feat: 读标补齐五类清单与项目基本信息，修复国动样本暴露的 4 个问题

- 文字层 v2：页眉页脚带单独成句；写到页尾的句子和下一页首句接续（出处两页各一条） - 结构识别（最小版）：十进制编号、冒号引出的列表找上级条款；上级条款写明无效的，条目照样认定 - 抽取：资格、评分、日期与递交、技术参数、项目基本信息各一份提示词；实质性条款出 v2 （认上级条款、补「做无效响应处理」等写法）；各类取值核对；按首句合并；拆开被合成一条的列举条款；描述标注方式的句子和表头碎片不当条款；* 号标注 - 项目：解析出的名称、编号、采购人、预算、截止时间、分包写回项目，用户改过的不覆盖 - 前端：清单项显示分值、参数要求、日期、证明材料 - 测试 40 个，新增两页虚构样例 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-03 19:32:25 +08:00
parent ee4bb23488
commit a95e27a290
29 changed files with 1971 additions and 230 deletions
--- a/server/src/bid_doctor/ai/providers/base.py
+++ b/server/src/bid_doctor/ai/providers/base.py
@@ -25,4 +25,4 @@ class TransientError(Exception):
 class Provider(Protocol):
    name: str

-    async def complete_json(self, system: str, user: str, capability: str) -> Completion: ...
+    async def complete_json(self, system: str, user: str, capability: str, task: str) -> Completion: ...
--- a/server/src/bid_doctor/ai/providers/fake.py
+++ b/server/src/bid_doctor/ai/providers/fake.py
@@ -1,40 +1,165 @@
 """确定性的假模型：开发和测试用，CI 不调用真实模型（架构 §8）。

-只认 extract.substantive：把带 ★ 或 <iu> 的句子当作实质性条款，逐字摘一段原文作为引文。
+按任务用关键词规则给出结构化结果，输出格式和真实模型一致，好让三道校验、合并、写库这些环节在测试里都跑到。
+它不理解文字，召回和准确率都没有参考价值。
 """

 from __future__ import annotations

 import json
 import re
+from collections.abc import Callable
+from typing import Any

 from bid_doctor.ai.providers.base import Completion

 LINE = re.compile(r"^\[([^\]]+)\] (.*)$")
 TAG = re.compile(r"</?(iu|u|i)>")
+PARENT = re.compile(r"〔上级：(.*)〕$")
+INVALID = re.compile(r"无效|否决|未实质性响应")
+DATE = re.compile(r"(\d{4})年(\d{1,2})月(\d{1,2})日\s*(\d{1,2})[时:：](\d{2})?")
+NUM_REQ = re.compile(
+    r"(不低于|不少于|≥|不小于|不大于|不超过|≤)\s*(\d+(?:\.\d+)?)\s*([A-Za-z%]+|[个台套项条次人天年月日分秒])?"
+)
+POINTS = re.compile(r"[（(](\d+(?:\.\d+)?)\s*分[）)]")
+
+
+def _lines(user: str) -> list[tuple[str, str, str, str]]:
+    """(句子编号, 带格式标记的文本, 纯文本, 上级条款)"""
+    out = []
+    for raw in user.split("## 待审分块", 1)[-1].splitlines():
+        m = LINE.match(raw.strip())
+        if not m:
+            continue
+        sid, marked = m.groups()
+        parent = ""
+        if pm := PARENT.search(marked):
+            parent, marked = pm.group(1), marked[: pm.start()]
+        out.append((sid, marked, TAG.sub("", marked), parent))
+    return out
+
+
+def _ev(sid: str, plain: str) -> list[dict[str, str]]:
+    return [{"sid": sid, "quote": plain.lstrip("★*＊ ")[:20] or plain}]
+
+
+def _substantive(lines: list[tuple[str, str, str, str]]) -> dict[str, Any]:
+    items = []
+    for sid, marked, plain, parent in lines:
+        if "★" in plain:
+            basis = "star"
+        elif "<iu>" in marked:
+            basis = "italic_underline"
+        elif INVALID.search(plain) or INVALID.search(parent):
+            basis = "explicit_invalid"
+        else:
+            continue
+        if any(k in plain for k in ("标注", "实质性要求", "条款响应", "斜体")) and len(plain) < 60:
+            continue  # 只是在说明标注规则的句子，不算条款（和提示词的「不要输出」一致）
+        items.append({"summary": plain.lstrip("★*＊ ")[:40], "basis": basis, "evidence": _ev(sid, plain)})
+    return {"items": items}
+
+
+def _qualification(lines: list[tuple[str, str, str, str]]) -> dict[str, Any]:
+    items = [
+        {
+            "summary": p[:40],
+            "kind": "credit" if "信用" in p else "specific",
+            "evidence_required": None,
+            "lot_no": None,
+            "evidence": _ev(sid, p),
+        }
+        for sid, _, p, _ in lines
+        if "资格" in p and any(k in p for k in ("具备", "须", "应", "具有"))
+    ]
+    return {"items": items}
+
+
+def _scoring(lines: list[tuple[str, str, str, str]]) -> dict[str, Any]:
+    items = []
+    for sid, _, p, _ in lines:
+        if m := POINTS.search(p):
+            items.append(
+                {
+                    "factor": p[: m.start()].strip("：: ")[-12:] or "评分项",
+                    "part": "technical",
+                    "points": m.group(1),
+                    "criteria": p[:80],
+                    "evidence_required": None,
+                    "evidence": _ev(sid, p),
+                }
+            )
+    return {"items": items}
+
+
+def _iso(y: str, mo: str, d: str, h: str, mi: str | None) -> str:
+    return f"{y}-{int(mo):02d}-{int(d):02d}T{int(h):02d}:{mi or '00'}"
+
+
+def _schedule(lines: list[tuple[str, str, str, str]]) -> dict[str, Any]:
+    items = []
+    for sid, _, p, _ in lines:
+        if m := DATE.search(p):
+            y, mo, d, h, mi = m.groups()
+            kind = "bid_deadline" if "截止" in p else "opening" if "开标" in p or "磋商时间" in p else "other"
+            items.append(
+                {
+                    "summary": p[:40],
+                    "date_kind": kind,
+                    "at": _iso(y, mo, d, h, mi),
+                    "deposit_amount": None,
+                    "submission_mode": None,
+                    "evidence": _ev(sid, p),
+                }
+            )
+    return {"items": items}
+
+
+def _tech_param(lines: list[tuple[str, str, str, str]]) -> dict[str, Any]:
+    ops = {"不低于": ">=", "不少于": ">=", "≥": ">=", "不小于": ">=", "不大于": "<=", "不超过": "<=", "≤": "<="}
+    items = []
+    for sid, _, p, _ in lines:
+        if m := NUM_REQ.search(p):
+            items.append(
+                {
+                    "param": p[: m.start()].lstrip("★*＊ ")[-10:] or "参数",
+                    "requirement": p.lstrip("★ ")[:60],
+                    "op": ops[m.group(1)],
+                    "value": m.group(2),
+                    "unit": m.group(3) or None,
+                    "device": None,
+                    "lot_no": None,
+                    "evidence": _ev(sid, p),
+                }
+            )
+    return {"items": items}
+
+
+def _project_info(lines: list[tuple[str, str, str, str]]) -> dict[str, Any]:
+    fields: dict[str, Any] = {}
+    for sid, _, p, _ in lines:
+        for key, label in (("name", "项目名称"), ("code", "项目编号"), ("purchaser", "采购人")):
+            if key not in fields and (m := re.search(label + r"[：:]\s*([^，。；]+)", p)):
+                fields[key] = {"value": m.group(1).strip(), "sid": sid, "quote": p[:30]}
+        if "budget_total" not in fields and (m := re.search(r"预算[^0-9]{0,8}(\d+(?:\.\d+)?)\s*(万元|元)", p)):
+            fields["budget_total"] = {"value": m.group(1), "unit": m.group(2), "sid": sid, "quote": p[:30]}
+    return {"fields": fields, "lots": []}
+
+
+TASKS: dict[str, Callable[[list[tuple[str, str, str, str]]], dict[str, Any]]] = {
+    "extract.substantive": _substantive,
+    "extract.qualification": _qualification,
+    "extract.scoring": _scoring,
+    "extract.schedule": _schedule,
+    "extract.tech_param": _tech_param,
+    "extract.project_info": _project_info,
+}


 class FakeProvider:
    name = "fake"

-    async def complete_json(self, system: str, user: str, capability: str) -> Completion:
+    async def complete_json(self, system: str, user: str, capability: str, task: str) -> Completion:
        del system, capability
-        items = []
-        for line in user.split("## 待审分块", 1)[-1].splitlines():
-            m = LINE.match(line.strip())
-            if not m:
-                continue
-            sid, marked = m.groups()
-            plain = TAG.sub("", marked)
-            if "★" in plain:
-                basis = "star"
-            elif "<iu>" in marked:
-                basis = "italic_underline"
-            else:
-                continue
-            if any(k in plain for k in ("标注", "实质性要求", "条款响应")) and len(plain) < 60:
-                continue  # 只是在说明标注规则的句子，不算条款（和提示词的「不要输出」一致）
-            body = plain.lstrip("★*＊ ").strip()
-            items.append({"summary": body[:40], "basis": basis, "evidence": [{"sid": sid, "quote": body[:20]}]})
-        data = {"items": items}
+        data = TASKS.get(task, lambda _lines: {"items": []})(_lines(user))
        return Completion(data=data, raw=json.dumps(data, ensure_ascii=False), model="fake-extractor")
--- a/server/src/bid_doctor/ai/providers/openai_compat.py
+++ b/server/src/bid_doctor/ai/providers/openai_compat.py
@@ -17,8 +17,8 @@ from bid_doctor.core.config import get_settings
 class OpenAICompatProvider:
    name = "openai_compat"

-    async def complete_json(self, system: str, user: str, capability: str) -> Completion:
-        del capability
+    async def complete_json(self, system: str, user: str, capability: str, task: str) -> Completion:
+        del capability, task
        s = get_settings()
        body = {
            "model": s.ai_model,
--- a/server/src/bid_doctor/ai/service.py
+++ b/server/src/bid_doctor/ai/service.py
@@ -35,7 +35,7 @@ async def run_json(
    error: TransientError | None = None
    for attempt in range(RETRIES + 1):
        try:
-            result = await prov.complete_json(p.body, user, p.capability)
+            result = await prov.complete_json(p.body, user, p.capability, task)
            error = None
            break
        except TransientError as e:  # 只对临时错误重试