Introducing Agents.·Read
OverviewAgentsWorkflowsExecutionDeveloper ToolsSecurityLibrary
Updated Oct 19, 2026

Utilities

Extract positioned text from one PDF

liteparse is a local PDF parser that returns native text and positioned text items as JSON.

Agent
Read PDF page textKeep text bounding boxesInspect returned page dimensions

overview

What liteparse does

Parse native text from the first 100 pages of one PDF into LiteParse JSON with page dimensions and text bounding boxes.

The packaged CLI uses PDFium locally. OCR and remote services are disabled, so scanned pages without native text can return empty text. total_pages describes the document while pages contains returned pages.

Input is limited to 50 MiB and parsing to 120 seconds on Linux amd64. PDF magic bytes identify the file; normalized filenames are accepted. Input bytes remain unchanged and failed parsing leaves no result.

This packaged operation uses 2.14.7. Upstream license: Apache-2.0. The linked source identifies the actual program; this page documents the adapter's exposed contract.

source github.com/run-llama/liteparse

use cases

Where liteparse fits

Read PDF page text

Extract native text without OCR or a remote parsing service.

Keep text bounding boxes

Retain positioned text items for a layout-aware next step.

Inspect returned page dimensions

Use returned page dimensions beside the text items.

reference

liteparse inputs and flags

1 inputs
NameTypeFlagDescription
inputFILE--inputLocal PDF file with native text, at most 50 MiB; only the first 100 pages are parsed

Showing key inputs. liteparse exposes 1 inputs in total.

example

Run liteparse

liteparse · command
/app/run.sh --input ./input.txt ./output

guidance

Choosing liteparse

Choose native text parsing when text positions matter. Embedded image extraction and rendered page screenshots produce different artifacts; image-only scans need a separate OCR step.

pdftotext

Returns plain PDF text rather than positioned JSON items.

faq

liteparse questions

input: Local PDF file with native text, at most 50 MiB; only the first 100 pages are parsed

Positioned text JSON. The packaged CLI uses PDFium locally. OCR and remote services are disabled, so scanned pages without native text can return empty text. total_pages describes the document while pages contains returned pages.

Input is limited to 50 MiB and parsing to 120 seconds on Linux amd64. PDF magic bytes identify the file; normalized filenames are accepted. Input bytes remain unchanged and failed parsing leaves no result.

Choose native text parsing when text positions matter. Embedded image extraction and rendered page screenshots produce different artifacts; image-only scans need a separate OCR step.

The public image repository includes download instructions for this version and a separate source/notices bundle. Its overview provides an immutable ORAS command, the expected bundle checksum and pinned upstream source directions. The bundle contains the original build recipe, runtime sources and retained component license notices. Image, source and license notice downloads.

related

More Utilities tools

  1. browser-fetch

    Headless Chromium render: requested URL, final URL, status, and title next to the page.

  2. cewl

    Spider a URL and return a wordlist for password crackers.

  3. csvkit

    SQL on one CSV. The table name is the file stem.

  4. dnsgen

    Wordlist and mined-word subdomain permutation.

  5. duckdb

    In-process SQL on CSV, JSON, and Parquet. No server.

  6. edge-tts

    Microsoft Edge neural speech from text or a script file. No API key.

Run liteparse yourself

The saved synthetic one-page PDF check verified sentinel text, page dimensions and nonzero text bounding boxes. The full native JSON is not retained as a page sample. The command shows the container entrypoint with example paths.

Open liteparse in the libraryBuild a workflow with it

Facts on this page come from the live Trickest tool library.