llamaindex_skill

Overview

This skill provides a data framework for building retrieval-augmented generation (RAG) applications and document-centric LLM systems. It focuses on robust document ingestion (300+ connectors), vector indexing, query engines, agents, and multi-modal support to power document Q&A, chatbots, and knowledge retrieval. Built for data-centric LLM workflows, it streamlines indexing, querying, and persistence for production and research use.

How this skill works

The skill ingests documents from many sources (files, web, databases, APIs, GitHub) and converts them into documents and metadata. It builds indices (vector, list, tree) and connects to vector stores (Chroma, FAISS, Pinecone) for fast retrieval. Query engines and retrievers return relevant chunks, which LLMs use to generate answers, and optional agents orchestrate tools and document search for decision-making. Persistence, filtering, streaming, and structured-output parsers are supported to make pipelines reliable and auditable.

When to use it

• Building RAG systems or document Q&A over private data
• Creating chatbots backed by enterprise documents or multi-modal assets
• Ingesting heterogeneous data sources or building a centralized knowledge base
• Needing fast semantic search with vector indices and metadata filters
• Prototyping or productionizing document-centric LLM pipelines

Best practices

• Prefer vector indices for semantic search; chunk documents ~512–1024 tokens
• Persist indices to disk or a managed vector store to avoid re-indexing
• Add metadata on ingestion to enable precise filtering and provenance
• Enable streaming during development for better UX on long responses
• Evaluate responses for relevance and faithfulness to reduce hallucinations
• Use chat engines for multi-turn conversations and built-in memory

Example use cases

• Document Q&A: ingest manuals, index them, and answer user questions with source citations
• Enterprise chatbot: combine company docs, databases, and APIs for an internal assistant
• RAG agent: let an agent decide when to search docs vs call helper tools (calculators, APIs)
• Multi-modal retrieval: index images and PDFs and query diagrams alongside text
• Evaluation pipeline: run relevancy and faithfulness evaluators to monitor model quality