Document Backends#

Backends are Docling's format-specific parsing layer. They read raw input files and produce a DoclingDocument directly, without going through the layout-analysis or OCR stages used for PDFs.

There are two distinct backend categories:

• DeclarativeDocumentBackend — for structured text formats (CSV, HTML, Markdown, DOCX, XLSX, etc.). These implement a single convert() → DoclingDocument method and are driven by SimplePipeline.
• PaginatedDocumentBackend — for page-based formats (PDF, TIFF, images). These expose a page_count() method and feed into the StandardPdfPipeline for staged recognition.

All backends are defined under docling/backend/.

Abstract Base Classes#

docling/backend/abstract_backend.py defines the three-level hierarchy:

unload() on the base class closes any open BytesIO stream and sets path_or_stream = None 1, providing a consistent resource-cleanup contract across all backends.

SimplePipeline Integration#

SimplePipeline is the pipeline that drives all declarative backends. Its _build_document() method:

1. Asserts the backend is a DeclarativeDocumentBackend; raises RuntimeError otherwise 2.
2. Calls conv_res.document = conv_res.input._backend.convert() directly 3 — no page-level loop, no recognition models.

The class method is_backend_supported() enforces this at pipeline-selection time. There is no enrichment or post-processing beyond convert() for declarative formats.

Backend Inventory#

All declarative backends are registered in DocumentConverter._get_default_option() with the corresponding InputFormat and pipeline_cls=SimplePipeline.

Paginated backends — DoclingParseDocumentBackend (PDF variants), ImageDocumentBackend, and MetsGbsDocumentBackend — use StandardPdfPipeline instead and are out of scope here.

CSV Backend: Dialect Detection and Table Generation#

CsvDocumentBackend is a representative example of the declarative pattern, illustrating how a backend handles format-specific parsing concerns end-to-end.

Dialect detection (lines 58–73)
csv.Sniffer().sniff(head, ",;\t|:") inspects the first line to auto-detect the delimiter from the set ,, ;, \t, |, :. If sniffing fails (e.g., single-column data with insufficient signal), it falls back to csv.excel (comma-delimited).

Row-uniformity check (lines 95–102)
If rows have different column counts, a warnings.warn() is issued but conversion continues using the max column count for the table dimensions.

Table structure generation (lines 107–129)
Each CSV cell becomes a TableCell with row_span=1, col_span=1 (CSV has no merged cells). The first row is automatically flagged column_header=True 4. The entire CSV becomes a single TableItem added to the document via doc.add_table(data=table_data) 5.

The backend does not support pagination (supports_pagination() → False, 6) and declares supported_formats() → {InputFormat.CSV} 7.

Backend Options#

Format-specific options are passed through the backend_options field on each FormatOption subclass and are routed to the backend at instantiation time. All options classes live in docling/datamodel/backend_options.py 8.

Notable options classes relevant to declarative backends:

The CSV backend exposes no backend_options subclass — its only configurable behavior is the automatic dialect detection at parse time 9.

See the DocumentConverter Configuration knowledge base article for the full options hierarchy and usage examples.