"""意图识别与字段抽取。 一次 OpenAI 兼容接口调用（DeepSeek / 千问 DashScope 都兼容），要求模型只输出 JSON： {"business_id": "...", "confidence": 0.9, "inputs": {"字段名": "值"}, "reply": "缺字段时的追问"} 接口不可用时退回关键词匹配 + 正则抽取，保证流程不中断。 """ import json import re from datetime import date, timedelta from typing import Any from backend.browser.llm_factory import get_llm_config WEEKDAYS = "一二三四五六日" # ---------- prompt ---------- def _field_desc(f: dict) -> str: parts = [f"{f['name']}（{f['label']}"] if f.get("required"): parts.append("，必填") if f.get("type") == "date": parts.append("，格式 YYYY-MM-DD") if f.get("hint"): parts.append(f"，{f['hint']}") return "".join(parts) + "）" def _catalog(businesses: list[dict]) -> str: lines = [] for b in businesses: fields = "、".join(_field_desc(f) for f in b.get("required_inputs", [])) or "无" lines.append(f"- id: {b['id']} | 名称: {b['name']} | 适用: {b.get('intent', '')} | 字段: {fields}") return "\n".join(lines) def _build_prompt(message: str, businesses: list[dict], existing: dict, fixed_business_id: str | None) -> str: today = date.today() lines = [ "你是学校办事助手的意图识别与资料抽取模块。", f"今天是 {today.isoformat()}（星期{WEEKDAYS[today.weekday()]}）。", "", "可选业务：", _catalog(businesses), "", ] if fixed_business_id: lines.append(f"业务已经确定为 {fixed_business_id}，不要更改，只需从用户这句话里补充抽取字段。") if existing: shown = {k: v for k, v in existing.items() if k not in ("goal", "business_id", "user_notes") and v} if shown: lines.append(f"已有资料（无需重复抽取，除非用户明确修改）：{json.dumps(shown, ensure_ascii=False)}") lines += [ f"用户说：{message}", "", "输出一个 JSON 对象，字段：", ' business_id：最匹配的业务 id；都不匹配时用 "generic_explore"', " confidence：0 到 1 的小数", " inputs：从用户话里抽到的字段，键用字段名。日期统一 YYYY-MM-DD，「今天/明天/昨天/下周一」按今天换算；没提到的字段不要编造，也不要输出空字符串", " reply：如果该业务还有必填字段没抽到，用一句自然的中文向用户追问，把缺的项列出来；否则为空字符串", "只输出 JSON，不要解释。", ] return "\n".join(lines) # ---------- LLM ---------- def _parse_json(text: str) -> dict | None: text = (text or "").strip() if text.startswith("```"): text = re.sub(r"^```(?:json)?\s*|\s*```$", "", text, flags=re.S) try: obj = json.loads(text) return obj if isinstance(obj, dict) else None except ValueError: pass m = re.search(r"\{.*\}", text, flags=re.S) if m: try: obj = json.loads(m.group(0)) return obj if isinstance(obj, dict) else None except ValueError: return None return None async def _call_llm(prompt: str, cfg: dict) -> dict | None: """一次 JSON 调用。主 Key 402 余额不足时切兜底 Key 重发（cfg 来自 llm_factory.get_llm_config）。""" from openai import AsyncOpenAI from backend.browser import llm_keys messages = [ {"role": "system", "content": "你只输出 JSON 对象，不输出任何其他文字。"}, {"role": "user", "content": prompt}, ] async def ask(api_key: str): client = AsyncOpenAI(api_key=api_key, base_url=cfg["base_url"], timeout=45) try: return await client.chat.completions.create( model=cfg["model"], messages=messages, temperature=0, response_format={"type": "json_object"} ) except Exception as error: # noqa: BLE001 千问部分模型不支持 response_format，去掉再试一次；余额不足直接抛 if llm_keys.is_balance_error(error): raise return await client.chat.completions.create(model=cfg["model"], messages=messages, temperature=0) key = cfg["api_key"] try: resp = await ask(key) if key and key == cfg.get("primary_api_key"): llm_keys.clear_primary_exhausted() except Exception as error: # noqa: BLE001 fallback = cfg.get("fallback_api_key") or "" if fallback and key != fallback and llm_keys.is_balance_error(error): llm_keys.mark_primary_exhausted(str(error)) resp = await ask(fallback) else: raise return _parse_json(resp.choices[0].message.content or "") # ---------- 关键词退路 ---------- def _resolve_relative_date(text: str) -> str | None: today = date.today() table = {"今天": 0, "今日": 0, "明天": 1, "明日": 1, "后天": 2, "昨天": -1, "昨日": -1, "前天": -2} for k, d in table.items(): if k in text: return (today + timedelta(days=d)).isoformat() m = re.search(r"(\d{4})[-/年.](\d{1,2})[-/月.](\d{1,2})", text) if m: return f"{m.group(1)}-{int(m.group(2)):02d}-{int(m.group(3)):02d}" m = re.search(r"(\d{1,2})月(\d{1,2})[日号]", text) if m: return f"{today.year}-{int(m.group(1)):02d}-{int(m.group(2)):02d}" return None def _phrase_score(phrase: str, message: str) -> int: """短语与用户原话的最长公共子串长度（少于 2 个字不算）。 不能要求整段短语原样出现：知识里写的是「每月查课日志」，用户说的是「帮我填今天的查课日志」， 整段匹配为 0，只有最长公共子串「查课日志」能把这条业务认出来。 """ import difflib if len(phrase) < 2: return 0 match = difflib.SequenceMatcher(None, phrase, message, autojunk=False).find_longest_match(0, len(phrase), 0, len(message)) return match.size if match.size >= 2 else 0 def _fallback(message: str, businesses: list[dict], fixed_business_id: str | None) -> dict: """模型不可用时的保底路由：按业务名 / intent 与用户原话的字面重合度打分。""" best, best_score = None, 0 if fixed_business_id: best = next((b for b in businesses if b["id"] == fixed_business_id), None) else: for b in businesses: if b.get("id") == "generic_explore": continue # 兜底业务不参与打分，避免它的 intent（「查找某个入口」等）抢走具体业务 phrases = [p for p in re.split(r"[、，,/ \s]+", f"{b.get('name', '')} {b.get('intent', '')}") if p] score = sum(_phrase_score(p, message) for p in phrases) if score > best_score: best, best_score = b, score if best is None: best = next((b for b in businesses if b["id"] == "generic_explore"), businesses[0]) inputs: dict[str, str] = {} for f in best.get("required_inputs", []): if f.get("type") == "date": d = _resolve_relative_date(message) if d: inputs[f["name"]] = d elif f["name"] in ("class_name", "class"): cleaned = re.sub(r"班级[是为:：]?", " ", message) m = re.search(r"([一-龥A-Za-z]{1,6}\d{3,4}班?)", cleaned) if m: inputs[f["name"]] = m.group(1) return {"business_id": best["id"], "confidence": 0.4 if best_score else 0.2, "inputs": inputs, "reply": ""} # ---------- 对外 ---------- async def classify_and_extract( message: str, businesses: list[dict], existing: dict | None = None, fixed_business_id: str | None = None, ) -> dict: """返回 {business_id, confidence, inputs, reply, source, error}。inputs 只含该业务声明过的字段。""" existing = existing or {} cfg = get_llm_config() result: dict | None = None source = "llm" error = "" if cfg.get("api_key"): try: result = await _call_llm(_build_prompt(message, businesses, existing, fixed_business_id), cfg) except Exception as e: # noqa: BLE001 error = f"{type(e).__name__}: {e}" if not result: result = _fallback(message, businesses, fixed_business_id) source = "fallback" ids = {b["id"] for b in businesses} bid = fixed_business_id or str(result.get("business_id") or "") if bid not in ids: bid = "generic_explore" if "generic_explore" in ids else businesses[0]["id"] business = next(b for b in businesses if b["id"] == bid) declared = {f["name"] for f in business.get("required_inputs", [])} raw_inputs = result.get("inputs") or {} inputs = { k: str(v).strip() for k, v in raw_inputs.items() if k in declared and v is not None and str(v).strip() and str(v).strip().lower() not in ("null", "none", "未提及") } try: confidence = float(result.get("confidence", 0.5)) except (TypeError, ValueError): confidence = 0.5 return { "business_id": bid, "confidence": max(0.0, min(1.0, confidence)), "inputs": inputs, "reply": str(result.get("reply") or "").strip(), "source": source, "error": error, }