feat: 添加反思；添加多合同支持；

68c08496 · ccran · 9fa3fd96 · 68c08496 · 68c08496 · 68c08496
Commit 68c08496 authored Feb 05, 2026 by ccran
11 changed files
--- a/__pycache__/main.cpython-312.pyc
+++ b/__pycache__/main.cpython-312.pyc
--- a/core/__pycache__/memory.cpython-312.pyc
+++ b/core/__pycache__/memory.cpython-312.pyc
--- a/core/memory.py
+++ b/core/memory.py
@@ -262,8 +262,8 @@ class MemoryStore:
    with self._lock:
      wb = Workbook()
-      ws_findings = wb.active
+      ws_final_findings = wb.active
-      ws_findings.title = "findings"
+      ws_final_findings.title = "findings"
      finding_headers = [
        ("rule_title", "规则标题"),
@@ -273,19 +273,19 @@ class MemoryStore:
        ("risk_level", "风险等级"),
        ("suggestion", "建议"),
      ]
-      ws_findings.append([label for _, label in finding_headers])
-      for f in self.findings:
-        ws_findings.append([
-          getattr(f, key, "") for key, _ in finding_headers
-        ])
-      ws_final_findings = wb.create_sheet("final_findings")
      ws_final_findings.append([label for _, label in finding_headers])
      for f in self.final_findings:
        ws_final_findings.append([
          getattr(f, key, "") for key, _ in finding_headers
        ])
+      ws_findings = wb.create_sheet("findings")
+      ws_findings.append([label for _, label in finding_headers])
+      for f in self.findings:
+        ws_findings.append([
+          getattr(f, key, "") for key, _ in finding_headers
+        ])
      ws_facts = wb.create_sheet("facts")
      if self.facts:
        fact_keys: List[str] = sorted({k for item in self.facts for k in item.keys()})
@@ -321,6 +321,7 @@ class MemoryStore:
    doc_obj: DocBase,
    file_name: Optional[str] = None,
    remove_prefix: bool = False,
+    export_final: bool = False,
  ) -> Dict[str, Any]:
    """Add all findings as comments to a document, upload, then delete the local file."""
    if doc_obj is None:
@@ -334,11 +335,14 @@ class MemoryStore:
      name = f"{name}{suffix}"
    output_path = Path(__file__).resolve().parent.parent / "tmp" / name
+    if export_final:
+      target_findings = self.final_findings
+    else:
+      target_findings = self.findings
    with self._lock:
      comments: List[Dict[str, Any]] = []
-      for idx, f in enumerate(self.findings, start=1):
+      for idx, f in enumerate(target_findings, start=1):
-      # 导出final_findings
-      # for idx, f in enumerate(self.final_findings, start=1):
        segment_id = int(f.segment_id or 0)
        chunk_id = max(segment_id - 1, 0)
        suggest_parts = []

--- a/core/tools/__pycache__/reflect_retry.cpython-312.pyc
+++ b/core/tools/__pycache__/reflect_retry.cpython-312.pyc
--- a/core/tools/__pycache__/segment_llm.cpython-312.pyc
+++ b/core/tools/__pycache__/segment_llm.cpython-312.pyc
--- a/core/tools/__pycache__/segment_review.cpython-312.pyc
+++ b/core/tools/__pycache__/segment_review.cpython-312.pyc
--- a/core/tools/reflect_retry.py
+++ b/core/tools/reflect_retry.py
@@ -8,21 +8,12 @@ from core.tools.segment_llm import LLMTool
 REFLECT_SYSTEM_PROMPT = '''
 你是一个合同审查反思智能体（ReviewReflection）。
+你要基于 facts 与全文上下文，对已有 findings 进行校正后，输出【最终可交付的 findings 列表】。
-你的任务是：基于 facts 与全文上下文，
+要求：
-对已有 findings 输出修改操作。
-你只能对 findings 执行以下四种操作：
- keep：确认该风险结论成立且无需修改
- update：修改一个已有风险
- add：新增一个由全文结构推导出的风险
- remove：删除一个不成立的风险
-【输出约束】
- 输出必须是 JSON
- 每条操作必须仅包含字段：op、id、findings
- findings 在 add / update 时必须是完整 finding
 - 严格按照输出 JSON Schema 返回结果，不得输出任何解释性文字
+- 最终 findings 中每条都必须证据充分，original_text 必须是合同原文直接引用
+- 不得引入新的审查维度，只能基于已有 findings 的范围做合并、修订、删除或系统性总结
 '''
 REFLECT_USER_PROMPT = '''
@@ -36,28 +27,23 @@ REFLECT_USER_PROMPT = '''
 【合同立场】
 站在 {party_role} 的立场进行反思审查。
-【反思原则】
+【任务】
- 不得引入新的审查维度
+输出反思后的最终 findings 列表（可直接用于最终审查报告）：
- 所有判断必须有 facts 或合同原文证据支持
+- 删除在全文上下文中不成立的 findings
- 仅对已有 findings 进行增add、改update、删remove、保留keep操作
+- 修订表述/严重性/建议不准确的 findings
- 若风险在全文上下文中不成立，必须 remove
+- 如需合并重复 findings，请合并成一条（保留全部原文证据引用）
- 若风险成立但表述、严重性或建议不准确，必须 update
+- 如可由全文结构推导出系统性风险，可新增 1~3 条 global findings（仍需原文证据）
- 若由多个 findings 或全文结构推导出新的系统性风险，可 add
-【输出要求】
- 输出必须是 JSON
- 每条操作必须仅包含字段：op、id、findings
- findings 在 add / update 时必须是完整 finding
 - 严格按照输出 JSON Schema 返回结果，不得输出任何解释性文字
 '''
 OUTPUT_FORMAT_SCHEMA = '''
 ```json
 {
-  "operations": [
+    "final_findings": [
        {
-      "op": "keep | add | update | remove",
+            "segment_id":"合同原文片段所在的段落ID",
-      "id": "string | null",
+            "issue": "详细的风险描述",
-      "findings": "object | null"
+            "original_text": "合同原文片段的直接引用",
+            "suggestion": "可直接替换原文或新增的条款措辞"
        }
    ]
 }
@@ -98,9 +84,13 @@ class ReflectRetryTool(LLMTool):
    except Exception:
      data = {}
-    operations = data.get("operations", []) or []
+    final_findings = data.get("final_findings", []) or []
-    final_findings = self._apply_operations(base_findings, operations)
+    for finding in final_findings:
+        try:
+          finding['segment_id'] = int(finding.get('segment_id', 0))
+        except Exception:
+          finding['segment_id'] = 0
+        finding['rule_title'] = rule.get('title','')
    return final_findings
  def _build_findings_with_ids(self, findings: List[Dict]) -> List[Dict[str, Any]]:
@@ -112,33 +102,6 @@ class ReflectRetryTool(LLMTool):
      res.append(item)
    return res
-  def _apply_operations(self, base_findings: List[Dict[str, Any]], operations: List[Dict]) -> List[Dict[str, Any]]:
-    by_id = {str(f.get("id")): dict(f) for f in base_findings if f.get("id") is not None}
-    added: List[Dict[str, Any]] = []
-    for op in operations:
-      action = (op.get("op") or "").strip().lower()
-      target_id = op.get("id")
-      payload = op.get("findings")
-      if action == "keep":
-        if target_id is not None and str(target_id) in by_id:
-          continue
-      elif action == "remove":
-        if target_id is not None:
-          by_id.pop(str(target_id), None)
-      elif action == "update":
-        if target_id is not None and isinstance(payload, dict):
-          payload = dict(payload)
-          payload["id"] = str(target_id)
-          by_id[str(target_id)] = payload
-      elif action == "add":
-        if isinstance(payload, dict):
-          added.append(dict(payload))
-    merged = list(by_id.values()) + added
-    return merged
 if __name__ == "__main__":
  tool = ReflectRetryTool()

--- a/core/tools/segment_review.py
+++ b/core/tools/segment_review.py
@@ -10,6 +10,9 @@ from utils.excel_util import ExcelUtil
 from core.tools.segment_llm import LLMTool
 import re
+DEFAULT_RULESET_ID = "通用"
+ALL_RULESET_IDS = ["通用","借款","担保","测试"]
 REVIEW_SYSTEM_PROMPT = '''
 你是一个专业的合同分段审查智能体（SegmentReview）。
 你的核心任务是对“当前分段”进行【法律风险识别】，并给出可落地的修改建议。
@@ -103,7 +106,7 @@ def _is_generic_suggestion(text: str) -> bool:
 class SegmentReviewTool(LLMTool):
    def __init__(self):
        super().__init__(REVIEW_SYSTEM_PROMPT)
-        self.rule_version = "通用"
+        self.default_ruleset_id = DEFAULT_RULESET_ID
        self.column_map = {
            "id": "ID",
            "title": "审查项",
@@ -113,8 +116,10 @@ class SegmentReviewTool(LLMTool):
            "suggestion_template": "建议模板",
        }
        rules_path = Path(__file__).resolve().parent.parent.parent / "data" / "rules.xlsx"
-        rules = ExcelUtil.load_mapped_excel(rules_path, sheet_name=self.rule_version, column_map=self.column_map)
+        self.rulesets: Dict[str, List[Dict]] = {}
-        self.rulesets: Dict[str, List[Dict]] = {self.rule_version: rules} if rules else {}
+        for rs_id in ALL_RULESET_IDS:
+            rules = ExcelUtil.load_mapped_excel(rules_path, sheet_name=rs_id, column_map=self.column_map)
+            self.rulesets[rs_id] = rules
    @tool_func(
        {
@@ -130,7 +135,7 @@ class SegmentReviewTool(LLMTool):
        }
    )
    def run(self, segment_id: str, segment_text: str, ruleset_id: str, party_role: str, context_memories: Optional[List[Dict]] = None) -> Dict:
-        rules = self.rulesets.get(ruleset_id) or self.rulesets.get(self.rule_version, []) or []
+        rules = self.rulesets.get(ruleset_id) or self.rulesets.get(self.default_ruleset_id, []) or []
        result = self._evaluate_rules(party_role,segment_id,segment_text,rules, context_memories)
        overall = "revise" if (result["findings"] ) else "pass"

--- a/data/rules.xlsx
+++ b/data/rules.xlsx
--- a/main.py
+++ b/main.py
@@ -67,7 +67,7 @@ async def parse_document(payload: DocumentParseRequest) -> DocumentParseResponse
 	# text = doc_obj.get_all_text()
 	chunk_ids = doc_obj.get_chunk_id_list()
 	# get ruleset items
-	ruleset_id = payload.ruleset_id or review_tool.rule_version
+	ruleset_id = payload.ruleset_id or review_tool.default_ruleset_id
 	ruleset_items = review_tool.rulesets.get(ruleset_id) or []
 	ruleset_review_items = [r.get('title') for r in ruleset_items]
 	return DocumentParseResponse(
@@ -198,15 +198,13 @@ class ReflectReviewResponse(BaseModel):
 @app.post("/segments/review/reflect", response_model=ReflectReviewResponse)
 def reflect_review(payload: ReflectReviewRequest) -> ReflectReviewResponse:
 	store = get_cached_memory(payload.conversation_id)
-    ruleset_id = payload.ruleset_id or review_tool.rule_version
+	ruleset_id = payload.ruleset_id or review_tool.default_ruleset_id
 	ruleset_items = review_tool.rulesets.get(ruleset_id) or []
 	rule = next((r for r in ruleset_items if r.get("title") == payload.rule_title), None)
 	if not rule:
 		raise HTTPException(status_code=404, detail=f"Rule not found: {payload.rule_title}")
 	facts = store.get_facts()
-    findings = [f.__dict__ for f in store.list_findings()]
+	findings = [f.__dict__ for f in store.search_findings("", rule_title=payload.rule_title)]
 	final_findings = reflect_tool.run(
 		party_role=payload.party_role,
 		rule=rule,
@@ -221,12 +219,12 @@ def reflect_review(payload: ReflectReviewRequest) -> ReflectReviewResponse:
 				"segment_id": f.get("segment_id", 0),
 				"original_text": f.get("original_text", ""),
 				"issue": f.get("issue", ""),
-                "risk_level": (f.get("risk_level") or f.get("level") or "").upper(),
+				"risk_level": (f.get("risk_level") or f.get("level") or "H").upper(),
 				"suggestion": f.get("suggestion", ""),
 			})
+			# print(f'len(store) final_findings:{len(store.final_findings)}')
 		except Exception:
 			continue
 	return ReflectReviewResponse(
 		conversation_id=payload.conversation_id,
 		rule_title=payload.rule_title,
@@ -250,6 +248,7 @@ class MemoryExportRequest(BaseModel):
 	conversation_id: str
 	file_ext: str
 	file_name: Optional[str] = None
+	export_final: Optional[bool] = False
 class MemoryExportResponse(BaseModel):
@@ -273,7 +272,7 @@ def export_memory(payload: MemoryExportRequest) -> MemoryExportResponse:
 		raise HTTPException(status_code=500, detail=f"Export failed: {exc}")
 	try:
-		doc_res = store.export_findings_to_doc_comments(doc_obj)
+		doc_res = store.export_findings_to_doc_comments(doc_obj,export_final=payload.export_final or False)
 	except Exception as exc:
 		traceback.print_exc()
 		raise HTTPException(status_code=500, detail=f"Export doc comments failed: {exc}")

--- a/utils/__pycache__/common_util.cpython-312.pyc
+++ b/utils/__pycache__/common_util.cpython-312.pyc