OmniParse
Ingests and parses unstructured documents, images, audio, video, and web pages into structured markdown for GenAI/RAG. Fully local, no external APIs, but GPU-heavy and GPL-3.0. 7.6K stars. Source: adithya-s-k/omniparse README, GitHub
OmniParse is a self-hostable ingestion platform that takes ~20 file types and returns high-quality structured markdown ready for LLM applications. Its differentiator is that everything runs locally, no API keys, and it fits in a single T4 GPU. Source: adithya-s-k/omniparse README, GitHub
What It Does
- Documents: PDF, DOCX, PPT/PPTX via the Surya OCR model series plus Florence-2 for vision.
- Media: audio/video transcription via Whisper.
- Web: dynamic web-page crawling via a Selenium crawler.
- Extraction: tables, image extraction and captioning, parallel crawling, caching.
- Deploy: Docker (
savatar101/omniparse:0.1) and SkyPilot, with a Gradio UI and a FastAPI surface (/parse_document,/parse_document/pdf,/parse_document/ppt,/parse_document/docs,/parse_media/image,/parse_media/video,/parse_media/audio,/parse_website). The server is Linux-only.
Source: adithya-s-k/omniparse README, GitHub
Version And Adoption Snapshot
Checked 2026-06-30:
| Surface | Current |
|---|---|
| GitHub | adithya-s-k/omniparse, GPL-3.0, 7,619 stars, 650 forks |
| Source revision | main HEAD 9d1ae83c46de777427e67b48f82eaca45ad7994a; no public Git tags or GitHub releases |
| Package surface | No npm/PyPI package checked into routing; Docker Hub image documented as savatar101/omniparse:0.1 |
| Skill decision | No upstream Agent Skill found. Keep as a tool route, not a local executable skill. |
| Commercial caution | The repo is GPL-3.0 and the README notes Marker/Surya OCR model-weight commercial restrictions; run license review before production use. |
The artifact screenshot is a GitHub README capture, not a product demo. It confirms the positioning and license badge, but not runtime quality. Source: X bookmark artifact, 2026-05-25; GitHub API/README, 2026-06-30
OmniParse vs MarkItDown
Both convert documents to markdown for LLMs, but they sit at opposite ends of the cost/quality curve. MarkItDown - Universal File-to-Markdown Converter (Microsoft) is a lightweight pure-Python utility, fast and dependency-light, ideal for clean text-based formats. OmniParse is heavier: it loads OCR, vision, and speech models and wants a GPU, in exchange for true multimodal parsing (scanned PDFs, image captioning, audio/video transcription, web crawling). Pick MarkItDown for quick text extraction; pick OmniParse when the corpus is image-heavy, multimedia, or needs OCR. Both feed RAG pipelines.
Install candidate for hard multimodal corpora only. Run the three-gate chain before adopting, per AGENTS.md, and treat license/commercial constraints as part of the gate.
Timeline
- 2026-06-30 | Deep review verified OmniParse as a tool route, not a skill import:
mainHEAD9d1ae83c46de777427e67b48f82eaca45ad7994a, no public tags/releases, 7,619 stars, GPL-3.0, documented Docker imagesavatar101/omniparse:0.1, and no upstream Agent Skill. Source: GitHub API/README, 2026-06-30 - 2026-05-25 | @tom_doerr surfaced OmniParse ("Converts documents and media into structured JSON for LLMs"), 1,156 likes, 1,391 bookmarks. Repo at ~7.5K stars. Source: X/@tom_doerr, 2026-05-25