YogeLiu/doc_processer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: optimize prompt

Browse Source
This commit is contained in:
yoge
2026-03-10 21:36:35 +08:00
parent d98fa7237c
commit a9d3a35dd7
1 changed files with 7 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
app/services/ocr_service.py
View File

@@ -150,9 +150,7 @@ def _clean_latex_syntax_spaces(expr: str) -> str:
# Strategy: remove spaces before \ and between non-command chars, # Strategy: remove spaces before \ and between non-command chars,
# but preserve the space after \command when followed by a non-\ char # but preserve the space after \command when followed by a non-\ char
cleaned = re.sub(r"\s+(?=\\)", "", content) # remove space before \cmd cleaned = re.sub(r"\s+(?=\\)", "", content) # remove space before \cmd
cleaned = re.sub( cleaned = re.sub(r"(?<!\\)(?<![a-zA-Z])\s+", "", cleaned) # remove space after non-letter non-\
r"(?<!\\)(?<![a-zA-Z])\s+", "", cleaned
) # remove space after non-letter non-\
return f"{operator}{{{cleaned}}}" return f"{operator}{{{cleaned}}}"
# Match _{ ... } or ^{ ... } # Match _{ ... } or ^{ ... }
@@ -630,9 +628,7 @@ class MineruOCRService(OCRServiceBase):
self.glm_ocr_url = glm_ocr_url self.glm_ocr_url = glm_ocr_url
self.openai_client = OpenAI(api_key="EMPTY", base_url=glm_ocr_url, timeout=3600) self.openai_client = OpenAI(api_key="EMPTY", base_url=glm_ocr_url, timeout=3600)
def _recognize_formula_with_paddleocr_vl( def _recognize_formula_with_paddleocr_vl(self, image: np.ndarray, prompt: str = "Formula Recognition:") -> str:
self, image: np.ndarray, prompt: str = "Formula Recognition:"
) -> str:
"""Recognize formula using PaddleOCR-VL API. """Recognize formula using PaddleOCR-VL API.
Args: Args:
@@ -673,9 +669,7 @@ class MineruOCRService(OCRServiceBase):
except Exception as e: except Exception as e:
raise RuntimeError(f"PaddleOCR-VL formula recognition failed: {e}") from e raise RuntimeError(f"PaddleOCR-VL formula recognition failed: {e}") from e
def _extract_and_recognize_formulas( def _extract_and_recognize_formulas(self, markdown_content: str, original_image: np.ndarray) -> str:
self, markdown_content: str, original_image: np.ndarray
) -> str:
"""Extract image references from markdown and recognize formulas. """Extract image references from markdown and recognize formulas.
Args: Args:
@@ -757,9 +751,7 @@ class MineruOCRService(OCRServiceBase):
markdown_content = result["results"]["image"].get("md_content", "") markdown_content = result["results"]["image"].get("md_content", "")
if "![](images/" in markdown_content: if "![](images/" in markdown_content:
markdown_content = self._extract_and_recognize_formulas( markdown_content = self._extract_and_recognize_formulas(markdown_content, original_image)
markdown_content, original_image
)
# Apply postprocessing to fix OCR errors # Apply postprocessing to fix OCR errors
markdown_content = _postprocess_markdown(markdown_content) markdown_content = _postprocess_markdown(markdown_content)
@@ -789,15 +781,11 @@ class MineruOCRService(OCRServiceBase):
# Task-specific prompts (from GLM-OCR SDK config.yaml) # Task-specific prompts (from GLM-OCR SDK config.yaml)
_TASK_PROMPTS: dict[str, str] = { _TASK_PROMPTS: dict[str, str] = {
"text": "Text Recognition:", "text": "Text Recognition. If the content is a formula, please ouput latex code, else output text",
"formula": "Formula Recognition:", "formula": "Formula Recognition:",
"table": "Table Recognition:", "table": "Table Recognition:",
} }
_DEFAULT_PROMPT = ( _DEFAULT_PROMPT = "Text Recognition. If the content is a formula, please ouput latex code, else output text"
"Recognize the text in the image and output in Markdown format. "
"Preserve the original layout (headings/paragraphs/tables/formulas). "
"Do not fabricate content that does not exist in the image."
)
class GLMOCREndToEndService(OCRServiceBase): class GLMOCREndToEndService(OCRServiceBase):
@@ -921,10 +909,7 @@ class GLMOCREndToEndService(OCRServiceBase):
# Parallel OCR calls # Parallel OCR calls
raw_results: dict[int, str] = {} raw_results: dict[int, str] = {}
with ThreadPoolExecutor(max_workers=min(self.max_workers, len(tasks))) as ex: with ThreadPoolExecutor(max_workers=min(self.max_workers, len(tasks))) as ex:
future_map = { future_map = {ex.submit(self._call_vllm, cropped, prompt): idx for idx, region, cropped, prompt in tasks}
ex.submit(self._call_vllm, cropped, prompt): idx
for idx, region, cropped, prompt in tasks
}
for future in as_completed(future_map): for future in as_completed(future_map):
idx = future_map[future] idx = future_map[future]
try: try: