Proposal: Add SoMark as a layout-aware PDF parser

[limuting233]

Hi maintainers,

I'm a developer at SoMark, an async document-parsing SaaS. I'd like to contribute SoMark as a new layout-aware PDF parser for RAGFlow, following the exact integration pattern of MinerU / Docling / PaddleOCR / OpenDataLoader.

A working branch rebased onto current main is ready, feel free to skim before deciding:
👉 limuting233/ragflow @ lty/feat/somark-model-service

What SoMark adds

A commercial document-parsing engine specializing in:

• High-quality layout detection on academic / report / form PDFs
• Image understanding — figures returned with auto-generated captions, which makes figures retrievable through semantic search (otherwise figures are typically lost in text-only embeddings)
• Native handling of equations, tables, stamps, QR codes, and chemical structures

Integration approach (PR-ready)

Strictly aligned with the existing OCR parser pattern, no new infrastructure:

• SoMarkOcrModel registered through the existing _FACTORY_NAME auto-registration — same as MinerU
• Hooked into both ingestion paths:
  • rag/app/naive.py (Built-in chunking, 2-tuple sections)
  • rag/flow/parser/parser.py (Pipeline DAG, 3-tuple sections — mirrors MinerU's parse_method="pipeline" contract exactly)
• Frontend: SoMark appears in the existing OCR model dropdown (same LayoutRecognizeFormField flow) plus a config modal mirroring MinerU's layout

Scope and non-goals

• PDF only — consistent with every other OCR parser in the codebase
• Spreadsheet / DOCX / image inputs are explicitly out of scope for this PR. SoMark supports them, but extending per-file-type dispatch to all OCR parsers is a separate, larger discussion that should benefit every parser, not just SoMark.
• No changes to tokenize_chunks or other shared chunking logic

Cost to RAGFlow operators: zero

• Each tenant brings their own SoMark API key via the standard model-provider flow — same as MinerU's API mode
• No bundled credentials, no rate-limit or billing liability on the RAGFlow side

End-to-end verified on a 69-page academic PDF

Path	Result
Built-in chunking (naive.py)	indexed, retrievable, figures cropped with captions
Pipeline DAG (Parser → Chunker → Tokenizer)	indexed, retrievable via q_*_vec
SoMark structural detection	21 figures + 18 tables + 24 equations that DeepDOC's text-only extraction would lose

Image-understanding captions enabled semantic retrieval on figures — e.g. searching "how many chapters does the paper have" correctly hit the figure-only "thesis structure" diagram.

Test coverage

• Integration test (test_somark_integration.py): 8 sections covering imports, factory registration, three configuration shapes, end-to-end parse, both L1/L2 chains, and DB-free image crop — all passing
• A pytest version (no-key auto-skip + mocked unit tests for CI) will be included in the PR

Ask

Would you welcome a PR for this? Happy to:

• Adjust scope or factor out anything that feels too specific
• Provide demo screenshots / a short video walkthrough
• Iterate on naming, config schema, or docs structure before opening the PR

Thanks for considering!

[yingfeng]

PRs are welcome! Additionally, RAGFlow is currently migrating from Python to Golang for all online retrieval services and select offline jobs. We encourage you to also provide a Golang model provider in accordance with this PR. Thank you!

[limuting233]

Thanks! Will open the PR shortly. Re: the Go counterpart — noted, happy to follow up on that as a separate contribution after the Python integration lands.