Figure Description Language Propagation#

The Bug#

RAGFlow datasets have a configurable language setting that flows through the chunking pipeline — including the vision-LLM prompts used to describe embedded figures. Two parsers, MinerU and Mistral, bypass the standard VisionFigureParser machinery and call vision_llm_figure_describe_prompt() directly without a language argument. Because the prompt function accepts no language parameter and renders a static template 1, all figure descriptions are generated in English regardless of the dataset's language setting.

Additionally, the four wrapper functions in figure_parser.py that do route through VisionFigureParser share the same underlying gap: VisionFigureParser.__init__ never stores self.language, and the prompt calls inside VisionFigureParser.__call__ also omit the language argument 2.

Architecture: The Two Paths#

Standard path — VisionFigureParser#

deepdoc/parser/figure_parser.py defines four public wrapper functions that orchestrate figure description for the deepdoc/rag-app layer:

Each wrapper constructs a VisionFigureParser instance and calls it. VisionFigureParser.__call__ selects between vision_llm_figure_describe_prompt() (no context) or vision_llm_figure_describe_prompt_with_context(...) (with surrounding text), then runs picture_vision_llm_chunk() per figure 3.

Bypass path — MinerU and Mistral#

MinerU: _enhance_images_with_vlm() calls vision_llm_figure_describe_prompt() at line 897 with no arguments, then passes the prompt string to vision_llm_chunk() for each image. No language is extracted from kwargs — even though parse_pdf() does receive lang via parser_config or kwargs 4.

Mistral: _describe_image() calls vision_llm_figure_describe_prompt() at line 329 with no arguments. parse_pdf() never captures a language/lang kwarg to forward 5.

Root Cause Chain#

The chunk() functions in rag/app/ declare lang as an explicit named parameter — so it is not captured in **kwargs. When the wrappers receive only **kwargs, lang is already absent 6. For MinerU and Mistral, the gap is simpler: neither parser's image-description method receives or forwards a language value.

Dataset language setting
  → task_executor.py → chunker (lang= parameter)
  → rag/app/*.py chunk() — lang declared as named param, absent from **kwargs
  → figure_parser.py wrappers receive **kwargs only (lang missing)
  → VisionFigureParser (no self.language)
  → vision_llm_figure_describe_prompt() ← no language parameter
  → prompt template rendered without {{ language }}
  → VLM produces English output regardless of dataset language

MinerU / Mistral bypass this chain entirely:
  parse_pdf() → _enhance_images_with_vlm() / _describe_image()
             → vision_llm_figure_describe_prompt() ← same static call

Affected Components#

Fix Status (as of 2026-08-05)#

Two open PRs address the figure_parser.py wrappers, but neither fixes MinerU or Mistral:

• PR #17677 — Superseding fix covering all four wrappers. Adds lang="English" to each wrapper signature, stores self.language in VisionFigureParser.__init__, passes language=self.language to prompt calls in __call__, and adds a language parameter to both prompt functions in generator.py. Adds ## OUTPUT LANGUAGE with {{ language }} to both prompt templates. Includes 15 unit tests.

• PR #17227 — Earlier partial fix for the _naive wrapper; superseded by #17677 for figure_parser.py. Also covers rag/flow/parser/ agent-flow paths not addressed by #17677.

How to complete the fix for MinerU / Mistral#

For MinerU: forward the lang already extracted at line 938 into _enhance_images_with_vlm(), then pass language=lang to vision_llm_figure_describe_prompt(language=lang) inside the worker.

For Mistral: pop language from kwargs in parse_pdf() (alongside vision_model at line 459), store it as self.language, and pass it to vision_llm_figure_describe_prompt(language=self.language) in _describe_image().

Both fixes depend on vision_llm_figure_describe_prompt() accepting a language parameter — which PR #17677 adds to generator.py.