How can I retrieve documents using LlamaIndex?

To retrieve documents using LlamaIndex, you’ll need to structure your data, build an index, and query it using natural language or specific parameters. LlamaIndex simplifies connecting large language models (LLMs) to external data by organizing documents into searchable indexes. The process typically involves three steps: loading data, creating an index, and querying it. Let’s break this down with practical examples.

First, load your documents into LlamaIndex using data connectors. For instance, you can use the SimpleDirectoryReader to ingest files from a local directory. This supports formats like PDFs, text files, or Markdown. Once loaded, documents are parsed into nodes—smaller chunks of text with metadata. For example, a 10-page PDF might be split into 20 nodes, each representing a section. Splitting documents into nodes improves retrieval accuracy by narrowing results to relevant sections. You can customize chunk size, overlap, or metadata (e.g., document titles) during this step.

Next, create an index tailored to your use case. The VectorStoreIndex is a common choice, which converts text into numerical embeddings (e.g., using OpenAI or Hugging Face models) and stores them for semantic search. To build it, pass your nodes to VectorStoreIndex.from_documents(nodes). For structured data, a ListIndex might be better, which retrieves documents based on keyword matches. Once indexed, use a QueryEngine to search. For example, query_engine = index.as_query_engine() lets you call query_engine.query("Find reports on Q3 sales"). LlamaIndex handles the retrieval by comparing the query’s embedding to node embeddings, returning the most relevant results.

Finally, customize retrieval with advanced options. You can adjust parameters like similarity_top_k to control how many nodes are returned, or use hybrid search that combines semantic and keyword-based results. For instance, VectorIndexRetriever(similarity_top_k=5) fetches the top five matches. To filter by metadata, add a node_filter to exclude nodes that don’t meet criteria like date ranges. For complex queries, use a RouterQueryEngine to direct questions to different indexes (e.g., routing financial data to a dedicated index). These tools let you balance speed, accuracy, and specificity based on your application’s needs.