Converting bytes that were never a file

Everything so far has been a path. A file arriving over HTTP is not a path, and writing it to disk just to convert it is a step nobody wants.

DocumentStream wraps bytes with a name. The name is not decoration: it is what Docling reads to decide which backend to use.

from io import BytesIO
from docling.datamodel.base_models import DocumentStream
from docling.document_converter import DocumentConverter

raw = b"# Notice\n\nDelivery is paused on public holidays.\n"
upload = DocumentStream(name="notice.md", stream=BytesIO(raw))
print(DocumentConverter().convert(upload).document.export_to_markdown())

No file was written. The same bytes with the name notice.html would go to the HTML backend and come out as one paragraph starting with a hash, because nothing about the bytes says Markdown.

Proving that

upload = DocumentStream(name="notice.html", stream=BytesIO(raw))
document = DocumentConverter().convert(upload).document
print(document.texts[0].label, "|", document.texts[0].text)

One text item holding the hash as a literal character, where the Markdown reading gave a title. Same bytes, different name, different document. In a web service that name comes from whatever the browser uploaded, so it is worth checking rather than trusting.

The result still knows the name

result = DocumentConverter().convert(
    DocumentStream(name="notice.md", stream=BytesIO(raw)))
print(result.input.file.name)
print(result.document.origin.filename)

So the citation machinery from lesson 14 works on an upload exactly as it does on a file. The handbook in lesson 29 never needs this, but a service built on the same code would use nothing else.