rag-toolkit

Tiny RAG toolkit for Node.js — chunking, embeddings abstraction, metadata, search, and context building.

npm install rag-toolkit

Provides

• Chunking
• Embeddings abstraction
• Metadata
• Similarity search
• Context builder

Zero dependencies. Swap in OpenAI / Azure / local embeddings when you go to production.

Quick start

import { RAG } from 'rag-toolkit';

const rag = new RAG();

await rag.ingest(
  'RAG retrieves relevant docs and feeds them to an LLM as context...',
  { source: 'docs/rag.md', product: 'search' },
);

const { hits, context, prompt } = await rag.context('How does RAG work?', {
  topK: 3,
});

console.log(hits);
console.log(prompt);

Chunking

import { chunkText, chunkDocuments } from 'rag-toolkit';

chunkText(longText, { size: 500, overlap: 50 });

chunkDocuments(longText, { source: 'manual.pdf' }, { size: 500, overlap: 50 });
// [{ id, text, index, metadata }]

Embeddings abstraction

Built-in hash embedder (great for demos/tests):

import { createHashEmbedding, createEmbeddingProvider } from 'rag-toolkit';

const embedder = createHashEmbedding(384);
const vector = await embedder.embed('hello world');

Plug in your own model (OpenAI, Azure, Gemini, local):

const embedder = createEmbeddingProvider(async (text) => {
  // return number[] from your provider
  return await openai.embeddings.create(...).then(/* ... */);
});

const rag = new RAG({ embedder });

Metadata + search

import { VectorStore, createHashEmbedding } from 'rag-toolkit';

const embedder = createHashEmbedding();
const store = new VectorStore();

store.add({
  text: 'Redis eviction policies',
  metadata: { topic: 'db', tenant: 'acme' },
  embedding: await embedder.embed('Redis eviction policies'),
});

const hits = store.search(await embedder.embed('cache eviction'), {
  topK: 5,
  minScore: 0.1,
  filter: (meta) => meta.tenant === 'acme',
});

Context builder

import { buildContext, buildPrompt } from 'rag-toolkit';

const context = buildContext(hits, {
  maxChars: 3000,
  includeMetadata: true,
});

const prompt = buildPrompt('What is eviction?', hits);

API map

| Need | API | | -------------- | ------------------------------------------------ | | Chunking | chunkText, chunkDocuments | | Embeddings | createHashEmbedding, createEmbeddingProvider | | Store / search | VectorStore, RAG#search | | Context | buildContext, buildPrompt, RAG#context | | All-in-one | new RAG() |

License

MIT

Introduction

rag-toolkit helps you ship reliable Node.js / TypeScript applications with a small, focused API.

Why this package exists

Popular stacks need small, trustworthy utilities with excellent DX. rag-toolkit exists to solve one problem well: clear APIs, strong typing, minimal dependencies, and production-ready defaults — without the overhead of larger frameworks.

Installation

npm install rag-toolkit
# or
pnpm add rag-toolkit
yarn add rag-toolkit

Requires Node.js 18+.

API Reference

See the exports from rag-toolkit and the inline TypeScript types for the full surface area. Primary entry points are documented in Quick Start and Examples above.

Examples

Minimal usage is shown in Quick Start. Prefer copying those snippets first, then expand into your app’s error handling and configuration patterns.

Advanced Examples

• Combine with environment validation, logging, and health checks in production services
• Prefer dependency injection / custom fetch / client injection in tests
• Keep configuration explicit; avoid hidden global state

Framework Integration

Works with Express, Fastify, Hono, NestJS, and plain Node HTTP servers. Import ESM (or CJS where published) and call the documented APIs from route handlers, middleware, or background jobs.

TypeScript Usage

import { /* symbols */ } from "rag-toolkit";

Types ship with the package (types / exports.types). Enable strict in your tsconfig for the best DX.

Error Handling

• Fail fast with typed / named errors where provided
• Never swallow errors silently in production paths
• Prefer returning structured error payloads in HTTP layers
• Surface actionable messages (what failed + how to fix)

Performance

• Minimal runtime work on the hot path
• Avoid unnecessary allocations and dependencies
• Tree-shakeable ESM entry points
• Prefer streaming / lazy work when dealing with large payloads

Best Practices

• Pin major versions with SemVer ranges you trust
• Validate configuration at process startup
• Add health checks and observability around I/O
• Write tests for failure modes (timeouts, bad input, missing credentials)

FAQ

Does it work with ESM and CommonJS?
Yes where the package publishes dual exports. Prefer ESM for new projects.

Is it production-ready?
Yes — tests, types, and SemVer releases are part of the maintenance model.

How do I report a bug?
Open a GitHub issue using the bug template.

Migration Guide

From 0.x / early drafts

This package follows SemVer. Breaking changes land in major releases and are called out in CHANGELOG.md.

Upgrading patch/minor

Patch and minor releases are backward compatible. Run your test suite after upgrading.

Troubleshooting

| Symptom | Likely cause | Fix | |---------|--------------|-----| | ERR_MODULE_NOT_FOUND | Wrong Node version / bad import path | Use Node 18+ and package exports | | Types not resolving | Old moduleResolution | Use bundler or node16+ | | Auth / network failures | Missing env or blocked egress | Check credentials and firewall | | Unexpected runtime errors | Invalid input | Validate options; read error message |

Contributing

See CONTRIBUTING.md. PRs with tests and docs are welcome.