baseline: 批次A-D 成果 + membership 半成品（测试红）

2026-09-11 23:11:35 +08:00
commit b3f3095d53
311 changed files with 40540 additions and 0 deletions
@@ -0,0 +1,183 @@
+"""AI 开单：自然语言文本 → 结构化订单行 → 匹配商品（批次 B2）。
+
+链路：文本 --LLM--> [{name, barcode?, qty, unit?}] --匹配--> {matched, unmatched}
+匹配优先级：barcode 精确 → 商品编码精确 → 品名精确 → 品名包含 → 相似度（difflib）
+**无 LLM KEY 时抛 LlmUnavailable**（视图翻译为 400 明确错误，不 500）。
+"""
+
+from __future__ import annotations
+
+import re
+from decimal import Decimal, InvalidOperation
+from difflib import SequenceMatcher
+
+from . import llm
+
+
+class LlmUnavailable(Exception):
+    """未配置 LLM（无 AI_API_KEY），AI 开单不可用。"""
+
+
+PARSE_PROMPT = """你是进销存系统的录单助手。把下面这段人类写的订货文本，抽取成 JSON 数组。
+
+规则：
+- 每行一个对象：{"name": "商品名", "barcode": "条码或空串", "qty": 数量数字, "unit": "单位或空串"}
+- 数量缺失时默认 1；"两箱"→2、"三瓶"→3 等中文数字要转成阿拉伯数字
+- 只输出 JSON 数组，不要任何解释
+
+文本：
+{text}
+"""
+
+
+_CN_NUM = {"零": 0, "一": 1, "二": 2, "两": 2, "三": 3, "四": 4, "五": 5,
+           "六": 6, "七": 7, "八": 8, "九": 9, "十": 10}
+
+
+def _cn_to_int(token: str) -> int | None:
+    """把"三"/"十二"/"两"等中文数字转 int（够用即可，不追求完整语法）。"""
+    if token in _CN_NUM:
+        return _CN_NUM[token]
+    if token.startswith("十") and len(token) == 2 and token[1] in _CN_NUM:
+        return 10 + _CN_NUM[token[1]]
+    if len(token) == 2 and token[0] in _CN_NUM and token[1] == "十":
+        return _CN_NUM[token[0]] * 10
+    if len(token) == 3 and token[0] in _CN_NUM and token[1] == "十" and token[2] in _CN_NUM:
+        return _CN_NUM[token[0]] * 10 + _CN_NUM[token[2]]
+    return None
+
+
+def fallback_parse(text: str) -> list:
+    """无 LLM 时的规则兜底：按行/逗号切分，抓 "商品名 数字单位" 模式。
+
+    仅在显式要求（allow_rule_fallback=True）时使用——默认严格走 LLM，
+    保证"没有 KEY 就明确报不可用"的验收标准。
+    """
+    items = []
+    for raw in re.split(r"[\n\r,，;；]+", text or ""):
+        line = raw.strip()
+        if not line:
+            continue
+        m = re.search(r"(\d+(?:\.\d+)?)\s*([^\s\d]*)", line)
+        qty = Decimal("1")
+        unit = ""
+        name = line
+        if m:
+            qty = Decimal(m.group(1))
+            unit = (m.group(2) or "").strip()
+            name = (line[:m.start()] + line[m.end():]).strip() or line
+        else:
+            # 中文数字：三箱 / 两瓶
+            m2 = re.search(r"([零一二两三四五六七八九十]+)\s*([^\s\d]*)", line)
+            if m2:
+                n = _cn_to_int(m2.group(1))
+                if n:
+                    qty = Decimal(n)
+                    unit = (m2.group(2) or "").strip()
+                    name = (line[:m2.start()] + line[m2.end():]).strip() or line
+        # 去掉常见量词残留
+        name = re.sub(r"(箱|瓶|件|个|包|袋|提|盒|罐|桶|斤|公斤|kg|Kg|KG)$", "", name).strip()
+        if name:
+            items.append({"name": name, "barcode": "", "qty": float(qty), "unit": unit})
+    return items
+
+
+def _norm_qty(value) -> Decimal:
+    try:
+        q = Decimal(str(value))
+    except (InvalidOperation, TypeError, ValueError):
+        q = Decimal("1")
+    return q if q > 0 else Decimal("1")
+
+
+def extract_items(text: str, *, allow_rule_fallback: bool = False) -> list:
+    """文本 → [{name, barcode, qty, unit}]。无 KEY 抛 LlmUnavailable。"""
+    if not text or not text.strip():
+        return []
+
+    if not llm.available():
+        if allow_rule_fallback:
+            return fallback_parse(text)
+        raise LlmUnavailable("未配置 AI 服务（AI_API_KEY），AI 录单暂不可用")
+
+    data = llm.extract_json(PARSE_PROMPT.format(text=text.strip()))
+    if not isinstance(data, list):
+        # LLM 偶发返回 {"items": [...]}
+        if isinstance(data, dict) and isinstance(data.get("items"), list):
+            data = data["items"]
+        else:
+            return []
+
+    items = []
+    for row in data:
+        if not isinstance(row, dict):
+            continue
+        name = str(row.get("name") or "").strip()
+        if not name:
+            continue
+        items.append({
+            "name": name,
+            "barcode": str(row.get("barcode") or "").strip(),
+            "qty": float(_norm_qty(row.get("qty"))),
+            "unit": str(row.get("unit") or "").strip(),
+        })
+    return items
+
+
+def match_products(tenant, items: list, *, unit_resolver=None) -> dict:
+    """把抽取结果匹配到商品档案。
+
+    返回 {matched: [{...item, product_id, product_code, product_name, match_by,
+    score, unit_id?, unit_name?, price?}], unmatched: [...]}
+    """
+    from apps.catalog.models import Product
+
+    products = list(Product.objects.filter(tenant=tenant, is_deleted=False))
+    by_barcode = {p.barcode: p for p in products if p.barcode}
+    by_code = {p.code.lower(): p for p in products}
+    by_name = {p.name: p for p in products}
+
+    matched, unmatched = [], []
+    for item in items:
+        hit, how, score = None, "", 0.0
+
+        if item.get("barcode") and item["barcode"] in by_barcode:
+            hit, how, score = by_barcode[item["barcode"]], "barcode", 100.0
+        elif item["name"].lower() in by_code:
+            hit, how, score = by_code[item["name"].lower()], "code", 100.0
+        elif item["name"] in by_name:
+            hit, how, score = by_name[item["name"]], "name_exact", 100.0
+        else:
+            # 包含匹配（长度优先，避免短名吃掉长名）
+            cands = [p for p in products if item["name"] and item["name"] in p.name]
+            if cands:
+                hit = sorted(cands, key=lambda p: len(p.name))[0]
+                how, score = "name_contains", 85.0
+            else:
+                # 相似度兜底
+                best, best_score = None, 0.0
+                for p in products:
+                    r = SequenceMatcher(None, item["name"], p.name).ratio()
+                    if r > best_score:
+                        best, best_score = p, r
+                if best is not None and best_score >= 0.6:
+                    hit, how, score = best, "fuzzy", round(best_score * 100, 1)
+
+        if hit is None:
+            unmatched.append(item)
+            continue
+
+        row = {
+            **item,
+            "product_id": hit.id,
+            "product_code": hit.code,
+            "product_name": hit.name,
+            "match_by": how,
+            "score": score,
+            "base_price": str(hit.sale_price),
+        }
+        if unit_resolver is not None:
+            row.update(unit_resolver(tenant, hit, item))
+        matched.append(row)
+
+    return {"matched": matched, "unmatched": unmatched}