical for PDFs and scanned documents | | document_date | Filter by recency for time-sensitive queries | | tenant_id / PROJECT_KEY | Multi-tenant isolation via metadata filters |

from llama_index.core import Document

def load_with_metadata(file_path: str) -> List[Document]:
    """Load documents with essential metadata for retrieval."""
    # Parse file to extract structure
    content, headings, page_numbers = parse_document(file_path)

    return [
        Document(
            text=chunk.text,
            metadata={
                "source": file_path,
                "section": chunk.section,
                "page": page_numbers[chunk.start_line],
                "tenant_id": "my-tenant",  # Replace with actual tenant
            },
        )
        for chunk in split_by_structure(content, headings)
    ]

Evaluating Chunking Changes

Never guess whether a chunking change helped. Measure:

1. Create a fixed eval set: 20-50 query-answer pairs hand-labeled from your actual use cases
2. Run retrieval before change: Record which chunks are retrieved for each query
3. Apply chunking change: Re-index the same documents
4. Run retrieval after change: Compare hit rates on the same queries
5. Measure:
   • Hit rate: % of queries where relevant chunk appears in top-k
   • MRR (Mean Reciprocal Rank): How high relevant chunks rank
   • Citation quality: Do answers reference the correct source sections?

# Simple hit rate evaluation
def evaluate_chunking(queries_with_relevant_chunks, retriever, top_k=5):
    hits = 0
    for query, relevant_ids in queries_with_relevant_chunks:
        nodes = retriever.retrieve(query)
        retrieved_ids = {n.node.node_id for n in nodes}
        if relevant_ids & retrieved_ids:  # Any overlap = hit
            hits += 1
    return hits / len(queries_with_relevant_chunks)

Vector Stores

Vector store setup follows the same pattern across backends:

from llama_index.core import VectorStoreIndex, StorageContext
from llama_index.vector_stores.chroma import ChromaVectorStore
import chromadb

# 1. Initialize backend client
db = chromadb.PersistentClient(path="./chroma_db")
chroma_collection = db.get_or_create_collection("my-index")

# 2. Wrap as LlamaIndex vector store
vector_store = ChromaVectorStore(chroma_collection=chroma_collection)
storage_context = StorageContext.from_defaults(vector_store=vector_store)

# 3. Create index with storage context
index = VectorStoreIndex.from_documents(documents, storage_context=storage_context)

Backend-specific notes:

• ChromaDB: Use PersistentClient for local storage, EphemeralClient for testing
• Pinecone: Create index with matching embedding dimensions (1536 for text-embedding-3-small)
• Qdrant: Specify collection_name and ensure vector size matches embed_model

For detailed setup per backend, see references/retrieval.md.

Retrieval Debugging

Symptom-first guide to fixing retrieval problems.

Answers Cite the Wrong Section

Symptom: Response references information from the wrong document or section.

Check:

1. Are metadata filters applied at query time?
2. Do chunks have accurate source and section metadata?
3. Is the LLM instructed to cite sources?

Fix:

# Apply metadata filters to narrow retrieval
query_engine = index.as_query_engine(
    filters=MetadataFilter(
        key="tenant_id", value="my-tenant"  # Replace with actual filter
    )
)

# Or filter at retriever level
retriever = index.as_retriever(
    filters=MetadataFilter(key="source", value="specific-file.pdf")
)

Retrieval Returns Nothing Relevant

Symptom: Query returns chunks that don't match the question intent.

Check:

1. Do embedding dimensions match at index and query time?
2. Is the embed_model the same for indexing and querying?
3. Is similarity_top_k too low?

Fix:

# Ensure consistent embedding model
from llama_index.embeddings.openai import OpenAIEmbedding

embed_model = OpenAIEmbedding(model="text-embedding-3-small")

# At index time
index = VectorStoreIndex.from_documents(documents, embed_model=embed_model)

# At query time (must match)
query_engine = index.as_query_engine(
    embed_model=embed_model,  # Same model as index time
    similarity_top_k=10,  # Retrieve more before reranking
)

Answers Are Generic

Symptom: Responses lack specificity, sound like boilerplate.

Check:

1. Is similarity_top_k too low (<5)?
2. Is reranking enabled?
3. Are retrieved chunks too large (diluted relevance)?

Fix:

from llama_index.core.postprocessor import SentenceTransformerRerank

# Retrieve more, then rerank
query_engine = index.as_query_engine(
    similarity_top_k=20,
    node_postprocessors=[
        SentenceTransformerRerank(
            model="cross-encoder/ms-marco-MiniLM-L-6-v2",
            top_n=5,
        ),
    ],
)

Hybrid Search Behaves Inconsistently

Symptom: Keyword + vector search gives unpredictable results.

Check:

1. Is the BM25 weight tuned for your data?
2. Is the keyword index fresh (rebuilt after document updates)?
3. Are you using AND vs OR mode appropriately?

Fix:

from llama_index.core.retrievers import QueryFusionRetriever

# Tune query fusion parameters
fusion_retriever = QueryFusionRetriever(
    retrievers=[vector_retriever, keyword_retriever],
    similarity_top_k=10,
    num_queries=3,
    mode="reciprocal_rerank",  # Or "weighted_sum"
)

# For strict matching, use AND mode
hybrid_retriever = HybridRetriever(
    vector_index, keyword_index, mode="AND"  # Only return nodes in both
)

Deep Dives

Advanced topics are split into reference files loaded on demand:

• references/retrieval.md — Advanced retrievers (hybrid, query fusion, auto-merging), rerankers, query engines (router, sub-question, multi-step)
• references/agents-workflows.md — Agents (ReAct, function calling, custom tools), streaming, workflow engine, chat engines
• references/evaluation-observability.md — Evaluation (retrieval hit rate, recall@k, faithfulness, latency/cost tracking), observability (callbacks, LangSmith)

Common Issues

Chunk Size Problems

# ❌ BAD: Too small chunks lose context
splitter = SentenceSplitter(chunk_size=64)

# ❌ BAD: Too large chunks dilute relevance
splitter = SentenceSplitter(chunk_size=8192)

# ✅ GOOD: Balanced chunk size
splitter = SentenceSplitter(
    chunk_size=512,  # For embeddings
    chunk_overlap=50,  # ~10% overlap
)

Retrieval Quality

# ❌ BAD: No reranking, few results
query_engine = index.as_query_engine(similarity_top_k=3)

# ✅ GOOD: Retrieve more, rerank
query_engine = index.as_query_engine(
    similarity_top_k=20,
    node_postprocessors=[
        SentenceTransformerRerank(top_n=5),
    ],
)

Memory Issues

# ❌ BAD: Load all documents at once
documents = SimpleDirectoryReader("./huge_folder").load_data()

# ✅ GOOD: Process in batches
from llama_index.core import StorageContext

for batch in document_batches:
    nodes = splitter.get_nodes_from_documents(batch)
    index.insert_nodes(nodes)

Embedding Dimension Mismatch

# ❌ BAD: Index created with different embedding
# Pinecone index: 1536 dimensions
# Using: 768 dimension embeddings

# ✅ GOOD: Match dimensions
embed_model = OpenAIEmbedding(model="text-embedding-3-small")  # 1536 dims
# Create Pinecone index with same dimensions

Best Practices

1. Use appropriate chunk sizes (512-1024 for most use cases)
2. Always add overlap (5-10% of chunk size)
3. Use rerankers for better retrieval quality
4. Enable streaming for better UX
5. Use async for parallel queries
6. Implement caching for repeated queries
7. Monitor token usage with callbacks
8. Test with evaluation before production
9. Use hybrid search for better recall
10. Keep context window in mind for chat

Resources

• Documentation: https://docs.llamaindex.ai/
• GitHub: https://github.com/run-llama/llama_index
• Examples: https://github.com/run-llama/llama_index/tree/main/docs/examples
• Discord: https://discord.gg/dGcwcsnxhU
• Blog: https://blog.llamaindex.ai/

Quick Reference

Common Imports

from llama_index.core import (
    VectorStoreIndex,
    SimpleDirectoryReader,
    Document,
    Settings,
    StorageContext,
)
from llama_index.core.node_parser import SentenceSplitter
from llama_index.core.query_engine import RetrieverQueryEngine
from llama_index.core.retrievers import VectorIndexRetriever
from llama_index.llms.openai import OpenAI
from llama_index.embeddings.openai import OpenAIEmbedding

Common Patterns

# Basic RAG
index = VectorStoreIndex.from_documents(documents)
query_engine = index.as_query_engine()
response = query_engine.query("query")

# With reranking
query_engine = index.as_query_engine(
    similarity_top_k=20,
    node_postprocessors=[reranker],
)

# Streaming
query_engine = index.as_query_engine(streaming=True)
for token in query_engine.query("query").response_gen:
    print(token)

# Chat
chat_engine = index.as_chat_engine()
response = chat_engine.chat("message")

# Agent
agent = ReActAgent.from_tools(tools, llm=llm)
response = agent.chat("message")