ai-orchestration-llamaindex

LlamaIndex.TS Patterns

Quick Guide: LlamaIndex.TS is a data framework for building context-aware LLM applications in TypeScript. Use Settings singleton to configure LLM and embedding models globally. Load documents with SimpleDirectoryReader, chunk with SentenceSplitter, index with VectorStoreIndex.fromDocuments(), and query with index.asQueryEngine(). For agents, use agent() from @llamaindex/workflow with tool() definitions using Zod schemas. All core operations are async -- every function returns a Promise. The llamaindex package re-exports most things, but LLM providers require separate packages like @llamaindex/openai or @llamaindex/ollama.

---

<critical_requirements>

CRITICAL: Before Using This Skill

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST configure Settings.llm and Settings.embedModel before any indexing or querying -- the Settings singleton is lazily initialized and defaults to OpenAI, which will fail without an API key)

(You MUST await all LlamaIndex operations -- fromDocuments(), asQueryEngine(), query(), chat(), loadData() are ALL async)

(You MUST install provider packages separately -- @llamaindex/openai, @llamaindex/ollama, @llamaindex/anthropic are NOT included in the base llamaindex package)

(You MUST use storageContextFromDefaults({ persistDir }) to persist indexes -- without persistence, indexes are rebuilt from scratch on every restart)

(You MUST never hardcode API keys -- use environment variables and dotenv/config)

</critical_requirements>

---

Auto-detection: LlamaIndex, llamaindex, VectorStoreIndex, SimpleDirectoryReader, Settings.llm, Settings.embedModel, asQueryEngine, asChatEngine, ContextChatEngine, SentenceSplitter, storageContextFromDefaults, @llamaindex/openai, @llamaindex/ollama, @llamaindex/workflow, FunctionTool, QueryEngineTool, agentStreamEvent

When to use:

• Building RAG (Retrieval-Augmented Generation) applications with custom documents
• Loading, chunking, and indexing documents for LLM consumption
• Creating query engines that answer questions from indexed data
• Building chat interfaces with conversation memory over your data
• Implementing agentic RAG with tool-calling agents that query indexes
• Working with multiple data sources (files, PDFs, markdown, code)
• Persisting vector indexes to avoid re-indexing on every restart

Key patterns covered:

• Settings singleton for LLM and embedding model configuration
• Document loading with SimpleDirectoryReader and custom readers
• VectorStoreIndex creation, persistence, and querying
• Query engines and chat engines
• Agent creation with agent() and tool() using Zod schemas
• Text splitting and chunking strategies
• Streaming responses from query and chat engines
• Storage context and index persistence

When NOT to use:

• Simple one-shot LLM calls without document context -- use the LLM provider SDK directly
• Applications that only need embeddings without indexing -- use the embedding API directly
• Client-side / browser applications -- LlamaIndex.TS is server-side focused (Node.js >= 20)

---

Examples Index

• Core: Setup, Indexing & Querying -- Settings config, document loading, VectorStoreIndex, query engines, persistence
• Agents & Tools -- FunctionTool, agent(), multi-agent workflows, QueryEngineTool
• Chat & Streaming -- Chat engines, ContextChatEngine, streaming responses
• Ingestion & Splitting -- Text splitters, node parsers, ingestion pipeline, custom readers
• Quick API Reference -- Package map, method signatures, response modes, model providers

---

Philosophy

LlamaIndex.TS is a data framework -- its core value proposition is connecting your data to LLMs through indexing, retrieval, and synthesis. It sits between raw LLM APIs and full application frameworks.

Core principles:

1. Context engineering -- Inject the right data into the LLM prompt at the right time. This drives RAG, agent memory, extraction, and summarization.
2. Modular provider system -- LLM providers, embedding models, vector stores, and readers are separate packages you compose. The base llamaindex package provides the framework; providers are installed separately.
3. Settings singleton -- Global configuration for LLM, embedding model, node parser, and other shared resources. Set once, used everywhere. Override locally when needed.
4. Async-first design -- Every I/O operation is async. Document loading, indexing, querying, and chat all return Promises.
5. Index as the core abstraction -- Documents are loaded, split into nodes, embedded, and stored in an index. Queries retrieve relevant nodes and synthesize responses.

When to use LlamaIndex.TS:

• You have documents/data that need to be indexed for LLM consumption
• You want structured RAG pipelines with configurable retrieval and synthesis
• You need agentic RAG where agents query multiple indexes with tools
• You want persistence and incremental updates to your index

When NOT to use:

• Simple LLM calls without data context -- use the provider SDK directly
• Browser-only applications -- LlamaIndex.TS requires Node.js >= 20
• You only need embeddings -- use the embedding API directly

---

Core Patterns

Pattern 1: Settings Configuration

The Settings singleton configures LLM, embedding model, and node parser globally. Set it once at application startup before any indexing or querying.

import { Settings } from "llamaindex";
import { openai, OpenAIEmbedding } from "@llamaindex/openai";

// Configure at app startup -- before any index operations
Settings.llm = openai({ model: "gpt-4o" });
Settings.embedModel = new OpenAIEmbedding({ model: "text-embedding-3-small" });

Why good: Single configuration point, provider packages are explicit imports, model names are visible

// BAD: No Settings configuration, relying on implicit defaults
import { VectorStoreIndex, SimpleDirectoryReader } from "llamaindex";

// This will silently try to use OpenAI with OPENAI_API_KEY from env
// Fails with cryptic error if key is missing
const documents = await new SimpleDirectoryReader().loadData({
  directoryPath: "./data",
});
const index = await VectorStoreIndex.fromDocuments(documents);

Why bad: Implicit defaults make failures confusing, no explicit provider, no model selection

See: examples/core.md for local LLM setup with Ollama, Anthropic configuration, and embedding model options

---

Pattern 2: Document Loading and Indexing

Load documents, create a vector index, and query it. This is the canonical RAG pipeline.

import { SimpleDirectoryReader, VectorStoreIndex, Settings } from "llamaindex";
import { openai, OpenAIEmbedding } from "@llamaindex/openai";

Settings.llm = openai({ model: "gpt-4o" });
Settings.embedModel = new OpenAIEmbedding({ model: "text-embedding-3-small" });

// Load all supported files from a directory
const documents = await new SimpleDirectoryReader().loadData({
  directoryPath: "./data",
});

// Create vector index -- embeds and stores all document chunks
const index = await VectorStoreIndex.fromDocuments(documents);

// Query the index
const queryEngine = index.asQueryEngine();
const response = await queryEngine.query({ query: "What is the main topic?" });
console.log(response.message.content);

Why good: Complete pipeline in minimal code, explicit Settings, clear data flow

See: examples/core.md for persistence, custom readers, and advanced indexing options

---

Pattern 3: Index Persistence

Persist indexes to disk to avoid re-indexing on every restart.

import {
  VectorStoreIndex,
  storageContextFromDefaults,
  SimpleDirectoryReader,
} from "llamaindex";

const PERSIST_DIR = "./storage";

// First run: create and persist
const storageContext = await storageContextFromDefaults({
  persistDir: PERSIST_DIR,
});
const documents = await new SimpleDirectoryReader().loadData({
  directoryPath: "./data",
});
const index = await VectorStoreIndex.fromDocuments(documents, {
  storageContext,
});

// Subsequent runs: load from storage
const loadedStorageContext = await storageContextFromDefaults({
  persistDir: PERSIST_DIR,
});
const loadedIndex = await VectorStoreIndex.init({
  storageContext: loadedStorageContext,
});

Why good: Named constant for path, separate create vs load paths, storage context reuse

// BAD: Rebuilding index on every request
async function handleQuery(question: string) {
  const docs = await new SimpleDirectoryReader().loadData({
    directoryPath: "./data",
  });
  const index = await VectorStoreIndex.fromDocuments(docs); // Expensive!
  const engine = index.asQueryEngine();
  return engine.query({ query: question });
}

Why bad: Re-indexes all documents on every call, wastes time and API credits on re-embedding

See: examples/core.md for load-or-create pattern

---

Pattern 4: Agents with Tool Definitions

Create agents that use tools defined with Zod schemas. Use agent() from @llamaindex/workflow.

import { tool, Settings } from "llamaindex";
import { agent, agentStreamEvent } from "@llamaindex/workflow";
import { openai } from "@llamaindex/openai";
import { z } from "zod";

Settings.llm = openai({ model: "gpt-4o" });

const weatherTool = tool({
  name: "getWeather",
  description: "Get current weather for a city",
  parameters: z.object({
    city: z.string({ description: "City name" }),
  }),
  execute: async ({ city }) => {
    // Your weather API call here
    return { temperature: 22, condition: "sunny" };
  },
});

const myAgent = agent({ tools: [weatherTool] });
const result = await myAgent.run("What's the weather in Paris?");
console.log(result.data);

Why good: Zod schema for type-safe parameters, description guides the LLM, async execute function

See: examples/agents.md for multi-agent workflows, QueryEngineTool, streaming agents

---

Pattern 5: Chat Engine

Build conversational interfaces over your indexed data with conversation memory.

import {
  VectorStoreIndex,
  ContextChatEngine,
  SimpleDirectoryReader,
} from "llamaindex";

const documents = await new SimpleDirectoryReader().loadData({
  directoryPath: "./data",
});
const index = await VectorStoreIndex.fromDocuments(documents);
const retriever = index.asRetriever({ similarityTopK: 3 });

const chatEngine = new ContextChatEngine({ retriever });

// Multi-turn conversation -- chat engine maintains history
const response1 = await chatEngine.chat({ message: "What is LlamaIndex?" });
console.log(response1.message.content);

const response2 = await chatEngine.chat({
  message: "How does it handle streaming?",
});
console.log(response2.message.content);

Why good: Retriever-based context injection, automatic conversation history, multi-turn support

See: examples/chat-streaming.md for streaming chat, system prompts, chat history management

---

Pattern 6: Streaming Responses

Stream responses for user-facing applications.

import { agentStreamEvent } from "@llamaindex/workflow";

// Agent streaming
const events = myAgent.runStream("Tell me about TypeScript");
for await (const event of events) {
  if (agentStreamEvent.include(event)) {
    process.stdout.write(event.data.delta);
  }
}

// Query engine streaming
const response = await queryEngine.query({
  query: "Summarize the document",
  stream: true,
});
for await (const chunk of response) {
  process.stdout.write(chunk.message.content);
}

Why good: Event-based agent streaming with typed filters, query engine streaming with for-await

See: examples/chat-streaming.md for response synthesizer streaming, chat engine streaming

---

Pattern 7: Text Splitting and Node Parsing

Configure how documents are chunked before indexing.

import { SentenceSplitter, Settings } from "llamaindex";

const CHUNK_SIZE = 512;
const CHUNK_OVERLAP = 50;

// Set globally via Settings
Settings.nodeParser = new SentenceSplitter({
  chunkSize: CHUNK_SIZE,
  chunkOverlap: CHUNK_OVERLAP,
});

// Or use standalone
const splitter = new SentenceSplitter({ chunkSize: CHUNK_SIZE });
const texts = splitter.splitText("Your long document text here...");

Why good: Named constants for chunk parameters, global vs standalone usage shown, sentence-aware splitting

// BAD: Using default chunk size without considering document characteristics
const index = await VectorStoreIndex.fromDocuments(documents);
// Default chunk size may be too large for short Q&A or too small for long narratives

Why bad: Default chunk size (1024 tokens) may not suit your data, causes poor retrieval quality

See: examples/ingestion.md for MarkdownNodeParser, CodeSplitter, custom chunk strategies

---

<decision_framework>

Decision Framework

Which Index Type to Use

What is your use case?
+-- Semantic search over documents -> VectorStoreIndex (most common)
+-- Summarization of all documents -> SummaryIndex
+-- Both search AND summarization -> Create both, use as separate tools in an agent
+-- Hierarchical document structure -> Use MarkdownNodeParser + VectorStoreIndex

Query Engine vs Chat Engine vs Agent

How should users interact with your data?
+-- Single question, single answer -> Query Engine (index.asQueryEngine())
+-- Multi-turn conversation -> Chat Engine (ContextChatEngine)
+-- Multiple tools/indexes + reasoning -> Agent (agent() from @llamaindex/workflow)
+-- Complex multi-step workflow -> Multi-agent with handoffs

Which LLM Provider

Which LLM provider are you using?
+-- OpenAI -> npm install @llamaindex/openai
+-- Anthropic -> npm install @llamaindex/anthropic
+-- Local (Ollama) -> npm install @llamaindex/ollama
+-- Groq -> npm install @llamaindex/groq
+-- Google Gemini -> npm install @llamaindex/gemini

Chunk Size Selection

What kind of documents are you indexing?
+-- Short Q&A pairs -> chunkSize: 256-512
+-- Technical documentation -> chunkSize: 512-1024
+-- Long narratives/reports -> chunkSize: 1024-2048
+-- Code files -> Use CodeSplitter (AST-aware)
+-- Markdown -> Use MarkdownNodeParser (structure-aware)

</decision_framework>

---

<red_flags>

RED FLAGS

High Priority Issues:

• Not configuring Settings.llm before indexing/querying -- defaults to OpenAI, fails silently without API key
• Forgetting to await async operations -- fromDocuments(), query(), chat() all return Promises
• Rebuilding indexes on every request instead of persisting with storageContextFromDefaults
• Hardcoding API keys instead of using environment variables
• Installing only llamaindex without provider packages (@llamaindex/openai, etc.)

Medium Priority Issues:

• Using default chunk size (1024) without considering document characteristics -- causes poor retrieval
• Not setting similarityTopK on retrievers -- default may return too few or too many results
• Ignoring the response sourceNodes -- they contain the retrieved context for debugging and citations
• Creating a new SimpleDirectoryReader per request instead of caching the loaded documents
• Not handling the case where response.message.content might be empty on retrieval failure

Common Mistakes:

• Confusing asQueryEngine() (single question) with ContextChatEngine (multi-turn conversation)
• Using VectorStoreIndex.fromDocuments() when you should use VectorStoreIndex.init() to load from storage
• Importing openai from llamaindex instead of @llamaindex/openai -- the llamaindex package may re-export some things but provider-specific imports are more reliable
• Passing messages array to query() -- query engines take { query: string }, not a messages array
• Using index.asQueryEngine() multiple times instead of storing the engine reference

Gotchas & Edge Cases:

• Settings is a global singleton -- setting it in one module affects all others. Override locally by passing llm directly to constructors when you need different models for different operations.
• SimpleDirectoryReader only works on Node.js -- it uses fs internally. For edge/serverless, load documents differently or use LlamaParse.
• storageContextFromDefaults creates four JSON files in the persist directory (docstore.json, graph_store.json, index_store.json, vector_store.json). If any are corrupted, delete the directory and re-index.
• Node.js >= 20 is required. Some modules use Web Stream API (ReadableStream, WritableStream), so add "DOM.AsyncIterable" to tsconfig.json lib if you get type errors.
• tsconfig.json must use "moduleResolution": "bundler" or "nodenext" -- the classic "node" resolution will fail to resolve LlamaIndex sub-packages.
• Default tokenizer is slow -- install gpt-tokenizer for 60x faster tokenization.
• SentenceSplitter chunk size is in tokens, not characters. A 512-token chunk is roughly 2000 characters.
• The llamaindex package is large (~2MB+). For production, consider importing specific sub-packages to reduce bundle size.
• VectorStoreIndex.fromDocuments() makes embedding API calls for every chunk. For large document sets, this can be expensive. Monitor costs.
• Chat engine conversation history grows unbounded -- implement history pruning for long-running sessions.

</red_flags>

---

<critical_reminders>

CRITICAL REMINDERS

All code must follow project conventions in CLAUDE.md (kebab-case, named exports, import ordering, import type, named constants)

(You MUST configure Settings.llm and Settings.embedModel before any indexing or querying -- the Settings singleton is lazily initialized and defaults to OpenAI, which will fail without an API key)

(You MUST await all LlamaIndex operations -- fromDocuments(), asQueryEngine(), query(), chat(), loadData() are ALL async)

(You MUST install provider packages separately -- @llamaindex/openai, @llamaindex/ollama, @llamaindex/anthropic are NOT included in the base llamaindex package)

(You MUST use storageContextFromDefaults({ persistDir }) to persist indexes -- without persistence, indexes are rebuilt from scratch on every restart)

(You MUST never hardcode API keys -- use environment variables and dotenv/config)

Failure to follow these rules will produce broken RAG pipelines, wasted embedding API credits, or cryptic runtime errors.

</critical_reminders>