Document Parsing Pipeline#

DeepDoc is RAGFlow's in-house document understanding subsystem located under deepdoc/. It splits into two cooperating layers:

• deepdoc/parser/ — format-specific parsers that extract structured content from raw files
• deepdoc/vision/ — a vision pipeline (OCR, layout recognition, table structure recognition) used by the PDF parser and image handling paths

DeepDoc is distinct from the MinerU integration, which is a separate PDF-only backend that calls an external API server (see deepdoc/parser/mineru_parser.py).

Format-Specific Parsers#

All parsers are exported from deepdoc/parser/__init__.py:

Key implementation notes per format:

• PDF (pdf_parser.py) — RAGFlowPdfParser is the heaviest parser: it initializes OCR, LayoutRecognizer, TableStructureRecognizer, and an XGBoost model (updown_concat_xgb.model) for text-block concatenation heuristics. Pages are rendered via pdfplumber at 216 DPI (72 × zoomin=3) 1. The output includes text chunks with page/position metadata, tables converted to HTML/natural language, and figures with captions.

• DOCX (docx_parser.py) — Uses the python-docx library. Extracts paragraphs, tables, and embedded images (via XPath into the XML part). Image blob extraction catches InvalidImageStreamError, UnrecognizedImageError, etc.

• Excel/CSV (excel_parser.py) — Reads .xlsx/.xls with openpyxl and falls back to pandas on parse error. CSV files are detected by magic bytes and converted to a workbook representation before processing.

• PPT (ppt_parser.py) — Uses python-pptx. Shapes are sorted by position (top then left) before text extraction; supports bulleted lists, tables, and text frames.

Vision Pipeline (PDF Path)#

The vision pipeline is invoked by RAGFlowPdfParser and handles image/scanned content. Its three stages run in sequence 2:

1. OCR — deepdoc/vision/ocr.py#

The OCR class wraps two ONNX models: a text detector (det.onnx) and a text recognizer (rec.onnx). Both are loaded via load_model() which selects CUDAExecutionProvider if torch.cuda.is_available(), otherwise CPUExecutionProvider. Multi-GPU parallelism is enabled via PARALLEL_DEVICES=N 3.

2. Layout Recognition — deepdoc/vision/layout_recognizer.py#

LayoutRecognizer4YOLOv10 (exported as LayoutRecognizer) classifies page regions into 10 label types 4:

• Text, Title, Figure, Figure caption, Table, Table caption, Header, Footer, Reference, Equation

Layout output drives downstream decisions: which blocks are contiguous text, which need TSR, and which are figures with captions. An AscendLayoutRecognizer variant is available for Huawei NPU deployments 5. The backend is selected at runtime via the LAYOUT_RECOGNIZER_TYPE env var (onnx or ascend) 6.

3. Table Structure Recognition (TSR) — deepdoc/vision/table_structure_recognizer.py#

TableStructureRecognizer (a subclass of Recognizer) classifies elements within a cropped table image into 6 labels 7:

• table, table column, table row, table column header, table projected row header, table spanning cell

TSR output is combined with OCR text boxes (which live in page-cumulative Y coordinates) via construct_table(), producing an HTML table or natural-language sentences. The coordinate space mismatch between TSR (image-local) and OCR boxes (cumulative) is resolved in _table_transformer_job(). TSR also supports Ascend via TABLE_STRUCTURE_RECOGNIZER_TYPE=ascend 8.

Table Auto-Rotation#

For scanned PDFs with rotated tables, _ocr_rotated_tables() evaluates 4 rotation angles (0°, 90°, 180°, 270°) using OCR confidence scores to find the best orientation before running TSR. This is on by default and controlled via the TABLE_AUTO_ROTATE env var 9.

PDF Parser Backends#

When used from the RAGFlow pipeline system, the PDF parse method is controlled by parse_method in ParserParam.setups["pdf"] 10. Available backends:

Key Source Files#