api-ai-llamaindex

What it does

Guides building context-aware LLM applications in TypeScript using LlamaIndex.TS. It instructs configuring LLM and embedding models via a Settings singleton, loading documents with SimpleDirectoryReader, splitting with SentenceSplitter, indexing with VectorStoreIndex.fromDocuments(), and querying with index.asQueryEngine(). For agents, it uses agent() with tool() definitions backed by Zod schemas, and supports chat engines, streaming, and multi-step workflows.

How it works

The skill describes a sequence: configure Settings.llm and Settings.embedModel, load documents via SimpleDirectoryReader, create a VectorStoreIndex from those documents, then obtain a query engine via index.asQueryEngine() to run queries. For agents, define tools with tool(...) using a Zod schema for parameters, then instantiate an agent with agent({ tools: [...] }) and run it. For chat, instantiate a ContextChatEngine from a retriever derived from the index and call .chat(...) for multi-turn conversations. It emphasizes all operations are async (Promises), and Persistence should be handled using storageContextFromDefaults({ persistDir }) when creating or loading the index. It also notes you must install provider packages separately (e.g., @llamaindex/openai).

When to use it

• Building Retrieval-Augmented Generation (RAG) applications with custom documents
• Loading, chunking, and indexing documents for LLM consumption
• Creating query engines or chat engines over indexed data
• Implementing agentic RAG with tool-calling agents that query indexes
• Working with multiple data sources and persisting vector indexes across restarts

What it can touch

• Uses Settings.llm and Settings.embedModel for global config
• Uses SimpleDirectoryReader for document loading
• Uses VectorStoreIndex for indexing and persistence
• Exposes asQueryEngine() for querying and ContextChatEngine for chat
• Supports agent tooling via tool(...) and agent()
• Utilizes storageContextFromDefaults({ persistDir }) for persistence

Caveats

• Requires explicit configuration of the LLM provider packages (e.g., @llamaindex/openai) as they are not included in the base llamaindex package
• All core operations are async and must be awaited
• API keys must be supplied via environment variables rather than hardcoding
• Server-side only; Node.js >= 20 required; not intended for client-side/browser usage