@llamaindex/liteparse-pi-extension

A pi-coding-agent extension that exposes LiteParse to the agent as a set of first-class tools.

LiteParse is a fast, local, open-source document parsing library written in Rust: PDFs, scanned images, and Office files in → text + spatial layout (bounding boxes) out. No LLMs, no cloud calls. This extension lets a pi agent reach for those capabilities the same way it would reach for read or grep.

Tools

| Tool | What it does | | -------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | liteparse_parse | Parse a local PDF / image / Office file and return its extracted text. Supports targetPages (e.g. "1-3,7"), optional OCR, custom DPI, passwords, and an includeItems flag for per-text-item bounding boxes. Text output is truncated to ~50 KB / 2000 lines. | | liteparse_screenshot | Render one or more pages of a local document as PNG images and return them as image content blocks the model can look at directly. Useful when text extraction isn't enough — figures, tables, signatures, layout. | | liteparse_search | Search a parsed document for a phrase and return every hit with its page number and bounding box. Great for building visual citations or jumping to the source of a quoted snippet. |

All three tools come with promptGuidelines so the model knows, for example, to prefer liteparse_parse over read when the user points it at a PDF.

Install / run

This repo uses Bun.

bun install

Run pi with this extension loaded:

bunx pi -e ./src/index.ts

The repo's package.json also declares the extension under the pi key, so if you install it as a dependency in another pi project it will be picked up automatically:

{
  "pi": {
    "extensions": ["./src/index.ts"],
  },
}

Example session

> summarize page 1 of data/whitepaper.pdf

[agent calls liteparse_parse with { path: "data/whitepaper.pdf", targetPages: "1" }]

The first page introduces …

> look at the chart on page 4 of data/report.pdf and tell me the y-axis label

[agent calls liteparse_screenshot with { path: "data/report.pdf", pages: [4] }]

The y-axis is labeled "Revenue (USD millions)".

> find every mention of "Section 4.2" in data/spec.pdf

[agent calls liteparse_search with { path: "data/spec.pdf", phrase: "Section 4.2" }]

Found 6 hits:
  p12  [102.3, 540.1  78.4×11.0]   Section 4.2
  p13  [102.3, 312.7  78.4×11.0]   Section 4.2
  …

Development

# Type-check
bun run typecheck

# Format
bun run format

# Lint
bun run lint

How it works

src/index.ts defines three tools with defineTool(...) from @earendil-works/pi-coding-agent, then registers them in the default export:

export default function (pi: ExtensionAPI) {
  pi.registerTool(parseTool);
  pi.registerTool(screenshotTool);
  pi.registerTool(searchTool);
}

Each tool wraps the LiteParse class from @llamaindex/liteparse:

const parser = new LiteParse({ quiet: true, ocrEnabled, targetPages, ... });
const result = await parser.parse(absPath);

Paths are resolved against ctx.cwd, with a leading @ stripped to play nice with pi's @-prefixed file references. Screenshots come back as Buffers and are returned as base64 image content blocks ({ type: "image", data, mimeType: "image/png" }) so the model can see them directly.