SQLite 持久连接 — sandbox 不再每次查询开关连接，改为 __init__ 时建连、close() 时释放
Explorer 的 system prompt 明确告知 sandbox 规则 — "每条 SQL 必须包含聚合函数或 LIMIT"，减少 LLM 生成违规 SQL 浪费轮次 LLM 客户端单例 — 所有组件共享一个 openai.OpenAI 实例，不再各建各的 sanitize 顺序修复 — 小样本抑制放在 float round 之前，避免被 round 干扰 quick_detect 从 O(n²) 改为 O(n) — 按列聚合一次，加去重，不再对每行重复算整列统计 历史上下文实际生效 — get_context_for 的结果现在会注入到 Explorer 的初始 prompt 里，多轮分析时 LLM 能看到之前的发现
This commit is contained in:
192
agent.py
192
agent.py
@@ -1,139 +1,175 @@
|
||||
"""
|
||||
Agent 编排层 —— 调度四层架构完成分析
|
||||
|
||||
Layer 1: Planner 意图规划
|
||||
Layer 2: Explorer 自适应探索
|
||||
Layer 3: Insight 异常洞察
|
||||
Layer 4: Context 上下文记忆
|
||||
Layer 1: Planner 意图规划
|
||||
Layer 1.5: Playbook 预设匹配
|
||||
Layer 2: Explorer 自适应探索
|
||||
Layer 3: Insight 异常洞察（与图表并行）
|
||||
Layer 4: Context 上下文记忆
|
||||
Output: Reporter + Chart + Consolidator
|
||||
"""
|
||||
import os
|
||||
from typing import Optional
|
||||
from concurrent.futures import ThreadPoolExecutor, as_completed
|
||||
|
||||
from config import DB_PATH, MAX_EXPLORATION_ROUNDS
|
||||
from schema_extractor import extract_schema, schema_to_text
|
||||
from sandbox_executor import SandboxExecutor
|
||||
from planner import Planner
|
||||
from explorer import Explorer
|
||||
from insights import InsightEngine, quick_detect
|
||||
from reporter import ReportGenerator
|
||||
from context import ContextManager
|
||||
from core.config import DB_PATH, MAX_EXPLORATION_ROUNDS, PLAYBOOK_DIR, CHARTS_DIR
|
||||
from core.schema import extract_schema, schema_to_text
|
||||
from core.sandbox import SandboxExecutor
|
||||
from layers.planner import Planner
|
||||
from layers.explorer import Explorer
|
||||
from layers.insights import InsightEngine, quick_detect
|
||||
from layers.context import ContextManager
|
||||
from layers.playbook import PlaybookManager
|
||||
from output.reporter import ReportGenerator
|
||||
from output.chart import ChartGenerator
|
||||
from output.consolidator import ReportConsolidator
|
||||
|
||||
|
||||
class DataAnalysisAgent:
|
||||
"""
|
||||
数据分析 Agent
|
||||
"""数据分析 Agent —— 四层架构编排"""
|
||||
|
||||
四层架构：
|
||||
1. Planner - 理解用户意图，生成分析计划
|
||||
2. Explorer - 基于计划多轮迭代探索
|
||||
3. Insights - 从结果中检测异常、输出主动洞察
|
||||
4. Context - 管理多轮对话上下文
|
||||
|
||||
Agent 负责编排这四层，从问题到报告。
|
||||
"""
|
||||
|
||||
def __init__(self, db_path: str):
|
||||
# 数据层
|
||||
def __init__(self, db_path: str = DB_PATH):
|
||||
self.db_path = db_path
|
||||
self.schema = extract_schema(db_path)
|
||||
self.schema_text = schema_to_text(self.schema)
|
||||
self.executor = SandboxExecutor(db_path)
|
||||
|
||||
# 四层组件
|
||||
# 各层组件
|
||||
self.planner = Planner()
|
||||
self.explorer = Explorer(self.executor)
|
||||
self.insight_engine = InsightEngine()
|
||||
self.reporter = ReportGenerator()
|
||||
self.context = ContextManager()
|
||||
self.playbook_mgr = PlaybookManager(PLAYBOOK_DIR)
|
||||
self.chart_gen = ChartGenerator(output_dir=CHARTS_DIR)
|
||||
self.consolidator = ReportConsolidator()
|
||||
|
||||
# 累积图表
|
||||
self._all_charts: list[dict] = []
|
||||
|
||||
# 自动生成 Playbook
|
||||
if not self.playbook_mgr.playbooks:
|
||||
print("\n🤖 [Playbook] 未发现预设剧本，AI 自动生成中...")
|
||||
generated = self.playbook_mgr.auto_generate(self.schema_text, save_dir=PLAYBOOK_DIR)
|
||||
if generated:
|
||||
print(f" ✅ 自动生成 {len(generated)} 个剧本：")
|
||||
for pb in generated:
|
||||
print(f" 📋 {pb.name} — {pb.description}")
|
||||
else:
|
||||
print(" ⚠️ 自动生成失败，将使用纯自适应模式")
|
||||
|
||||
def analyze(self, question: str, max_rounds: Optional[int] = None) -> str:
|
||||
"""
|
||||
完整分析流程
|
||||
|
||||
Args:
|
||||
question: 用户分析问题
|
||||
max_rounds: 最大探索轮数（默认用配置值）
|
||||
|
||||
Returns:
|
||||
格式化的分析报告
|
||||
"""
|
||||
"""完整分析流程"""
|
||||
max_rounds = max_rounds or MAX_EXPLORATION_ROUNDS
|
||||
|
||||
print(f"\n{'='*60}")
|
||||
print(f"📊 {question}")
|
||||
print(f"{'='*60}")
|
||||
|
||||
# ── Layer 0: 检查上下文 ──────────────────────────
|
||||
prev_context = self.context.get_context_for(question)
|
||||
if prev_context:
|
||||
print("📎 发现历史分析上下文，将结合之前的发现")
|
||||
# Layer 0: 上下文
|
||||
prev = self.context.get_context_for(question)
|
||||
if prev:
|
||||
print("📎 发现历史上下文")
|
||||
|
||||
# ── Layer 1: 意图规划 ────────────────────────────
|
||||
# Layer 1: 意图规划
|
||||
print("\n🎯 [Layer 1] 意图规划...")
|
||||
plan = self.planner.plan(question, self.schema_text)
|
||||
|
||||
analysis_type = plan.get("analysis_type", "unknown")
|
||||
dimensions = plan.get("dimensions", [])
|
||||
# 注入历史上下文到 plan 中，让 Explorer 能看到
|
||||
if prev:
|
||||
plan["_prev_context"] = prev
|
||||
print(f" 类型: {plan.get('analysis_type', 'unknown')}")
|
||||
print(f" 维度: {', '.join(plan.get('dimensions', [])) or '自动发现'}")
|
||||
rationale = plan.get("rationale", "")
|
||||
print(f" 类型: {analysis_type}")
|
||||
print(f" 维度: {', '.join(dimensions) if dimensions else '自动发现'}")
|
||||
print(f" 理由: {rationale[:80]}{'...' if len(rationale) > 80 else ''}")
|
||||
|
||||
# ── Layer 2: 自适应探索 ──────────────────────────
|
||||
print(f"\n🔍 [Layer 2] 自适应探索 (最多 {max_rounds} 轮)...")
|
||||
steps = self.explorer.explore(plan, self.schema_text, max_rounds=max_rounds)
|
||||
# Layer 1.5: Playbook 匹配
|
||||
playbook_result = None
|
||||
if self.playbook_mgr.playbooks:
|
||||
print(f"\n📋 [Playbook] 匹配预设剧本 ({len(self.playbook_mgr.playbooks)} 个可用)...")
|
||||
playbook_result = self.playbook_mgr.match(plan, self.schema_text)
|
||||
if playbook_result:
|
||||
n = len(playbook_result.get("preset_queries", []))
|
||||
print(f" ✅ 匹配: {playbook_result['playbook_name']} ({n} 条预设查询)")
|
||||
else:
|
||||
print(" ❌ 无匹配，走纯自适应路径")
|
||||
|
||||
# Layer 2: 自适应探索
|
||||
print(f"\n🔍 [Layer 2] 自适应探索 (最多 {max_rounds} 轮)...")
|
||||
steps = self.explorer.explore(plan, self.schema_text, max_rounds=max_rounds, playbook_result=playbook_result)
|
||||
successful = sum(1 for s in steps if s.success)
|
||||
print(f"\n 完成: {len(steps)} 轮, {successful} 条成功查询")
|
||||
|
||||
# ── Layer 3: 异常洞察 ────────────────────────────
|
||||
print("\n🔎 [Layer 3] 异常洞察...")
|
||||
|
||||
# 先做规则检测
|
||||
# Layer 3 + Charts: 并行执行洞察和图表生成
|
||||
print("\n🔎 [Layer 3] 异常洞察 + 📊 图表生成（并行）...")
|
||||
rule_alerts = quick_detect(steps)
|
||||
for alert in rule_alerts:
|
||||
print(f" {alert}")
|
||||
|
||||
# 再做 LLM 深度分析
|
||||
insights = self.insight_engine.analyze(steps, question)
|
||||
if insights:
|
||||
print(f" 发现 {len(insights)} 条洞察")
|
||||
for insight in insights:
|
||||
print(f" {insight}")
|
||||
else:
|
||||
print(" 未发现异常")
|
||||
insights = []
|
||||
charts = []
|
||||
|
||||
# ── 生成报告 ────────────────────────────────────
|
||||
with ThreadPoolExecutor(max_workers=2) as pool:
|
||||
future_insights = pool.submit(self.insight_engine.analyze, steps, question)
|
||||
future_charts = pool.submit(self.chart_gen.generate, steps, question)
|
||||
|
||||
for future in as_completed([future_insights, future_charts]):
|
||||
try:
|
||||
result = future.result()
|
||||
if future == future_insights:
|
||||
insights = result
|
||||
if insights:
|
||||
print(f" 🔎 发现 {len(insights)} 条洞察")
|
||||
else:
|
||||
print(" 🔎 未发现异常")
|
||||
else:
|
||||
charts = result
|
||||
if charts:
|
||||
print(f" 📊 生成 {len(charts)} 张图表：")
|
||||
for c in charts:
|
||||
print(f" 🖼️ {c['title']} → {c['path']}")
|
||||
else:
|
||||
print(" 📊 无需生成图表")
|
||||
except Exception as e:
|
||||
print(f" ⚠️ 并行任务出错: {e}")
|
||||
|
||||
if charts:
|
||||
self._all_charts.extend(charts)
|
||||
|
||||
# 生成报告
|
||||
print("\n📝 正在生成报告...")
|
||||
report = self.reporter.generate(question, plan, steps, insights)
|
||||
report = self.reporter.generate(question, plan, steps, insights, charts=charts)
|
||||
|
||||
# 追加主动洞察
|
||||
if insights:
|
||||
insight_text = self.insight_engine.format_insights(insights)
|
||||
report += f"\n\n---\n\n{insight_text}"
|
||||
report += f"\n\n---\n\n{self.insight_engine.format_insights(insights)}"
|
||||
|
||||
# ── Layer 4: 记录上下文 ──────────────────────────
|
||||
self.context.add_session(
|
||||
question=question,
|
||||
plan=plan,
|
||||
steps=steps,
|
||||
insights=insights,
|
||||
report=report,
|
||||
)
|
||||
# Layer 4: 记录上下文
|
||||
self.context.add_session(question=question, plan=plan, steps=steps, insights=insights, report=report)
|
||||
|
||||
return report
|
||||
|
||||
def full_report(self, question: str = "") -> str:
|
||||
"""整合所有历史分析为综合报告"""
|
||||
sessions = self.context.sessions
|
||||
if not sessions:
|
||||
return "（还没有分析记录，请先执行分析）"
|
||||
print(f"\n📑 整合 {len(sessions)} 次分析为综合报告...")
|
||||
report = self.consolidator.consolidate(sessions=sessions, question=question, charts=self._all_charts or None)
|
||||
print(" ✅ 综合报告生成完成")
|
||||
return report
|
||||
|
||||
def get_schema(self) -> str:
|
||||
"""获取 Schema 文本"""
|
||||
return self.schema_text
|
||||
|
||||
def get_history(self) -> str:
|
||||
"""获取分析历史摘要"""
|
||||
return self.context.get_history_summary()
|
||||
|
||||
def get_audit(self) -> str:
|
||||
"""获取执行审计日志"""
|
||||
return self.executor.get_execution_summary()
|
||||
|
||||
def clear_history(self):
|
||||
"""清空分析历史"""
|
||||
self.context.clear()
|
||||
self._all_charts.clear()
|
||||
|
||||
def close(self):
|
||||
"""释放资源"""
|
||||
self.executor.close()
|
||||
|
||||
Reference in New Issue
Block a user