Is it possible for Docling to recognise Header depth through controls in Docling Serve?

[JonasPapinigis]

Currently in the progress of developing a custom chunker for OpenWebUI, a frontend which utilises Docling Serve. I've not been happy with the way OpenWebUI does chunking so I've written my langchain script.

Problem is, OWUI uses Docling Serve to ingest documents, and I'm finding that all headers are recognised as being of level 2 (##) which makes it extremely difficult to gather useful metadata, and is something I can't seem to find a way to fix after testing DS independantly.

Furthermore, headings that are multi-line are recognised seperately (an issue I fixed by fusing them together after the fact). It would also be nice to fix this too at digestion.

I'm really eager to not alter source code in order to not have to patch Docling Serve manually in production, but can do need-be.

Appreciate any help!

[dosubot]

Hi @JonasPapinigis! I'm Dosu and I’m helping the docling-serve team.

Docling Serve doesn't currently offer any configuration or controls to adjust header depth recognition or to fuse multi-line headings during ingestion or chunking. Header levels are determined by the parsing logic in the docling-core library, specifically by how the input document is structured and interpreted by chunkers like HierarchicalChunker. If all headers are coming through as level 2 or multi-line headings are split, that's due to the internal parsing heuristics, not a setting you can tweak in Docling Serve itself (source).

To change this behavior, you'd need to modify or extend the chunker logic in docling-core and expose those controls in Docling Serve, which would require patching the source. For now, post-processing remains necessary unless such changes are made upstream.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[dolfim-ibm]

Regarding chunking, you might be interested in the upcoming feature which will allow to return chunks (via the Docling chunkers) directly from the API.

The detection of the headers is currently not available in Docling. It will be when new models are ready. You can follow up about it in the Docling issues.