落地陆运群内协同：按填写样例拉群并只列本单协同项，进群把私聊已抽到的字段带上。

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-09-18 13:38:02 +08:00
parent be95d823a2
commit 982b3a2317
25 changed files with 1451 additions and 315 deletions
@@ -23,12 +23,31 @@ _MODE_WORDS = (
 _LABEL_LINE = re.compile(
    r"^(起运地|起运港|目的地|目的港|品名|件数|毛重|体积|重量|重量\(KG\)|体积\(CBM\)|"
    r"包装方式|包装类型|货量|货物数量|整柜或拼柜|箱型箱量|报价日期|车型/数量|车型数量|"
-    r"贸易条款|货好时间|商品海关编码|海关编码|HS编码|HS|是否含油|是否含电|是否含磁)\s*[:：]\s*(.+)$"
+    r"贸易条款|货好时间|商品海关编码|海关编码|HS编码|HS|是否含油|是否含电|是否含磁|"
+    r"客户名称|货值|是否为危险品)\s*[:：]\s*(.+)$"
 )
 _LABEL_INLINE = re.compile(
    r"(起运地|起运港|目的地|目的港|品名|件数|毛重|体积|重量|重量\(KG\)|体积\(CBM\)|"
    r"包装方式|包装类型|货量|货物数量|整柜或拼柜|箱型箱量|报价日期|车型/数量|车型数量|"
-    r"贸易条款|货好时间|商品海关编码|海关编码|HS编码|HS|是否含油|是否含电|是否含磁)\s*[:：]\s*(\S+)"
+    r"贸易条款|货好时间|商品海关编码|海关编码|HS编码|HS|是否含油|是否含电|是否含磁|"
+    r"客户名称|货值|是否为危险品)\s*[:：]\s*(\S+)"
+)
+# 群里协同标签，冒号可省略：「客户名称张三」「货值 10万」。长词优先。
+_COLLAB_LABELS = (
+    "是否为危险品",
+    "商品海关编码",
+    "车型/数量",
+    "包装方式",
+    "客户名称",
+    "海关编码",
+    "贸易条款",
+    "货好时间",
+    "是否含油",
+    "是否含电",
+    "是否含磁",
+    "HS编码",
+    "货值",
+    "HS",
 )


@@ -64,6 +83,23 @@ def parse_labeled_facts(text: str) -> dict[str, str]:
    return facts


+def _merge_spoken_collab(text: str, facts: dict[str, str]) -> dict[str, str]:
+    """
+    私聊已经说到的协同项写进事实，进群才能预填。
+
+    不覆盖已有询价键。不拿去补问、不挡查价。
+    """
+    from agent.schema.field_validate import harvest_oral_collab, normalize_facts
+
+    merged = dict(facts or {})
+    extra = dict(harvest_oral_collab(text))
+    extra.update(normalize_facts(harvest_collab_labels(text)))
+    for key, val in extra.items():
+        if val and not str(merged.get(key) or "").strip():
+            merged[key] = val
+    return merged
+
+
 def extract_inquiry_snapshot(
    text: str,
    *,
@@ -80,11 +116,12 @@ def extract_inquiry_snapshot(
    标签行覆盖 B 的同名字段。B 未出现的字段保持空，不本地猜港口。
    件数/毛重/体积若原话已带 TMS 单位（件/KGS/CBM），空字段可机读补上。
    """
-    from agent.schema.field_validate import harvest_oral_measures, normalize_facts
+    from agent.schema.field_validate import harvest_oral_collab, harvest_oral_measures, normalize_facts

    if injected_facts is not None:
        mode = (injected_mode or detect_transport_mode(text) or "").upper()
        facts = harvest_oral_measures(text, normalize_facts(dict(injected_facts)))
+        facts = _merge_spoken_collab(text, facts)
        subtype = str(facts.get("运输类型") or "").strip()
        return {
            "business_line": mode,
@@ -109,6 +146,7 @@ def extract_inquiry_snapshot(
        mode = mode or str(b_payload.get("business_line") or "")
        source = "deepseek_b"
    facts = harvest_oral_measures(text, normalize_facts(facts))
+    facts = _merge_spoken_collab(text, facts)
    land_subtype = str((b_payload or {}).get("land_subtype") or facts.get("运输类型") or "").strip()
    if land_subtype and not str(facts.get("运输类型") or "").strip():
        facts["运输类型"] = land_subtype
@@ -123,37 +161,90 @@ def extract_inquiry_snapshot(
    }


+def harvest_collab_labels(text: str) -> dict[str, str]:
+    """
+    群里协同标签：有无冒号都收。
+
+    「客户名称张三 包装方式编织袋 货值 10万」切成三项。
+    只说「纸箱」没有标签则不收。值取到下一个标签之前。
+    """
+    raw = text or ""
+    hits: list[tuple[int, str]] = []
+    for label in _COLLAB_LABELS:
+        start = 0
+        while True:
+            pos = raw.find(label, start)
+            if pos < 0:
+                break
+            hits.append((pos, label))
+            start = pos + len(label)
+    if not hits:
+        return {}
+    hits.sort()
+    # 同一起点只留最长标签，避免 HS 抢 HS编码。
+    kept: list[tuple[int, str]] = []
+    for pos, label in hits:
+        if kept and kept[-1][0] == pos and len(label) <= len(kept[-1][1]):
+            continue
+        if kept and kept[-1][0] == pos:
+            kept[-1] = (pos, label)
+            continue
+        kept.append((pos, label))
+    out: dict[str, str] = {}
+    for i, (pos, label) in enumerate(kept):
+        begin = pos + len(label)
+        end = kept[i + 1][0] if i + 1 < len(kept) else len(raw)
+        chunk = raw[begin:end]
+        chunk = re.sub(r"^[\s:：,，、]+", "", chunk)
+        chunk = re.sub(r"[\s,，、]+$", "", chunk)
+        if chunk:
+            out[label] = chunk
+    return out
+
+
 def extract_collab_fields(
    text: str,
    *,
    allow_b: bool = False,
    injected: Optional[dict[str, str]] = None,
+    allowed_keys: Optional[tuple[str, ...]] = None,
 ) -> dict[str, str]:
    """
-    群里协同字段：标签「键：值」+ 口语含油/含电/含磁、货好时间。
+    群里协同字段：标签（冒号可省）+ 口语含油/含电/含磁/危险品、货好时间。

-    闲聊、报价金额不进这里。商品海关编码 / HS 折成 HS编码。
-    allow_b 保留给调用方；协同六项不走询价七项的 B 合同。
+    allowed_keys 传入则只返回本单该有的项。不传则海运六项 ∪ 陆运协同键，给意图分类。
+    闲聊、只说「纸箱」不进这里。商品海关编码 / HS 折成 HS编码。
    """
    from agent.schema.field_validate import (
-        SEA_COLLAB_FIELDS,
+        ALL_COLLAB_KEYS,
        harvest_oral_collab,
        normalize_facts,
    )

    del allow_b
+    keys = tuple(allowed_keys) if allowed_keys is not None else ALL_COLLAB_KEYS
    if injected:
        return {
            k: str(v).strip()
            for k, v in normalize_facts(dict(injected)).items()
-            if k in SEA_COLLAB_FIELDS and str(v or "").strip()
+            if k in keys and str(v or "").strip()
        }
    labeled = normalize_facts(parse_labeled_facts(text))
+    labeled.update(normalize_facts(harvest_collab_labels(text)))
    oral = harvest_oral_collab(text)
+    if "是否为危险品" in oral:
+        if allowed_keys is not None and "是否为危险品" not in keys:
+            oral.pop("是否为危险品", None)
+        elif allowed_keys is None:
+            raw = text or ""
+            if "普货" in raw and not any(
+                w in raw for w in ("危险", "非危", "有危", "含危", "是否为危险品")
+            ):
+                oral.pop("是否为危险品", None)
    merged = dict(oral)
    merged.update({k: v for k, v in labeled.items() if v})
    out: dict[str, str] = {}
-    for key in SEA_COLLAB_FIELDS:
+    for key in keys:
        val = (merged.get(key) or "").strip()
        if val:
            out[key] = val