I want to use Docling for legal text processing, and have some questions

[gustavotrapp]

I want to specifically use the HTML output, so that I can easily use a HTML library to extract specific texts and do post-processing

Every single document should follow the same patterns (headers, text layout, etc)

How much can I trust Docling with having consistent HTML outputs, because I will need to perform operations like:
"Getting all clauses"

Essentially my question is, does it stay consistent, is there any way I can make it output custom tags like, for example:
<clause></clause>

[dosubot[bot]]

A few things that might help here:

HTML consistency — Docling's export_to_html() produces a structured HTML output based on the document element types it detects (headings, paragraphs, tables, lists, etc.) [1]. For documents that follow the same layout, the output structure should be consistent, but it ultimately depends on how well Docling's layout model recognizes each element. If your legal documents have a uniform format (e.g., Word templates), you should get fairly reliable results.

Custom tags like <clause> — Docling does not natively support custom HTML tags. The HTML exporter uses a fixed set of standard HTML tags [1]. For your use case, the recommended approach would be to post-process the HTML (e.g., with BeautifulSoup) to identify and wrap clauses with custom tags.

Useful alternatives to consider:

1. doc.iterate_items() — You can iterate over all document items programmatically and build fully custom output with whatever tag structure you need [1]:

   for item, level in doc.iterate_items():
       # Apply your own clause-detection logic and emit custom HTML
       pass

2. Experimental Doclang serializer — There's an experimental XML-like serializer in docling_core/experimental/doclang.py that emits structural tags with layer annotations, which might give you more semantic structure to work with than plain HTML [1].

For legal clause extraction specifically, you'd likely want to combine Docling's structural parsing with your own post-processing logic — e.g., detecting clause patterns based on heading hierarchy and numbering, then wrapping them in <clause> tags after export.

To reply, just mention @dosu.

---

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