Parser Configuration#

The Parser component in RAGFlow's DSL chunker pipeline parses raw files into structured sections before downstream chunking. Its behavior is controlled by a ParserParam object whose defaults can be selectively overridden by the pipeline DSL configuration via param.update().

Key Files#

ParserParam Structure#

ParserParam.__init__() defines two top-level dicts 1:

• setups — per-file-type config dictionaries (one key per supported type: pdf, spreadsheet, doc, docx, markdown, text&code, html, slides, image, email, audio, video, epub)
• allowed_output_format — per-type lists of valid output_format values used exclusively during validation

Each entry in setups contains fields like:

DSL Override via param.update()#

When the canvas loads a pipeline, it runs this sequence for every component 2:

param = ParserParam() # creates defaults
param.update(dsl["params"]) # merges DSL overrides
param.check() # validates constraints
Parser(canvas, id, param) # instantiates component

ComponentParamBase.update() performs a recursive merge 3: it walks the DSL dict and, for each key, either sets the attribute directly (builtins and None) or recurses into the sub-object. Because setups is a plain dict (a builtin type), the entire setups value from the DSL replaces the default setups rather than merging individual file-type entries. A file type omitted from the DSL's setups will therefore lose its defaults entirely — include all file-type configs you need.

A minimal DSL override enabling a VLM method for PDF:

{
  "setups": {
    "pdf": {
      "parse_method": "gpt-4o@openai",
      "lang": "English",
      "output_format": "json"
    }
  }
}

See general_pdf_all.json for a complete multi-type example.

Validation Constraints (param.check())#

ParserParam.check() validates each present file-type config 4. The main constraints are:

PDF (setups["pdf"])

• parse_method must not be empty 5
• If parse_method is not one of deepdoc, plain_text, mineru, docling, opendataloader, tcadp parser, paddleocr, somark, it is treated as a VLM model name and lang must be non-empty 6
• output_format must be in ["json", "markdown"] 7

Spreadsheet — output_format ∈ ["json", "markdown", "html"] 8

Doc / DOCX — output_format ∈ ["json", "markdown"] 9

Slides — output_format must be "json" 10

Image — if parse_method is not "ocr", lang must be non-empty 11

Audio / Video — vlm.llm_id must be non-empty 12

Email — output_format ∈ ["text", "json"] 13

EPUB — output_format ∈ ["text", "json"] 14

Validation helpers (check_empty, check_valid_value) are inherited from ComponentParamBase 15 and raise ValueError on failure; Canvas.load() wraps this in a ValueError with the component name for clear error attribution 16.

Parse Method Selection (PDF)#

The PDF handler normalises parse_method at runtime 17:

Suffix-based routing in _invoke() matches a file's extension against each setups[type]["suffix"] list and dispatches to the first matching handler 18.