Breaking

Self-hosted open-source "text-extract-api" turns documents and images into JSON or Markdown

July 2, 2026 at 06:17 EDT

Open Source
AI Agents
Software Dev & Coding

An open-source API called "text-extract-api" that converts PDFs, images and Office documents into structured JSON or Markdown using OCR and LLMs is drawing attention among developers. Published by CatchTheTornado, the project runs fully locally and even ships with a feature to strip personally identifiable information (PII) (GitHub).

Open Source · CatchTheTornado

text-extract-api: Documents → Structured Markdown & JSON, Fully Local

An open-source API that pairs OCR with locally run LLMs to turn PDFs, images and scans into layout-preserving Markdown or prompt-defined JSON — no cloud required.

switchable OCR strategies

30+

languages via EasyOCR

90B

Llama 3.2 Vision — high accuracy, slow

self-hosted, MIT-centered

The pipeline: from raw file to meaning

Input

PDF · Word · PPTX · images · scans

→

OCR / Vision

EasyOCR · MiniCPM-V · Llama Vision

→

LLM via Ollama

correction · PII removal · structuring

→

Output

Markdown or JSON with tables & formulas

Four OCR strategies, chosen by use case

Selectable via CLI or URL parameter

easyocr

30+ languages, strong in English

Apache

minicpm_v

MiniCPM-V, requires Ollama pull

Apache

llama_vision

Llama 3.2 Vision (90B), high accuracy but slow

Meta Community

remote

Uses external APIs such as Marker

Separate server

Why it appeals

Context & layout kept vs. OCR-only
Tables, numbers and formulas preserved
PII removal for invoices & medical reports
Fully self-hosted — Docker Compose, no cloud

The trade-offs

Large 90B models tend to be slow
Ollama models & Docker/GPU setup take effort
Possible accuracy gap vs. cloud services
No pricing or official benchmarks published

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.