feat: add party role

.gitignore

@@ -7,12 +7,13 @@
 # Keep Python source files
 !**/*.py
 
-!**/*.doc
-!**/*.docx
-!**/*.xlsx
-!**/*.pdf
-!**/*.xls
+# !**/*.doc
+# !**/*.docx
+# !**/*.xlsx
+# !**/*.pdf
+# !**/*.xls
 !workflow/**
+!data/**.xlsx
 
 !README.md
 

core/config.py

@@ -12,16 +12,27 @@ use_docker = False
 #     model: str = "deepseek-v4-pro"
 
 
+# @dataclass
+# class LLMConfig:
+#     base_url: str = "http://172.21.107.80:9002/v1"
+#     api_key: str = "none"
+#     model: str = "Qwen2-72B-Instruct"
+
 @dataclass
 class LLMConfig:
     base_url: str = "http://192.168.252.71:9002/v1"
     api_key: str = "none"
     model: str = "Qwen2-72B-Instruct"
 
+# 最大分片数量
+min_single_chunk_size = 2000
+max_single_chunk_size = 100000
+max_chunk_page = 10
 
-# MAX_SINGLE_CHUNK_SIZE=100000
-MERGE_RULE_PROMPT = False
+# MAX_SINGLE_CHUNK_SIZE = 100000
 MAX_SINGLE_CHUNK_SIZE = 5000
+# MAX_SINGLE_CHUNK_SIZE = 2000
+MERGE_RULE_PROMPT = False
 META_KEY = "META"
 DEFAULT_RULESET_ID = "通用"
 ## 规则集ID列表，需与rules.xlsx中的sheet名称保持一致!!!
@@ -33,7 +44,7 @@ ALL_RULESET_IDS = [
     "金盘",
     "金盘简化",
     "麓发测试",
-    "麓发标准",
+    "通用（原始）",
     "金盘B类",
     "合同信息提取（合同组）",
     "合同信息提取（技术部）",
@@ -45,11 +56,13 @@ MAX_WORKERS = 10
 FILE_SUFFIX = "-审核批注"
 
 ## 关键参数**
-use_non_fastgpt_llm = False
+use_non_fastgpt_llm = True
 use_lufa = False
 use_jp_machine = True
+debug_mode = False
 
 ## 关键参数**
+max_model_len = 131072
 ocr_url = 'http://192.168.252.71:8202/openapi/ocrUploadFile'
 if use_lufa:
     outer_backend_url = "http://znkf.lgfzgroup.com:48081"
@@ -67,9 +80,9 @@ else:
         base_fastgpt_url = "http://192.168.252.71:18088"
         base_backend_url = "http://192.168.252.71:48080"
     else:
-        outer_backend_url = "http://172.21.107.45:48080"
+        outer_backend_url = "https://172.21.107.45:48080"
         base_fastgpt_url = "http://172.21.107.45:3030"
-        base_backend_url = "http://172.21.107.45:48080"
+        base_backend_url = "http://172.21.107.45:1122"
         ocr_url = "http://172.21.107.45:8202/openapi/ocrUploadFile"
     segment_review_api_key = (
         "fastgpt-vLu2JHAfqwEq5FUQhvATFDK0yDS6fs804v7KwWBMyU4sRrHzh4UGl89Zpa"
@@ -108,7 +121,3 @@ LLM = {
 }
 doc_support_formats = [".docx", ".doc", ".wps"]
 pdf_support_formats = [".txt", ".md", ".pdf"]
-# 最大分片数量
-min_single_chunk_size = 2000
-max_single_chunk_size = 20000
-max_chunk_page = 10

core/tools/party_role.py

+from __future__ import annotations
+
+from typing import Dict
+
+from loguru import logger
+
+from core.tool import tool, tool_func
+from core.tools.segment_llm import LLMTool
+
+
+PARTY_ROLE_SYSTEM_PROMPT = """
+你是一名专业合同审查律师。
+"""
+
+
+PARTY_ROLE_USER_PROMPT = """
+请分析指定公司在合同中的实际商业角色。
+
+注意：
+- 不要仅依据“甲方/乙方”字样判断
+- 必须结合实际权利义务关系判断
+
+重点考虑：
+- 谁提出需求
+- 谁支付费用
+- 谁接受服务或交付成果
+- 谁拥有验收、管理、监督权
+- 谁负责提供产品或服务
+
+输出要求：
+
+只输出一个 JSON：
+
+{{
+  "party_role": "demand_side | supplier_side | unclear",
+  "reason": ""
+}}
+
+其中：
+- demand_side：需求方、委托人、采购方、客户、甲方等
+- supplier_side：供应商、服务方、承包方、实施方、乙方等
+- unclear：无法明确判断
+
+reason 要用一句话简要说明核心依据。
+
+禁止输出任何Markdown、代码块或额外文本。
+
+待分析公司：
+{company_name}
+
+合同内容：
+{contract_text}
+"""
+
+
+VALID_PARTY_ROLES = {"demand_side", "supplier_side", "mixed", "unclear"}
+
+PARTY_ROLES_MAP={
+    "demand_side": "需求方、委托人、采购方、客户、甲方等",
+    "supplier_side": "供应商、服务方、承包方、实施方、乙方等",
+    "unclear": "",
+}
+
+
+@tool("party_role", "识别指定公司在合同中的实际商业角色")
+class PartyRoleTool(LLMTool):
+    def __init__(self) -> None:
+        super().__init__(PARTY_ROLE_SYSTEM_PROMPT)
+
+    @tool_func(
+        {
+            "type": "object",
+            "properties": {
+                "company_name": {"type": "string"},
+                "contract_text": {"type": "string"},
+            },
+            "required": ["company_name", "contract_text"],
+        }
+    )
+    def run(self, company_name: str, contract_text: str) -> Dict[str, str]:
+        company_name = (company_name or "").strip()
+        contract_text = (contract_text or "").strip()
+        if not company_name or not contract_text:
+            return {"party_role": "", "reason": ""}
+
+        user_content = PARTY_ROLE_USER_PROMPT.format(
+            company_name=company_name,
+            contract_text=contract_text,
+        )
+        try:
+            resp = self.run_with_loop(self.chat_async(self.build_messages(user_content)))
+            data = self.parse_first_json(resp)
+            party_role = str(data.get("party_role", "")).strip()
+            reason = str(data.get("reason", "")).strip()
+            if party_role in VALID_PARTY_ROLES:
+                return {"party_role": PARTY_ROLES_MAP.get(party_role, ""), "reason": reason}
+        except Exception as exc:
+            logger.error("Party role detection LLM failed: %s", exc)
+
+        return {"party_role": "", "reason": ""}
+
+
+if __name__ == "__main__":
+    tool = PartyRoleTool()
+    print(
+        tool.run(
+            company_name="麓发公司",
+            contract_text="麓发委托乙公司开发系统，甲公司支付费用并负责验收。",
+        )
+    )

demo/demo.py

-for _ in range(input()):
-    try:
-        eval(raw_input())
-        print("YES")
-    except TypeError:
-        print("NO")
-    except:
-        print("NO")

main.py

@@ -19,6 +19,7 @@ from core.config import (
     pdf_support_formats,
     MERGE_RULE_PROMPT,
     use_lufa,
+    max_model_len
 )
 from core.tools.segment_summary import SegmentSummaryTool
 from core.tools.segment_review import SegmentReviewTool
@@ -29,6 +30,7 @@ from core.tools.reflect_retry import ReflectRetryTool
 from core.tools.segment_merger import SegmentMergerTool
 from core.tools.fact_merger import FactMergerTool
 from core.tools.ruleset_router import RulesetRouterTool
+from core.tools.party_role import PartyRoleTool
 from core.memory import Finding
 from core.memory import FINDING_KEY_MERGE, FINDING_KEY_REFLECT, FINDING_KEY_REVIEW
 
@@ -44,6 +46,7 @@ reflect_tool = ReflectRetryTool()
 merger_tool = SegmentMergerTool()
 fact_merger_tool = FactMergerTool()
 ruleset_router_tool = RulesetRouterTool()
+party_role_tool = PartyRoleTool()
 
 
 @app.post("/sleep")
@@ -84,6 +87,17 @@ class RulesetRouteResponse(BaseModel):
     reason: str = ""
 
 
+class PartyRoleRequest(BaseModel):
+    company_name: str = Field(..., description="待分析公司名称")
+    contract_text: str = Field(..., description="合同全文或合同片段内容")
+
+
+class PartyRoleResponse(BaseModel):
+    company_name: str
+    party_role: str
+    reason: Optional[str] = ""
+
+
 @app.post("/rulesets/route", response_model=RulesetRouteResponse)
 def route_ruleset(payload: RulesetRouteRequest) -> RulesetRouteResponse:
     question = (payload.question or "").strip()
@@ -98,6 +112,26 @@ def route_ruleset(payload: RulesetRouteRequest) -> RulesetRouteResponse:
     )
 
 
+@app.post("/contracts/party-role", response_model=PartyRoleResponse)
+def detect_party_role(payload: PartyRoleRequest) -> PartyRoleResponse:
+    company_name = (payload.company_name or "").strip()
+    contract_text = (payload.contract_text or "").strip()
+    if not company_name:
+        raise HTTPException(status_code=400, detail="company_name cannot be empty")
+    if not contract_text:
+        raise HTTPException(status_code=400, detail="contract_text cannot be empty")
+
+    result = party_role_tool.run(
+        company_name=company_name,
+        contract_text=contract_text[:max_model_len],  # TODO 合同文本过长时，优先保留包含公司名称的片段内容，后续优化
+    )
+    return PartyRoleResponse(
+        company_name=company_name,
+        party_role=result.get("party_role", ""),
+        reason=result.get("reason", ""),
+    )
+
+
 @app.post("/documents/parse", response_model=DocumentParseResponse)
 async def parse_document(payload: DocumentParseRequest) -> DocumentParseResponse:
     if not payload.urls:

utils/tesseract_ocr_util.py

+import asyncio
+import os
+import re
+from urllib import parse
+from urllib.parse import urlparse
+
+import fitz
+from loguru import logger
+
+from utils.common_util import random_str
+from utils.http_util import download_file, url_replace_fastgpt
+
+
+class TesseractOCRUtil:
+    def __init__(self, lang='chi_sim+eng'):
+        self.lang = lang
+
+    def ocr_image(self, file_path):
+        import pytesseract
+        from PIL import Image
+
+        with Image.open(file_path) as img:
+            return pytesseract.image_to_string(img, lang=self.lang)
+
+    async def ocr_image_async(self, path_list):
+        tasks = [asyncio.to_thread(self.ocr_image, file_path) for file_path in path_list]
+        responses = await asyncio.gather(*tasks)
+
+        res_dict = {}
+        for file_path, content in zip(path_list, responses):
+            page_num = int(self.get_pdf_2_img_page_num(file_path))
+            res_dict[page_num] = content
+
+        logger.info(f'tesseract ocr_image_async finish. all pages:{len(res_dict)}')
+        return [res_dict[key] for key in sorted(res_dict)]
+
+    def set_pdf_2_img_page(self, path, page_idx):
+        return f'{path}_{page_idx + 1}.png'
+
+    def get_pdf_2_img_page_num(self, path):
+        match = re.search(r'_(\d+)\.png$', path)
+        if not match:
+            raise ValueError(f'Invalid pdf page image path: {path}')
+        return match.group(1)
+
+    def pdf_2_img(self, path, zoom_x=2, zoom_y=2):
+        pdf = fitz.open(path)
+        pdf_list = []
+        for pg in range(0, pdf.page_count):
+            page = pdf[pg]
+            trans = fitz.Matrix(zoom_x, zoom_y)
+            pm = page.get_pixmap(matrix=trans, alpha=False)
+            dest_png = self.set_pdf_2_img_page(path, pg)
+            pm.save(dest_png)
+            pdf_list.append(dest_png)
+        pdf.close()
+        return pdf_list
+
+    def ocr_download_path(self, url):
+        logger.info(f'tesseract ocr url:{url}')
+        url = url_replace_fastgpt(url)
+        url_parsed = urlparse(url)
+        query_dict = parse.parse_qs(url_parsed.query)
+        if 'filename' in query_dict:
+            filename = query_dict.get('filename')[0]
+        else:
+            filename = f'{random_str()}.pdf'
+        dest_path = f'ocr/{filename}'
+        download_file(url, dest_path)
+        return dest_path
+
+    async def ocr_result_pdf(self, dest_path):
+        pdf_list = self.pdf_2_img(dest_path)
+        try:
+            return await self.ocr_image_async(pdf_list)
+        finally:
+            for pdf in pdf_list:
+                if os.path.exists(pdf):
+                    os.remove(pdf)
+
+
+if __name__ == '__main__':
+    ocr_util = TesseractOCRUtil()
+    result = asyncio.run(ocr_util.ocr_result_pdf('../demo/2020100593中建大成建筑（B类）.pdf'))
+    print(f'len(result):{len(result)}')
+    print(result)