zlei6/ragflow
1
0
Fork 0
mirror of https://github.com/infiniflow/ragflow.git synced 2026-08-15 05:04:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(mistral): honor dataset language in figure prompts (#18021)

Browse Source 
### Summary

Refs #17885.

Mistral figure enrichment now receives the dataset language through the
production parsing path. `by_mistral_ocr` forwards `lang` to
`MistralParser.parse_pdf`; the parser stores the normalized language and
passes it to the figure-description prompt. Empty or missing values
still fall back to English.
This commit is contained in:
Ziyang Guo
2026-08-11 20:38:24 +08:00
committed by GitHub
parent ad6fdfd7b4
commit 75363af54b
4 changed files with 38 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
deepdoc/parser/mistral_parser.py
View File

@@ -78,6 +78,7 @@ class MistralParser(RAGFlowPdfParser):
# Optional tenant vision model (LLMBundle) for figure description;
# set from parse_pdf's vision_model kwarg. None -> no enrichment.
self.vision_model = None
self.language = "English"
# ------------------------------------------------------------------
# Page image rendering
@@ -326,7 +327,8 @@ class MistralParser(RAGFlowPdfParser):
# vision_llm_chunk expects a PIL Image (it calls img.size / img.save),
# not raw bytes — pass the crop directly.
desc = vision_llm_chunk(binary=img, vision_model=self.vision_model, prompt=vision_llm_figure_describe_prompt())
self.logger.debug("[Mistral OCR] describing figure with language=%s", self.language)
desc = vision_llm_chunk(binary=img, vision_model=self.vision_model, prompt=vision_llm_figure_describe_prompt(language=self.language))
return (desc or "").strip()
except Exception:
self.logger.info("[Mistral OCR] figure description skipped", exc_info=True)
@@ -457,6 +459,7 @@ class MistralParser(RAGFlowPdfParser):
def parse_pdf(self, filepath: str | PathLike[str], binary=None, callback=None, parse_method: str = "raw", from_page: int = 0, to_page: int = MAXIMUM_PAGE_NUMBER, **kwargs) -> tuple[list, list]:
# Optional tenant vision model for figure description (best-effort).
self.vision_model = kwargs.pop("vision_model", None)
self.language = kwargs.pop("lang", None) or "English"
# Load bytes.
if binary is not None: