Getting started

RAGLite turns documents into a searchable index, then answers questions from it with any LLM. The TypeScript package needs Node.js 20 or later; the Python package needs Python 3.11 or later.

Install

The core package is small. Add the SDK for each provider you use; Choosing providers lists them all.

TypeScript

npm install raglite-toolkit
npm install @huggingface/transformers   # local embeddings (the default), no API key
npm install @ai-sdk/anthropic           # Claude answers

Python

pip install 'raglite-toolkit[local,anthropic]'   # local embeddings (the default) + Claude answers
pip install 'raglite-toolkit[all]'               # every provider

The local extra installs sentence-transformers with PyTorch. The all-MiniLM-L6-v2 model (about 90 MB) is downloaded on first use.

If a package is missing, RAGLite throws a ConfigError that names the command to run.

Build, search, ask

TypeScript

import { Document } from "raglite-toolkit";

const doc = new Document("./policy.pdf", {
  llm: { provider: "anthropic", apiKey: process.env.ANTHROPIC_API_KEY },
});

await doc.build(); // load → chunk → embed → save

const hits = await doc.search("refund policy", { topK: 3 });
for (const hit of hits) console.log(hit.score.toFixed(2), hit.metadata.source, hit.text);

const answer = await doc.ask("What is the refund policy?");
console.log(answer.text); // "Refunds are issued within 30 days [1]."
console.log(answer.citations); // the passages behind [1], [2], ...

Python

import os
from raglite import Document

doc = Document("./policy.pdf", {
    "llm": {"provider": "anthropic", "apiKey": os.environ["ANTHROPIC_API_KEY"]},
})

doc.build()  # load → chunk → embed → save

hits = doc.search("refund policy", top_k=3)
for hit in hits:
    print(round(hit.score, 2), hit.metadata.source, hit.text)

answer = doc.ask("What is the refund policy?")
print(answer.text)       # "Refunds are issued within 30 days [1]."
print(answer.citations)  # the passages behind [1], [2], ...

• build() reuses the saved index when nothing has changed, so calling it on every start is cheap. See How caching works.
• A Document that was never built throws FileNotIndexedError from search() and ask(). A saved index from an earlier run is loaded without calling build().
• Without an apiKey, each provider reads its usual environment variable, such as ANTHROPIC_API_KEY. See Choosing providers.

Supported files: PDF, DOCX, Markdown, TXT and JSON, plus web pages by URL. To index many files at once, see Collections.

Index text you already have

No file needed: index a database row, an upload or CMS content. The id names the index (reuse it to reuse the index) and becomes each chunk’s source. Changed text is re-indexed on the next build().

TypeScript

const doc = Document.fromText("faq-42", faqText, { embeddings: { provider: "openai" } });
await doc.build();

Python

doc = Document.from_text("faq-42", faq_text, {"embeddings": {"provider": "openai"}})
doc.build()

Where the index is stored

Indexes are saved under ./.raglite, one folder per document, so two documents never overwrite each other. Set storeDir to keep them elsewhere, or use a vector database. Add .raglite/ to your .gitignore.

Next steps

• Choosing providers: embeddings and LLMs, API keys, running fully local.
• Citations: pages and sections for each cited passage.
• MCP server: give your documents to Claude Code, Cursor or VS Code.