agents-llamaindex_skill

Overview

This skill exposes the LlamaIndex framework for building data-centric LLM applications and retrieval-augmented generation (RAG) workflows. It provides tools for document ingestion, indexing, querying, and composing retrieval-driven chatbots or agents that combine document search with tool calls. Use it to turn files, databases, web pages, and APIs into searchable knowledge bases for LLMs.

How this skill works

The skill ingests data via 300+ connectors (files, web, databases, APIs) and creates indices (vector, list, tree) to structure content for semantic search. Query engines and retrievers find relevant chunks, and optional chat engines provide multi-turn conversational memory. Agents can wrap query engines as tools and combine document search with custom functions or calculators for hybrid RAG agents.

When to use it

• Building RAG applications or document question-answering over private data
• Creating knowledge bases or search-driven chatbots from heterogeneous sources
• Ingesting many file types, web pages, databases, or API endpoints
• Needing structured outputs (Pydantic) or metadata-filtered retrieval
• Combining document search with tools in an agent for decision logic

Best practices

• Use vector indices for most semantic search use cases and persist storage to avoid re-indexing
• Chunk documents into ~512–1024 token pieces and add metadata for filtering and traceability
• Evaluate responses for relevance and faithfulness to detect hallucinations
• Enable streaming for better UX on long responses and verbose mode during development
• Choose an appropriate vector store (Chroma, Pinecone, FAISS) based on scale and latency

Example use cases

• Document Q&A system: index internal docs and answer user questions with source attribution
• Conversational chatbot with memory: multi-turn support using chat_engine (condense_plus_context)
• RAG agent: wrap a query engine as a tool and combine with calculation or lookup functions
• Multi-modal retrieval: ingest images and text to answer questions about diagrams or scanned pages
• Evaluation pipelines: run relevancy and faithfulness evaluators to benchmark responses