● Beta — KubeEra is in active development. Profiles are live; AI-perception measurement is rolling out.
KubeEra
D

Docling

AI Agents
Open source · MIT

Docling parses and converts documents — PDF, DOCX, PPTX, XLSX, HTML, images — into structured formats (Markdown, JSON, HTML) that are actually usable by LLM pipelines.

updated 2026-07-01

What it is

Docling parses and converts documents — PDF, DOCX, PPTX, XLSX, HTML, images — into structured formats (Markdown, JSON, HTML) that are actually usable by LLM pipelines. It handles layout analysis, table structure recognition, reading order, and OCR for scanned content, then exports to a unified document representation. The pitch is "get your documents ready for gen AI" — meaning it sits in the ingestion layer of a RAG or agent pipeline, not in the model or retrieval layer itself.

It's tagged as an "AI Agent" category tool here, but functionally it's a document conversion/parsing library. Agent frameworks call it as a tool; it doesn't reason or plan on its own.

Who builds it and why

Docling comes out of IBM Research (Deep Search team). The motivation is straightforward: enterprise gen AI projects die on bad document parsing long before they die on bad prompts. IBM needed a robust, layout-aware conversion layer for its own watsonx and RAG offerings, and open-sourced the core engine under MIT. 249 contributors on a project of this scope suggests real external adoption beyond IBM's own use, not just a single-vendor side project.

Production readiness signal

62,450 GitHub stars and 249 contributors is a strong adoption signal for a document-processing library — this is well past "interesting weekend project" territory. Latest release v2.108.0 indicates a fast, mature release cadence (this isn't a project still on v0.x figuring out its API). Last commit dated 2026-07-01 shows active, ongoing maintenance rather than an abandoned repo with stale stars.

What's missing: no CNCF maturity signal (not applicable — this isn't a CNCF project), and founding date isn't publicly documented, so you can't easily gauge how long the core architecture has been battle-tested versus how much is recent rewrite. MIT license removes any adoption friction on the legal side.

Who should use this

  • Teams building RAG pipelines who are hitting quality ceilings because of naive PDF-to-text extraction (garbled tables, lost reading order, missing figures).
  • Python shops that want a single library handling multiple document formats instead of stitching together PyPDF2, python-docx, Tesseract, and custom table-parsing heuristics.
  • Anyone who needs table structure preserved (not just flattened text) for financial reports, scientific papers, or technical manuals feeding into an LLM.
  • Teams that need an open-source, self-hostable conversion layer for compliance reasons — no data leaves your infrastructure.

Who should NOT use this

  • Teams needing real-time, sub-100ms document processing at high throughput — layout models and OCR add latency; Docling isn't built for streaming use cases.
  • Non-Python shops. There's no first-class Go/Java/Node SDK; you're either running Python natively or wrapping it as a microservice.
  • Teams with simple, well-structured input (clean HTML, plain text, already-structured JSON) — Docling's layout/table intelligence is overhead you don't need.
  • Anyone expecting an end-to-end agent framework. Docling doesn't do retrieval, chunking strategy, embedding, or orchestration — you still need LangChain/LlamaIndex/Haystack (or your own code) around it.
  • Heavy scanned-document/handwriting OCR workloads where you need best-in-class OCR accuracy — Docling's OCR integrations (Tesseract, EasyOCR, RapidOCR) are solid but not purpose-built OCR specialists like ABBYY or Google Document AI.

Alternatives

  • Unstructured.io — broader format coverage and a hosted API option, but historically weaker table structure fidelity than Docling.
  • LlamaParse (LlamaIndex) — hosted, GPT-4V-backed parsing with strong table/chart extraction; not open source, usage-based pricing.
  • Apache Tika — battle-tested, format-agnostic extraction with a huge install base, but text-only output with much weaker layout/table understanding.

Pricing

Fully open source, MIT licensed. No paid tier, no hosted SaaS version from IBM as of now — you run it yourself. Costs are your own compute (CPU/GPU for layout models and OCR) and engineering time to integrate it into your pipeline.

Frequently asked

What is Docling?+
Docling parses and converts documents — PDF, DOCX, PPTX, XLSX, HTML, images — into structured formats (Markdown, JSON, HTML) that are actually usable by LLM pipelines. It handles layout analysis, table structure recognition, reading order, and OCR for scanned content, then exports to a unified document representation.
Who builds Docling?+
Docling comes out of IBM Research (Deep Search team). The motivation is straightforward: enterprise gen AI projects die on bad document parsing long before they die on bad prompts. IBM needed a robust, layout-aware conversion layer for its own watsonx and RAG offerings, and open-sourced the core engine under MIT.
Is Docling production ready?+
62,450 GitHub stars and 249 contributors is a strong adoption signal for a document-processing library — this is well past "interesting weekend project" territory. Latest release v2.108.0 indicates a fast, mature release cadence (this isn't a project still on v0.x figuring out its API).
Who should use Docling?+
Teams building RAG pipelines who are hitting quality ceilings because of naive PDF-to-text extraction (garbled tables, lost reading order, missing figures).
Who should not use Docling?+
Teams needing real-time, sub-100ms document processing at high throughput — layout models and OCR add latency; Docling isn't built for streaming use cases. Non-Python shops. There's no first-class Go/Java/Node SDK; you're either running Python natively or wrapping it as a microservice.
What are the alternatives to Docling?+
Unstructured.io — broader format coverage and a hosted API option, but historically weaker table structure fidelity than Docling. LlamaParse (LlamaIndex) — hosted, GPT-4V-backed parsing with strong table/chart extraction; not open source, usage-based pricing.
How much does Docling cost?+
Fully open source, MIT licensed. No paid tier, no hosted SaaS version from IBM as of now — you run it yourself. Costs are your own compute (CPU/GPU for layout models and OCR) and engineering time to integrate it into your pipeline.