Llamaindex

LlamaIndex

Overview

LlamaIndex is a data framework for building RAG pipelines, knowledge assistants, and data-augmented LLM applications. It provides document loading from 300+ sources, flexible chunking strategies, multiple index types, hybrid retrieval with reranking, and production evaluation tools for question-answering systems.

Instructions

• When ingesting documents, use SimpleDirectoryReader for local files or LlamaHub connectors for SaaS platforms, and run through an IngestionPipeline with metadata extractors (title, summary) and deduplication.
• When chunking, start with SentenceSplitter at 1024 tokens with 200 token overlap, use MarkdownNodeParser for structured documents, CodeSplitter for code, and adjust based on evaluation results.
• When indexing, use VectorStoreIndex as the default for most RAG, KnowledgeGraphIndex for entity relationships, and DocumentSummaryIndex for per-document summaries.
• When retrieving, implement hybrid retrieval (vector + keyword) for production, add a reranker (CohereRerank) after retrieval for improved relevance, and set similarity_top_k based on context window (3-5 for large models, 2-3 for smaller).
• When building query engines, use RetrieverQueryEngine for standard RAG, CitationQueryEngine for responses with source attribution, and SubQuestionQueryEngine for complex multi-part queries.
• When creating agents, use ReActAgent with tools wrapping query engines (QueryEngineTool), functions, and other agents for multi-step reasoning.
• When evaluating, use CorrectnessEvaluator, FaithfulnessEvaluator, and RelevancyEvaluator on a test set before deploying.

Examples

Example 1: Build a RAG pipeline over company documentation

User request: "Create a question-answering system over our internal docs"

Actions:

1. Load documents with SimpleDirectoryReader and extract metadata (title, summary)
2. Chunk with SentenceSplitter (1024 tokens, 200 overlap) through an IngestionPipeline
3. Create VectorStoreIndex with OpenAI embeddings and configure hybrid retrieval
4. Build CitationQueryEngine for answers with source references

Output: A RAG system that answers questions with citations from company documentation.

Example 2: Create a multi-source research agent

User request: "Build an agent that can search across our docs, database, and web"

Actions:

1. Create separate query engines for each data source (vector index, SQL, web search)
2. Wrap each engine as a QueryEngineTool with descriptive tool descriptions
3. Build a ReActAgent that routes questions to the appropriate tool
4. Add SubQuestionQueryEngine for complex queries requiring multiple sources

Output: An intelligent agent that reasons about which data source to query and synthesizes multi-source answers.

Guidelines

• Use SentenceSplitter with 1024 token chunks and 200 token overlap as the starting point.
• Always add metadata extractors to the ingestion pipeline; title and summary metadata improve retrieval significantly.
• Use hybrid retrieval (vector + keyword) for production; pure vector search misses exact term matches.
• Add a reranker (CohereRerank) after retrieval to improve result relevance for small cost.
• Evaluate with CorrectnessEvaluator on a test set before deploying; subjective quality assessment does not scale.
• Set similarity_top_k based on context window: 3-5 chunks for large models, 2-3 for smaller models.
• Use IngestionPipeline with deduplication for incremental data updates; do not re-embed unchanged documents.