feat(server): add workbench API, DSH planner worker and RAG pipeline

2026-09-16 19:06:20 +08:00
parent 72fc5b47c7
commit d4b5b3f43a
122 changed files with 31906 additions and 0 deletions
@@ -0,0 +1,89 @@
+"""任务级成本/Token/耗时预估（交付报告 P2-1）。
+
+思路：用本组织近期已结算的真实用量外推（历史均值 × 本次段数），
+没有历史数据时用保守常量估计。不依赖任何外部服务。
+"""
+from __future__ import annotations
+
+from sqlmodel import Session, select
+
+from .models import Campaign, LLMCall, Segment, UsageEvent
+from .billing import POINTS_PER_TOKEN
+
+
+# 无历史数据时的保守单段估计（deepseek 级别的小上下文生产段）
+FALLBACK_INPUT_TOKENS = 6000
+FALLBACK_OUTPUT_TOKENS = 2500
+FALLBACK_SECONDS = 90
+
+
+def estimate_campaign_cost(session: Session, campaign: Campaign, single_segment_id: int | None = None) -> dict:
+    """single_segment_id：逐段生成模式——只按指定段估算。"""
+    if single_segment_id:
+        segments = [s for s in session.exec(select(Segment).where(
+            Segment.campaign_id == campaign.id)).all() if s.id == int(single_segment_id)]
+    else:
+        segments = session.exec(select(Segment).where(
+            Segment.campaign_id == campaign.id)).all()
+    seg_count = len(segments)
+    if not seg_count:
+        return {"ready": False, "reason": "任务还没有内容段，先完成导图拆解",
+                "segments": 0}
+
+    org_id = int(campaign.organization_id or 1)
+
+    # 历史均值：本组织最近 200 条已测量 LLM 用量事件
+    events = session.exec(select(UsageEvent).where(
+        UsageEvent.organization_id == org_id,
+        UsageEvent.service_type == "llm",
+    ).order_by(UsageEvent.id.desc()).limit(200)).all()
+    measured = [e for e in events
+                if e.status == "measured" and (e.input_quantity or 0) + (e.output_quantity or 0) > 0]
+
+    if measured:
+        avg_in = sum(e.input_quantity or 0 for e in measured) / len(measured)
+        avg_out = sum(e.output_quantity or 0 for e in measured) / len(measured)
+        source = "history"
+    else:
+        avg_in, avg_out = float(FALLBACK_INPUT_TOKENS), float(FALLBACK_OUTPUT_TOKENS)
+        source = "fallback"
+
+    # 耗时：近 200 次 LLM 调用的平均时长 × 每段约 3 次调用（生产+质检）+ 组合余量
+    durations = [float(row.duration_ms or 0) for row in session.exec(
+        select(LLMCall).order_by(LLMCall.id.desc()).limit(200)).all()
+        if (row.duration_ms or 0) > 0]
+    avg_seconds = FALLBACK_SECONDS
+    if durations:
+        avg_seconds = (sum(durations) / len(durations)) / 1000 * 3 + 15
+
+    total_in = avg_in * seg_count
+    total_out = avg_out * seg_count
+    total_tokens = int((total_in + total_out) * 1.6)   # 1.6 倍：每段附带质检与组合分摊
+    est_points = total_tokens * POINTS_PER_TOKEN
+    est_seconds = int(avg_seconds * seg_count + 20)
+
+    # 费用：历史事件费率均值
+    if measured:
+        avg_in_rate = sum(e.input_rate or 0 for e in measured) / len(measured)
+        avg_out_rate = sum(e.output_rate or 0 for e in measured) / len(measured)
+    else:
+        avg_in_rate = avg_out_rate = 0.0
+    amount_minor = int(round((total_in / 1e6 * avg_in_rate + total_out / 1e6 * avg_out_rate) * 100))
+
+    return {
+        "ready": True,
+        "source": source,
+        "segments": seg_count,
+        "estimated_input_tokens": int(total_in),
+        "estimated_output_tokens": int(total_out),
+        "estimated_tokens": total_tokens,
+        "estimated_points": est_points,
+        "estimated_amount_minor": amount_minor,
+        "estimated_seconds": est_seconds,
+        "estimated_duration_text": (f"{est_seconds // 60} 分 {est_seconds % 60} 秒"
+                                    if est_seconds >= 60 else f"{est_seconds} 秒"),
+    }
+
+
+__all__ = ["estimate_campaign_cost"]
+