611 lines
27 KiB
Python
611 lines
27 KiB
Python
|
# 在 FeatureRetriever 类内部新增以下常量和方法
|
|||
|
|
import json
|
|||
|
|
import os
|
|||
|
|
import re
|
|||
|
|
from typing import Dict, List, Optional
|
|||
|
|
import logging
|
|||
|
|
|
|||
|
|
import faiss
|
|||
|
|
import numpy as np
|
|||
|
|
|
|||
|
|
from config import TOP_K, METADATA_PATH, VECTOR_DB_PATH, KNOWLEDGE_GRAPH_PATH, MIN_SIMILARITY_THRESHOLD
|
|||
|
|
from llm_processor import call_qwen_max, get_qwen_embedding
|
|||
|
|
|
|||
|
|
logger = logging.getLogger(__name__)
|
|||
|
|
class KnowledgeGraphAnalyzer:
|
|||
|
|
"""知识图谱分析器（简化版）"""
|
|||
|
|
|
|||
|
|
def __init__(self, graph_data: Dict):
|
|||
|
|
"""初始化图谱分析器"""
|
|||
|
|
self.nodes = {node["id"]: node for node in graph_data.get("nodes", [])}
|
|||
|
|
self.edges = graph_data.get("edges", [])
|
|||
|
|
|
|||
|
|
# 构建邻接表
|
|||
|
|
self.calls_graph = self._build_calls_graph()
|
|||
|
|
self.reverse_calls_graph = self._build_reverse_calls_graph()
|
|||
|
|
|
|||
|
|
# 构建名称到ID的映射
|
|||
|
|
self.name_to_id = self._build_name_mapping()
|
|||
|
|
|
|||
|
|
def _build_calls_graph(self) -> Dict[str, List[str]]:
|
|||
|
|
"""构建函数调用图"""
|
|||
|
|
calls_graph = {}
|
|||
|
|
for edge in self.edges:
|
|||
|
|
if edge["type"] == "CALLS":
|
|||
|
|
source_id = edge["source_id"]
|
|||
|
|
target_id = edge["target_id"]
|
|||
|
|
if source_id not in calls_graph:
|
|||
|
|
calls_graph[source_id] = []
|
|||
|
|
calls_graph[source_id].append(target_id)
|
|||
|
|
return calls_graph
|
|||
|
|
|
|||
|
|
def _build_reverse_calls_graph(self) -> Dict[str, List[str]]:
|
|||
|
|
"""构建反向调用图"""
|
|||
|
|
reverse_graph = {}
|
|||
|
|
for edge in self.edges:
|
|||
|
|
if edge["type"] == "CALLS":
|
|||
|
|
source_id = edge["source_id"]
|
|||
|
|
target_id = edge["target_id"]
|
|||
|
|
if target_id not in reverse_graph:
|
|||
|
|
reverse_graph[target_id] = []
|
|||
|
|
reverse_graph[target_id].append(source_id)
|
|||
|
|
return reverse_graph
|
|||
|
|
|
|||
|
|
def _build_name_mapping(self) -> Dict[str, str]:
|
|||
|
|
"""构建函数名到节点ID的映射"""
|
|||
|
|
name_to_id = {}
|
|||
|
|
for node_id, node in self.nodes.items():
|
|||
|
|
if node["type"] == "Function":
|
|||
|
|
name_to_id[node["name"]] = node_id
|
|||
|
|
return name_to_id
|
|||
|
|
|
|||
|
|
def find_function_by_name(self, function_name: str) -> Optional[Dict]:
|
|||
|
|
"""根据函数名查找节点"""
|
|||
|
|
node_id = self.name_to_id.get(function_name)
|
|||
|
|
if node_id:
|
|||
|
|
return self.nodes.get(node_id)
|
|||
|
|
return None
|
|||
|
|
|
|||
|
|
def analyze_function_context(self, function_id: str) -> Dict:
|
|||
|
|
"""分析函数的上下文（调用者和被调用者）"""
|
|||
|
|
if function_id not in self.nodes:
|
|||
|
|
return {"error": f"未找到函数节点: {function_id}"}
|
|||
|
|
|
|||
|
|
result = {
|
|||
|
|
"function_info": self.nodes[function_id],
|
|||
|
|
"callers": [],
|
|||
|
|
"callees": [],
|
|||
|
|
"call_depth": 0
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
# 分析调用者
|
|||
|
|
if function_id in self.reverse_calls_graph:
|
|||
|
|
for caller_id in self.reverse_calls_graph[function_id]:
|
|||
|
|
if caller_id in self.nodes:
|
|||
|
|
result["callers"].append(self.nodes[caller_id])
|
|||
|
|
|
|||
|
|
# 分析被调用者
|
|||
|
|
if function_id in self.calls_graph:
|
|||
|
|
for callee_id in self.calls_graph[function_id]:
|
|||
|
|
if callee_id in self.nodes:
|
|||
|
|
result["callees"].append(self.nodes[callee_id])
|
|||
|
|
|
|||
|
|
# 计算调用深度
|
|||
|
|
result["call_depth"] = self._calculate_max_depth(function_id)
|
|||
|
|
|
|||
|
|
return result
|
|||
|
|
|
|||
|
|
def _calculate_max_depth(self, func_id: str, visited: set = None) -> int:
|
|||
|
|
"""计算函数的最大调用深度"""
|
|||
|
|
if visited is None:
|
|||
|
|
visited = set()
|
|||
|
|
|
|||
|
|
if func_id in visited or func_id not in self.calls_graph:
|
|||
|
|
return 0
|
|||
|
|
|
|||
|
|
visited.add(func_id)
|
|||
|
|
|
|||
|
|
max_depth = 0
|
|||
|
|
for callee_id in self.calls_graph.get(func_id, []):
|
|||
|
|
depth = 1 + self._calculate_max_depth(callee_id, visited.copy())
|
|||
|
|
max_depth = max(max_depth, depth)
|
|||
|
|
|
|||
|
|
return max_depth
|
|||
|
|
class FeatureRetriever:
|
|||
|
|
"""功能需求检索器主类（简化版）"""
|
|||
|
|
|
|||
|
|
def __init__(self, config: Optional[Dict] = None):
|
|||
|
|
"""初始化检索器"""
|
|||
|
|
self.llm_judgment_prompt_template = self._get_judgment_prompt_template()
|
|||
|
|
|
|||
|
|
def _get_judgment_prompt_template(self) -> str:
|
|||
|
|
"""返回用于LLM推理的提示词模板"""
|
|||
|
|
return r"""
|
|||
|
|
你是一个资深的软件架构师和代码评审专家。你需要根据用户提出的功能需求，结合给定的代码知识库分析结果，判断该需求在项目中是否已被实现。
|
|||
|
|
|
|||
|
|
# 任务
|
|||
|
|
请严格遵循以下步骤和约束条件进行分析，并给出最终裁决。
|
|||
|
|
|
|||
|
|
# 输入信息
|
|||
|
|
1. **用户需求**： “{user_query}”
|
|||
|
|
2. **候选函数分析档案**：
|
|||
|
|
{analysis_dossiers}
|
|||
|
|
|
|||
|
|
# 你必须遵循的推理约束（知识图谱约束）
|
|||
|
|
请基于以下四个约束条件，对每个候选函数进行分析，并综合判断需求实现状态：
|
|||
|
|
|
|||
|
|
**约束1：功能完整性约束**
|
|||
|
|
- 检查候选函数的调用链（callers & callees）是否能完整实现用户需求所隐含的所有子步骤或业务流程。
|
|||
|
|
- 关键问题：是否存在逻辑断点？调用链是否覆盖了需求中“输入-处理-输出”的全过程？
|
|||
|
|
|
|||
|
|
**约束2：架构/设计模式契合度约束**
|
|||
|
|
- 识别候选函数所体现的设计模式（如工厂、策略、观察者、责任链、模板方法等）。
|
|||
|
|
- 判断该模式是否与用户需求所描述的系统行为模式、扩展性要求或解耦意图相匹配。
|
|||
|
|
|
|||
|
|
**约束3：模块协调性约束**
|
|||
|
|
- 分析候选函数与其相关函数（调用者/被调用者）是否共同构成了一个高内聚、低耦合的功能模块。
|
|||
|
|
- 关键问题：相关函数是否在语义上属于同一业务领域？它们之间的数据流是否清晰合理？
|
|||
|
|
|
|||
|
|
**约束4：语义覆盖深度约束**
|
|||
|
|
- 超越简单的文本相似度，深入分析函数在代码上下文中的**实际语义**。
|
|||
|
|
- 关键问题：函数的名称、摘要、调用关系所体现的核心功能，是否精准地**覆盖**了用户需求的**核心意图**？是否存在语义偏差或功能泛化/缩窄？
|
|||
|
|
|
|||
|
|
你的输出**必须且只能**是一个完全符合 RFC 8259 标准的 JSON 对象。该 JSON 必须能被 Python 的 `json.loads()` 函数**直接解析**，无需任何预处理。
|
|||
|
|
为了确保这一点，你必须严格遵守以下规则：
|
|||
|
|
1. **禁用反引号**：在整个 JSON 对象的任何字符串值中，**绝对不允许**使用反引号 (\`)。所有需要引用标识符（如函数名、文件名）的地方，应直接将其作为字符串的一部分，**无需用任何特殊字符包围**。例如，应写成 `StarCheckGyro`，而非 \`StarCheckGyro\`。
|
|||
|
|
2. **正确转义双引号**：如果在字符串内部**必须**使用双引号（"），则必须使用反斜杠进行转义，写作 `\"`。例如，描述“主函数”的字符串应写作 `"主函数"`，若需在内部提及带引号的文本，应写作 `"他说：\\"你好\\""`。
|
|||
|
|
3. **正确的布尔值和空值**：布尔值必须使用小写的 `true` 或 `false`。空值必须使用小写的 `null`。禁止使用 Python 的 `True`, `False`, `None` 或大写的 `Null`。
|
|||
|
|
4. **无注释**：JSON 对象中不允许包含任何形式的注释（`//`, `/* */`）。
|
|||
|
|
5. **无尾随逗号**：在对象或数组的最后一个元素后，不允许有逗号。
|
|||
|
|
|
|||
|
|
违反以上任何一条规则，都将导致你的输出无法被解析，任务失败。
|
|||
|
|
|
|||
|
|
# 输出格式
|
|||
|
|
你必须以以下JSON格式输出，且仅输出此JSON对象：
|
|||
|
|
{{
|
|||
|
|
"verdict": "完整实现 | 部分实现 | 未实现",
|
|||
|
|
"confidence": 0.0到1.0之间的一个浮点数,
|
|||
|
|
"primary_reasoning": "一段文字，综合阐述依据约束1-4得出的主要判断理由。",
|
|||
|
|
"constraint_analysis": {{
|
|||
|
|
"functional_completeness": {{
|
|||
|
|
"status": "满足 | 部分满足 | 不满足",
|
|||
|
|
"reason": "基于约束1的分析说明。"
|
|||
|
|
}},
|
|||
|
|
"architectural_fit": {{
|
|||
|
|
"status": "匹配 | 部分匹配 | 不匹配",
|
|||
|
|
"reason": "基于约束2的分析说明。"
|
|||
|
|
}},
|
|||
|
|
"module_coordination": {{
|
|||
|
|
"status": "协调 | 部分协调 | 不协调",
|
|||
|
|
"reason": "基于约束3的分析说明。"
|
|||
|
|
}},
|
|||
|
|
"semantic_coverage": {{
|
|||
|
|
"status": "覆盖 | 部分覆盖 | 未覆盖",
|
|||
|
|
"reason": "基于约束4的分析说明。"
|
|||
|
|
}}
|
|||
|
|
}},
|
|||
|
|
"call_chain_analysis": {{
|
|||
|
|
"description": "对实现需求的核心调用链的详细文字描述。包括起点函数、关键处理函数、终点函数，以及数据流或控制流的简要说明。",
|
|||
|
|
"chain": ["函数名1", "函数名2", "...", "函数名N"] # 按调用顺序列出关键函数名
|
|||
|
|
}},
|
|||
|
|
"design_pattern_analysis": {{
|
|||
|
|
"identified_patterns": ["模式1", "模式2", ...], # 识别出的设计模式列表
|
|||
|
|
"explanation": "对识别出的设计模式进行解释，说明其在代码中如何体现，以及为何适用于当前需求。"
|
|||
|
|
}},
|
|||
|
|
"most_relevant_functions": [
|
|||
|
|
{{
|
|||
|
|
"name": "函数名",
|
|||
|
|
"role": "此函数在满足需求中扮演的角色（如：主入口、关键计算、数据提供者、校验器、协调者等）",
|
|||
|
|
"supporting_evidence": "来自分析档案的简要证据"
|
|||
|
|
}}
|
|||
|
|
],
|
|||
|
|
"gap_analysis": "如果 verdict 不是 '完整实现'，请详细说明缺失的功能组件、断裂的调用链或语义偏差。否则为空字符串。"
|
|||
|
|
}}
|
|||
|
|
现在，请开始你的分析。
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
def _generate_analysis_dossier(self, func_info: Dict) -> str:
|
|||
|
|
"""为单个候选函数生成详细的文本分析档案"""
|
|||
|
|
dossier_lines = []
|
|||
|
|
|
|||
|
|
# 1. 基础信息
|
|||
|
|
dossier_lines.append(f"## 函数: {func_info.get('name', 'N/A')}")
|
|||
|
|
dossier_lines.append(f"- **文件**: {os.path.basename(func_info.get('file', 'N/A'))}")
|
|||
|
|
dossier_lines.append(f"- **语义相似度**: {func_info.get('similarity', 0):.3f}")
|
|||
|
|
dossier_lines.append(f"- **功能摘要**: {func_info.get('summary', '无')}")
|
|||
|
|
|
|||
|
|
# 2. 调用链深度分析 (利用知识图谱)
|
|||
|
|
func_name = func_info.get('name')
|
|||
|
|
call_chain_info = " (知识图谱未加载，无法分析)"
|
|||
|
|
if self.kg_analyzer and func_name:
|
|||
|
|
node = self.kg_analyzer.find_function_by_name(func_name)
|
|||
|
|
if node:
|
|||
|
|
ctx = self.kg_analyzer.analyze_function_context(node['id'])
|
|||
|
|
dossier_lines.append(f"- **调用深度**: {ctx.get('call_depth', 0)}")
|
|||
|
|
dossier_lines.append(f"- **被以下函数调用 (Callers, {len(ctx.get('callers', []))}个)**:")
|
|||
|
|
for caller in ctx.get('callers', [])[:3]: # 最多显示3个
|
|||
|
|
dossier_lines.append(f" - {caller.get('name', 'N/A')}")
|
|||
|
|
dossier_lines.append(f"- **调用以下函数 (Callees, {len(ctx.get('callees', []))}个)**:")
|
|||
|
|
for callee in ctx.get('callees', [])[:5]: # 最多显示5个
|
|||
|
|
dossier_lines.append(f" - {callee.get('name', 'N/A')}")
|
|||
|
|
call_chain_info = f" 深度{ctx.get('call_depth', 0)}， {len(ctx.get('callers', []))}个调用者， {len(ctx.get('callees', []))}个被调用者。"
|
|||
|
|
|
|||
|
|
# 3. 架构模式分析 (增强版)
|
|||
|
|
pattern_info = self._analyze_design_pattern(func_info)
|
|||
|
|
dossier_lines.append(f"- **识别的设计模式/架构特征**: {', '.join(pattern_info) if pattern_info else '未识别到明显模式'}")
|
|||
|
|
|
|||
|
|
dossier_lines.append("") # 空行分隔
|
|||
|
|
return "\n".join(dossier_lines)
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _analyze_design_pattern(self, func_info: Dict) -> List[str]:
|
|||
|
|
"""增强的设计模式分析，返回识别出的模式列表"""
|
|||
|
|
patterns = []
|
|||
|
|
func_name = func_info.get('name', '')
|
|||
|
|
|
|||
|
|
if not self.kg_analyzer:
|
|||
|
|
return patterns
|
|||
|
|
|
|||
|
|
# 关键修正：使用您原有的 find_function_by_name 方法
|
|||
|
|
node = self.kg_analyzer.find_function_by_name(func_name)
|
|||
|
|
if not node:
|
|||
|
|
return patterns
|
|||
|
|
|
|||
|
|
# 关键修正：使用您原有的 analyze_function_context 方法获取调用关系
|
|||
|
|
ctx = self.kg_analyzer.analyze_function_context(node['id'])
|
|||
|
|
callers = ctx.get('callers', [])
|
|||
|
|
callees = ctx.get('callees', [])
|
|||
|
|
callers_cnt = len(callers)
|
|||
|
|
callees_cnt = len(callees)
|
|||
|
|
|
|||
|
|
# 基于调用关系推断模式
|
|||
|
|
if callers_cnt == 0 and callees_cnt > 2:
|
|||
|
|
patterns.append("“工厂方法”或“构建器” (可能是对象创建入口)")
|
|||
|
|
elif callers_cnt > 2 and callees_cnt == 0:
|
|||
|
|
patterns.append("“观察者”的通知方法 或 “策略”的接口实现")
|
|||
|
|
elif callers_cnt == 1 and callees_cnt > 1:
|
|||
|
|
patterns.append("“模板方法”中的步骤定义 或 “外观”模式接口")
|
|||
|
|
elif callers_cnt > 1 and callees_cnt > 1:
|
|||
|
|
patterns.append("“协调者”或“中介者” (复杂逻辑协调)")
|
|||
|
|
if ctx.get('call_depth', 0) >= 3:
|
|||
|
|
patterns.append("“责任链”或“管道” (多级处理)")
|
|||
|
|
|
|||
|
|
# 基于函数名关键词的启发式规则 (可选)
|
|||
|
|
name_lower = func_name.lower()
|
|||
|
|
if 'factory' in name_lower or 'create' in name_lower or 'build' in name_lower:
|
|||
|
|
patterns.append("函数名暗示“创建型”模式")
|
|||
|
|
if 'handler' in name_lower or 'processor' in name_lower or 'service' in name_lower:
|
|||
|
|
patterns.append("函数名暗示“行为型”模式")
|
|||
|
|
if 'adapter' in name_lower or 'wrapper' in name_lower:
|
|||
|
|
patterns.append("函数名暗示“结构型”模式")
|
|||
|
|
|
|||
|
|
return list(set(patterns)) # 去重
|
|||
|
|
|
|||
|
|
def analyze_with_multiple_constraints(self, query: str) -> Dict:
|
|||
|
|
"""
|
|||
|
|
基于多推理约束分析需求实现状态 (新版)
|
|||
|
|
流程：1. 语义检索 -> 2. 为每个结果生成分析档案 -> 3. LLM基于约束推理
|
|||
|
|
"""
|
|||
|
|
if not self.faiss_index:
|
|||
|
|
logger.error("知识库未加载")
|
|||
|
|
return {"error": "知识库未完全加载"}
|
|||
|
|
|
|||
|
|
# 阶段1: 语义检索 (不变)
|
|||
|
|
relevant_functions = self.search_by_semantic(query, top_k=TOP_K)
|
|||
|
|
if not relevant_functions:
|
|||
|
|
return {
|
|||
|
|
"implemented": False,
|
|||
|
|
"reason": "在代码库中未找到语义相似度足够高的相关函数",
|
|||
|
|
"confidence": 0.1,
|
|||
|
|
"analysis_method": "semantic_search_only",
|
|||
|
|
"details": "语义检索未命中。"
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
# 阶段2: 为每个相关函数生成深度分析档案
|
|||
|
|
analysis_dossiers = []
|
|||
|
|
for func_info in relevant_functions:
|
|||
|
|
dossier_text = self._generate_analysis_dossier(func_info)
|
|||
|
|
analysis_dossiers.append(dossier_text)
|
|||
|
|
|
|||
|
|
combined_dossiers = "\n---\n".join(analysis_dossiers)
|
|||
|
|
|
|||
|
|
# 阶段3: 调用LLM，基于约束进行综合推理判断
|
|||
|
|
llm_judgment_result = self._make_llm_judgment(query, combined_dossiers)
|
|||
|
|
|
|||
|
|
# 阶段4: 格式化最终输出，兼容新旧接口
|
|||
|
|
final_result = self._format_final_result(llm_judgment_result, relevant_functions)
|
|||
|
|
return final_result
|
|||
|
|
|
|||
|
|
def _make_llm_judgment(self, query: str, analysis_dossiers: str) -> Dict:
|
|||
|
|
"""核心：调用LLM，强制其在知识图谱约束下进行推理评判"""
|
|||
|
|
try:
|
|||
|
|
# 构建提示词
|
|||
|
|
prompt = self.llm_judgment_prompt_template.format(
|
|||
|
|
user_query=query,
|
|||
|
|
analysis_dossiers=analysis_dossiers
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# 调用LLM
|
|||
|
|
llm_response_text = call_qwen_max(prompt, temperature=0.1)
|
|||
|
|
|
|||
|
|
# --- 新增：健壮的JSON解析与修复 ---
|
|||
|
|
import json
|
|||
|
|
judgment_data = self._robust_json_parse(llm_response_text)
|
|||
|
|
if judgment_data is not None:
|
|||
|
|
return judgment_data
|
|||
|
|
else:
|
|||
|
|
# 如果修复后解析仍然失败，则回退
|
|||
|
|
logger.error(f"LLM响应JSON修复后解析仍失败。原始响应:\n{llm_response_text}")
|
|||
|
|
return self._get_fallback_judgment()
|
|||
|
|
|
|||
|
|
except Exception as e:
|
|||
|
|
logger.error(f"LLM推理过程出错: {e}")
|
|||
|
|
return self._get_fallback_judgment()
|
|||
|
|
|
|||
|
|
def _robust_json_parse(self, raw_text: str) -> Optional[Dict]:
|
|||
|
|
"""
|
|||
|
|
尝试从原始文本中稳健地解析JSON。
|
|||
|
|
步骤：1. 提取JSON块。 2. 清理常见问题。 3. 尝试解析。 4. 如失败，尝试修复常见问题后重试。
|
|||
|
|
"""
|
|||
|
|
import json
|
|||
|
|
import re
|
|||
|
|
|
|||
|
|
# 步骤1: 提取最可能包含JSON的部分
|
|||
|
|
# 去除可能的Markdown代码块标记
|
|||
|
|
text = raw_text.strip()
|
|||
|
|
if text.startswith('```'):
|
|||
|
|
# 移除开头的 ```json 或 ```
|
|||
|
|
lines = text.split('\n', 1)
|
|||
|
|
if len(lines) > 1:
|
|||
|
|
text = lines[1].strip()
|
|||
|
|
if text.endswith('```'):
|
|||
|
|
text = text[:-3].strip()
|
|||
|
|
|
|||
|
|
# 查找第一个 '{' 和最后一个 '}' 的位置
|
|||
|
|
start = text.find('{')
|
|||
|
|
end = text.rfind('}')
|
|||
|
|
|
|||
|
|
if start == -1 or end == -1 or start >= end:
|
|||
|
|
logger.warning(f"在LLM响应中未找到有效的JSON结构。")
|
|||
|
|
return None
|
|||
|
|
|
|||
|
|
json_str = text[start:end + 1] # 提取从 { 到 } 的字符串
|
|||
|
|
|
|||
|
|
# 步骤2: 首次尝试直接解析
|
|||
|
|
try:
|
|||
|
|
return json.loads(json_str)
|
|||
|
|
except json.JSONDecodeError as e:
|
|||
|
|
logger.warning(f"首次JSON解析失败 (位置 {e.pos})，尝试清理和修复。错误: {e.msg}")
|
|||
|
|
# 记录出错位置附近的文本，便于调试
|
|||
|
|
snippet_start = max(0, e.pos - 20)
|
|||
|
|
snippet_end = min(len(json_str), e.pos + 20)
|
|||
|
|
logger.debug(f"错误位置附近文本: ...{json_str[snippet_start:snippet_end]}...")
|
|||
|
|
|
|||
|
|
# 步骤3: 进行一系列文本清理和修复尝试
|
|||
|
|
repaired_json = json_str
|
|||
|
|
pattern = r'(?<!\\)"(\w+)"'
|
|||
|
|
def replace_inner_quotes(match):
|
|||
|
|
# 将匹配到的 "word" 替换为 \"word\"
|
|||
|
|
return r'\"' + match.group(1) + r'\"'
|
|||
|
|
|
|||
|
|
repaired_json = re.sub(pattern, replace_inner_quotes, repaired_json)
|
|||
|
|
|
|||
|
|
# 尝试2: 修复不匹配的引号 (例如，字符串以单引号开始但以双引号结束)
|
|||
|
|
# 这里我们只是简单地移除字符串开头/结尾可能存在的杂散引号或换行符
|
|||
|
|
repaired_json = repaired_json.strip()
|
|||
|
|
|
|||
|
|
# 尝试3: 修复不规范的布尔值或null (例如 True, False, Null)
|
|||
|
|
repaired_json = re.sub(r':\s*True\b', ': true', repaired_json)
|
|||
|
|
repaired_json = re.sub(r':\s*False\b', ': false', repaired_json)
|
|||
|
|
repaired_json = re.sub(r':\s*Null\b', ': null', repaired_json, flags=re.IGNORECASE)
|
|||
|
|
repaired_json = re.sub(r':\s*None\b', ': null', repaired_json)
|
|||
|
|
|
|||
|
|
# 步骤4: 尝试解析修复后的JSON
|
|||
|
|
try:
|
|||
|
|
return json.loads(repaired_json)
|
|||
|
|
except json.JSONDecodeError as e2:
|
|||
|
|
logger.error(f"修复后JSON解析仍然失败 (位置 {e2.pos})。修复后文本片段:\n...{repaired_json[max(0, e2.pos-50):e2.pos+50]}...")
|
|||
|
|
return None
|
|||
|
|
|
|||
|
|
def _get_fallback_judgment(self) -> Dict:
|
|||
|
|
"""LLM调用失败时的回退方案"""
|
|||
|
|
return {
|
|||
|
|
"verdict": "未知",
|
|||
|
|
"confidence": 0.0,
|
|||
|
|
"primary_reasoning": "LLM推理过程发生错误，使用回退逻辑。",
|
|||
|
|
"constraint_analysis": {
|
|||
|
|
"functional_completeness": {"status": "未知", "reason": "错误"},
|
|||
|
|
"architectural_fit": {"status": "未知", "reason": "错误"},
|
|||
|
|
"module_coordination": {"status": "未知", "reason": "错误"},
|
|||
|
|
"semantic_coverage": {"status": "未知", "reason": "错误"}
|
|||
|
|
},
|
|||
|
|
"most_relevant_functions": [],
|
|||
|
|
"gap_analysis": "系统内部错误。"
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
def _format_final_result(self, llm_judgment: Dict, relevant_functions: List[Dict]) -> Dict:
|
|||
|
|
"""将LLM的评判结果格式化为兼容原接口的最终输出"""
|
|||
|
|
# 映射 verdict 到 implemented
|
|||
|
|
verdict_map = {
|
|||
|
|
"完整实现": True,
|
|||
|
|
"部分实现": True, # 部分实现也视为某种程度的实现
|
|||
|
|
"未实现": False,
|
|||
|
|
"未知": False
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
implemented = verdict_map.get(llm_judgment.get("verdict", "未知"), False)
|
|||
|
|
|
|||
|
|
# 构建详细的约束分析结果，兼容旧版格式的同时提供更丰富的信息
|
|||
|
|
constraint_scores = {}
|
|||
|
|
constraint_status = llm_judgment.get("constraint_analysis", {})
|
|||
|
|
# 为每个约束生成一个模拟分数（基于状态），用于兼容性
|
|||
|
|
for cons_name, cons_info in constraint_status.items():
|
|||
|
|
status_score_map = {"满足": 1.0, "匹配": 1.0, "协调": 1.0, "覆盖": 1.0,
|
|||
|
|
"部分满足": 0.5, "部分匹配": 0.5, "部分协调": 0.5, "部分覆盖": 0.5,
|
|||
|
|
"不满足": 0.0, "不匹配": 0.0, "不协调": 0.0, "未覆盖": 0.0,
|
|||
|
|
"未知": 0.0}
|
|||
|
|
score = status_score_map.get(cons_info.get("status", "未知"), 0.0)
|
|||
|
|
constraint_scores[cons_name] = {
|
|||
|
|
"score": score,
|
|||
|
|
"passed": score >= 0.5, # 0.5作为通过的阈值
|
|||
|
|
"details": cons_info.get("reason", "")
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
# 选取相似度最高的函数作为 most_relevant_function
|
|||
|
|
most_relevant = None
|
|||
|
|
if relevant_functions:
|
|||
|
|
best_func = max(relevant_functions, key=lambda x: x["similarity"])
|
|||
|
|
most_relevant = {
|
|||
|
|
"name": best_func["name"],
|
|||
|
|
"file": os.path.basename(best_func.get("file", "未知文件")),
|
|||
|
|
"similarity": best_func["similarity"],
|
|||
|
|
"summary": best_func.get("summary", "无摘要")[:100]
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
# 构建最终结果
|
|||
|
|
final_result = {
|
|||
|
|
"implemented": implemented,
|
|||
|
|
"reason": llm_judgment.get("primary_reasoning", "无理由提供"),
|
|||
|
|
"confidence": llm_judgment.get("confidence", 0.0),
|
|||
|
|
"total_score": llm_judgment.get("confidence", 0.0), # 用confidence作为total_score
|
|||
|
|
"passed_constraints": sum(1 for cs in constraint_scores.values() if cs.get("passed", False)),
|
|||
|
|
"total_constraints": len(constraint_scores),
|
|||
|
|
"constraint_scores": constraint_scores,
|
|||
|
|
"most_relevant_function": most_relevant,
|
|||
|
|
"relevant_functions_count": len(relevant_functions),
|
|||
|
|
# 新增字段，提供更丰富的LLM推理结果
|
|||
|
|
"llm_judgment": {
|
|||
|
|
"verdict": llm_judgment.get("verdict", "未知"),
|
|||
|
|
"gap_analysis": llm_judgment.get("gap_analysis", ""),
|
|||
|
|
"most_relevant_functions_roles": llm_judgment.get("most_relevant_functions", [])
|
|||
|
|
},
|
|||
|
|
"analysis_method": "llm_constraint_reasoning" # 标识使用了新方法
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
return final_result
|
|||
|
|
|
|||
|
|
def __init__(self, config: Optional[Dict] = None):
|
|||
|
|
"""初始化检索器"""
|
|||
|
|
self.config = config or {}
|
|||
|
|
self.faiss_index = None
|
|||
|
|
self.contexts = None
|
|||
|
|
self.metadatas = None
|
|||
|
|
self.knowledge_graph = None
|
|||
|
|
self.kg_analyzer = None
|
|||
|
|
|
|||
|
|
# 从配置加载路径
|
|||
|
|
self.vector_db_path = config.get("vector_db_path", VECTOR_DB_PATH)
|
|||
|
|
self.metadata_path = config.get("metadata_path", METADATA_PATH)
|
|||
|
|
self.knowledge_graph_path = config.get("knowledge_graph_path", KNOWLEDGE_GRAPH_PATH)
|
|||
|
|
|
|||
|
|
# 新增：初始化提示词模板
|
|||
|
|
self.llm_judgment_prompt_template = self._get_judgment_prompt_template()
|
|||
|
|
|
|||
|
|
def load_knowledge_base(self) -> bool:
|
|||
|
|
"""加载知识库文件"""
|
|||
|
|
try:
|
|||
|
|
logger.info("正在加载知识库...")
|
|||
|
|
|
|||
|
|
# 1. 加载FAISS索引
|
|||
|
|
if not os.path.exists(self.vector_db_path):
|
|||
|
|
logger.error(f"向量数据库文件不存在: {self.vector_db_path}")
|
|||
|
|
return False
|
|||
|
|
self.faiss_index = faiss.read_index(self.vector_db_path)
|
|||
|
|
|
|||
|
|
# 2. 加载元数据
|
|||
|
|
if not os.path.exists(self.metadata_path):
|
|||
|
|
logger.error(f"元数据文件不存在: {self.metadata_path}")
|
|||
|
|
return False
|
|||
|
|
with open(self.metadata_path, "r", encoding="utf-8") as f:
|
|||
|
|
self.metadatas = json.load(f)
|
|||
|
|
|
|||
|
|
# 3. 加载知识图谱
|
|||
|
|
if os.path.exists(self.knowledge_graph_path):
|
|||
|
|
with open(self.knowledge_graph_path, "r", encoding="utf-8") as f:
|
|||
|
|
self.knowledge_graph = json.load(f)
|
|||
|
|
|
|||
|
|
# 初始化知识图谱分析器
|
|||
|
|
self.kg_analyzer = KnowledgeGraphAnalyzer(self.knowledge_graph)
|
|||
|
|
else:
|
|||
|
|
logger.warning(f"知识图谱文件不存在: {self.knowledge_graph_path}")
|
|||
|
|
self.knowledge_graph = None
|
|||
|
|
self.kg_analyzer = None
|
|||
|
|
|
|||
|
|
# 4. 重建上下文文本
|
|||
|
|
self.contexts = []
|
|||
|
|
for meta in self.metadatas:
|
|||
|
|
func_name = meta.get("name", "未知函数")
|
|||
|
|
file_path = meta.get("file", "未知文件")
|
|||
|
|
summary = meta.get("summary", "无摘要")
|
|||
|
|
|
|||
|
|
context_str = (
|
|||
|
|
f"【函数名】{func_name}\n"
|
|||
|
|
f"【所在文件】{os.path.basename(file_path)}\n"
|
|||
|
|
f"【功能摘要】{summary}"
|
|||
|
|
)
|
|||
|
|
self.contexts.append(context_str)
|
|||
|
|
|
|||
|
|
logger.info(f"知识库加载成功: {len(self.metadatas)} 个函数")
|
|||
|
|
return True
|
|||
|
|
|
|||
|
|
except Exception as e:
|
|||
|
|
logger.error(f"加载知识库失败: {e}")
|
|||
|
|
return False
|
|||
|
|
|
|||
|
|
def search_by_semantic(self, query: str, top_k: int = TOP_K) -> List[Dict]:
|
|||
|
|
"""语义搜索：基于向量相似度查找相关函数"""
|
|||
|
|
if not self.faiss_index or not self.contexts:
|
|||
|
|
logger.error("知识库未加载")
|
|||
|
|
return []
|
|||
|
|
|
|||
|
|
try:
|
|||
|
|
# 生成查询向量
|
|||
|
|
query_vec = get_qwen_embedding(query)
|
|||
|
|
query_vec = np.expand_dims(query_vec, axis=0).astype(np.float32)
|
|||
|
|
|
|||
|
|
# 搜索最近邻
|
|||
|
|
distances, indices = self.faiss_index.search(query_vec, top_k)
|
|||
|
|
|
|||
|
|
# 处理结果
|
|||
|
|
relevant_functions = []
|
|||
|
|
for i, idx in enumerate(indices[0]):
|
|||
|
|
similarity = 1 - 0.5 * float(distances[0][i])
|
|||
|
|
if similarity >= MIN_SIMILARITY_THRESHOLD:
|
|||
|
|
func_info = {
|
|||
|
|
"index": int(idx),
|
|||
|
|
"similarity": similarity,
|
|||
|
|
"context": self.contexts[idx]
|
|||
|
|
}
|
|||
|
|
# 提取元数据
|
|||
|
|
func_info.update(self._get_function_metadata(idx))
|
|||
|
|
relevant_functions.append(func_info)
|
|||
|
|
|
|||
|
|
return relevant_functions
|
|||
|
|
|
|||
|
|
except Exception as e:
|
|||
|
|
logger.error(f"语义搜索失败: {e}")
|
|||
|
|
return []
|
|||
|
|
|
|||
|
|
def _get_function_metadata(self, index: int) -> Dict:
|
|||
|
|
"""获取函数的元数据"""
|
|||
|
|
if index < 0 or index >= len(self.metadatas):
|
|||
|
|
return {}
|
|||
|
|
|
|||
|
|
meta = self.metadatas[index]
|
|||
|
|
func_name = meta.get("name", "未知函数")
|
|||
|
|
file_path = meta.get("file", "未知文件")
|
|||
|
|
|
|||
|
|
return {
|
|||
|
|
"name": func_name,
|
|||
|
|
"file": file_path,
|
|||
|
|
"summary": meta.get("summary", "无摘要"),
|
|||
|
|
"calls": meta.get("calls", []),
|
|||
|
|
"called_by": meta.get("called_by", [])
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
def _extract_function_name(self, context: str) -> str:
|
|||
|
|
"""从上下文文本中提取函数名"""
|
|||
|
|
match = re.search(r"【函数名】(.*)", context)
|
|||
|
|
if match:
|
|||
|
|
return match.group(1).strip()
|
|||
|
|
return "未知函数"
|