LlamaIndex RAG

End-to-end RAG using llamaindex + @nalinor/mupdf4llm/llama.

npm i @nalinor/mupdf4llm llamaindex @llamaindex/openai

import { PDFMarkdownReader } from "@nalinor/mupdf4llm/llama";
import { Settings, VectorStoreIndex } from "llamaindex";
import { OpenAI, OpenAIEmbedding } from "@llamaindex/openai";

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

const reader = new PDFMarkdownReader();
const docs = await reader.loadData("paper.pdf");

const index = await VectorStoreIndex.fromDocuments(docs);
const engine = index.asQueryEngine();

const answer = await engine.query({
  query: "What's the main contribution of the paper?",
});
console.log(answer.message);

Each page becomes a Document. Per-page metadata (page, total_pages, title, file_path) is attached automatically, so the LLM can cite source pages in its answer.

Filtering metadata

If you want to control what makes it into the embedding model's context window:

const reader = new PDFMarkdownReader({
  metaFilter: (m) => ({
    page: m.page,
    title: m.title,
    file_path: m.file_path,
  }),
});

Forwarding extraction options

To embed images or extract per-word coords:

const docs = await reader.loadData(
  "paper.pdf",
  /* extraInfo */ { collection: "papers/2026" },
  /* MarkdownOptions */ { embedImages: true },
);