LlamaIndex Starter

Pre-configured template implementing retrieval-augmented generation (RAG) for question-answering over document collections. Uses LlamaIndex's document ingestion pipeline to parse files (PDF, TXT, Markdown), chunk them with configurable strategies, embed chunks via vector stores, and retrieve relevant context before passing to an LLM for answer generation. Abstracts away index construction, retrieval configuration, and prompt engineering boilerplate.

Template implementing stateful conversation over documents using LlamaIndex's chat engine, which maintains conversation history while retrieving relevant document context for each turn. Handles context window management by summarizing or filtering conversation history, retrieves fresh context from the document index per query, and passes both history and context to the LLM to generate contextually-aware responses that reference previous turns.

Template providing utilities to evaluate RAG system quality across multiple dimensions: retrieval quality (precision, recall, NDCG), answer quality (relevance, factuality, citation accuracy), and end-to-end performance. Includes evaluation datasets, metrics computation, and comparison tools to measure impact of configuration changes. Supports both automated metrics (embedding-based similarity) and human evaluation workflows.

Template providing utilities to monitor and optimize LLM API costs and latency in RAG pipelines. Tracks token usage per component (retrieval, synthesis, tool calls), identifies bottlenecks, and suggests optimizations (smaller models, caching, batching). Implements caching strategies (semantic caching, exact-match caching) to reduce redundant LLM calls, and provides cost estimation before execution.

Template using LlamaIndex's structured output capabilities (via Pydantic schema definitions) to extract typed data from documents. Defines a Pydantic model representing desired output structure (e.g., invoice fields, entity lists), passes documents through LlamaIndex's extraction pipeline which uses the LLM to parse content and map it to the schema, and returns validated structured objects. Handles schema validation, type coercion, and optional field handling automatically.

Template implementing an agentic loop where an LLM reasons over multiple documents and tools to answer complex queries. Uses LlamaIndex's agent framework to define tools (document search, calculation, external API calls), implements a ReAct-style loop where the agent plans actions, executes tools, observes results, and refines its approach. Manages context across multiple document indexes and tool invocations, handling tool selection, parameter binding, and result integration into the reasoning loop.

Template exposing LlamaIndex's chunking and indexing configuration options (chunk size, overlap, separator strategy, node post-processors) as configurable parameters. Allows developers to experiment with different chunking strategies (fixed-size, semantic, hierarchical) and index types (vector, keyword, tree-based) without code changes. Includes utilities to evaluate chunking quality and measure retrieval performance across configurations.

Template supporting indexing of multi-modal documents (PDFs with images, scanned documents, mixed text/image content) using LlamaIndex's image extraction and OCR capabilities. Automatically extracts images from documents, generates descriptions or embeddings for images, indexes both text and image content separately, and enables retrieval that matches queries against both text and visual content. Handles image-to-text mapping to preserve document structure.

Template implementing hybrid retrieval that combines semantic vector search with keyword/BM25 search to improve recall and precision. Uses LlamaIndex's retriever composition to run both vector and keyword queries in parallel, ranks results using configurable fusion strategies (RRF, weighted scoring), and returns a merged result set. Enables fallback to keyword search when vector search fails and vice versa, improving robustness across different query types.

Template using LlamaIndex's metadata filtering capabilities to enable retrieval constrained by document metadata (date ranges, categories, source, author, etc.). Defines metadata schemas, attaches metadata to indexed nodes, and uses metadata filters in retrieval queries to narrow search space. Supports complex filter expressions (AND, OR, NOT) and enables faceted search where users can filter by multiple metadata dimensions simultaneously.

Template implementing query preprocessing techniques (expansion, rewriting, decomposition) to improve retrieval quality. Uses LlamaIndex's query transformation modules to generate multiple query variants (synonyms, paraphrases, sub-questions), retrieves results for each variant, and merges results. Handles query decomposition for complex multi-part questions, enabling the system to retrieve context for each sub-question separately before synthesis.

Template implementing response synthesis that generates LLM answers while tracking and attributing source documents. Uses LlamaIndex's response synthesizer to combine retrieved context with LLM generation, maintains source-to-content mappings, and generates citations or footnotes in the final response. Supports multiple synthesis modes (refine, compact, tree-summarize) with different trade-offs between quality and token usage.