"""新闻/政策催化归因。 边界： - LLM 只负责把文本映射到题材、提炼催化类型和解释。 - 行情、资金、最终动作仍由规则引擎决定。 """ from __future__ import annotations import json import logging import re from datetime import datetime, timezone from app.analysis.theme_mapper import THEME_ALIASES, THEME_NAMES, resolve_theme from app.catalyst.models import CatalystAnalysis, CatalystInput, CatalystTheme from app.config import settings logger = logging.getLogger(__name__) CATALYST_TYPE_KEYWORDS = { "policy": ["政策", "工信部", "发改委", "国务院", "证监会", "财政部", "规划", "指导意见", "补贴"], "industry": ["订单", "需求", "涨价", "产能", "景气", "出口", "交付", "装机", "销量"], "event": ["大会", "发布会", "展会", "会议", "试点", "招标", "中标", "事故"], "earnings": ["业绩", "净利润", "营收", "预增", "扭亏", "年报", "季报"], "announcement": ["公告", "重组", "并购", "定增", "回购", "签订合同"], } STRENGTH_KEYWORDS = { 18: ["重大", "重磅", "首次", "超预期", "全面", "国家级"], 12: ["政策", "补贴", "涨价", "订单", "中标", "突破"], 8: ["试点", "规划", "发布", "扩产", "合作"], } async def analyze_catalyst(item: CatalystInput, use_llm: bool = True) -> CatalystAnalysis: """分析单条催化文本，LLM 不可用时使用规则归因。""" rule_result = _analyze_by_rules(item) if not use_llm or not settings.deepseek_api_key: return rule_result llm_result = await _analyze_by_llm(item, rule_result) return llm_result or rule_result def _analyze_by_rules(item: CatalystInput) -> CatalystAnalysis: text = f"{item.title}\n{item.content}".strip() themes = _match_themes(text) catalyst_type = _infer_catalyst_type(text) strength = _score_strength(text, themes, catalyst_type) freshness = _score_freshness(item.published_at) confidence = 45 + min(len(themes) * 12, 35) if catalyst_type in {"policy", "announcement"}: confidence += 8 return CatalystAnalysis( title=item.title, summary=_summarize_text(item.content or item.title), source=item.source, url=item.url, published_at=item.published_at, catalyst_type=catalyst_type, strength=min(strength, 100), freshness=freshness, confidence=min(confidence, 90), themes=themes, raw_text=text, generated_by="rules", ) async def _analyze_by_llm(item: CatalystInput, baseline: CatalystAnalysis) -> CatalystAnalysis | None: from app.llm.client import get_client client = get_client() if not client: return None aliases_text = "\n".join( f"- {THEME_NAMES[theme_id]}: {', '.join(aliases[:8])}" for theme_id, aliases in THEME_ALIASES.items() ) user_text = f"""\ 请把下面新闻/政策/公告归因到 A 股题材。只做语义归因，不给买卖建议。 ## 可选系统题材 {aliases_text} ## 文本 标题: {item.title} 来源: {item.source} 正文: {(item.content or '')[:1600]} 请严格输出 JSON： {{ "summary": "一句话摘要", "catalyst_type": "policy | industry | event | earnings | announcement | news", "strength": 0-100, "confidence": 0-100, "themes": [ {{"theme_name": "系统题材名或新题材名", "relevance": 0-100, "reason": "一句话"}} ], "reason": "为什么这么归因" }}""" try: response = await client.chat.completions.create( model=settings.deepseek_model, messages=[ { "role": "system", "content": ( "你是A股新闻催化归因器。" "你只能做题材归因、催化类型和强度判断，不能输出买入卖出建议。" "必须返回合法JSON。" ), }, {"role": "user", "content": user_text}, ], max_tokens=700, temperature=0.1, ) data = _extract_json(response.choices[0].message.content or "") if not data: return None themes = [] for raw_theme in data.get("themes", [])[:5]: theme_name = str(raw_theme.get("theme_name", "")).strip() if not theme_name: continue theme_id, resolved_name, _ = resolve_theme(theme_name) themes.append(CatalystTheme( theme_id=theme_id, theme_name=resolved_name, relevance=_clamp_float(raw_theme.get("relevance"), 0, 100, 60), reason=str(raw_theme.get("reason", "")).strip(), )) if not themes: themes = baseline.themes return CatalystAnalysis( title=item.title, summary=str(data.get("summary", "")).strip() or baseline.summary, source=item.source, url=item.url, published_at=item.published_at, catalyst_type=_normalize_type(data.get("catalyst_type")) or baseline.catalyst_type, strength=_clamp_float(data.get("strength"), 0, 100, baseline.strength), freshness=baseline.freshness, confidence=_clamp_float(data.get("confidence"), 0, 100, baseline.confidence), themes=themes, raw_text=baseline.raw_text, llm_reason=str(data.get("reason", "")).strip(), generated_by="llm", ) except Exception as e: logger.warning("LLM 催化归因失败: %s", e) return None def _match_themes(text: str) -> list[CatalystTheme]: clean_text = _clean(text) matched: list[CatalystTheme] = [] for theme_id, aliases in THEME_ALIASES.items(): hits = [] for alias in aliases: alias_clean = _clean(alias) if alias_clean and alias_clean in clean_text: hits.append(alias) if not hits: continue relevance = min(55 + len(hits) * 12, 95) matched.append(CatalystTheme( theme_id=theme_id, theme_name=THEME_NAMES[theme_id], relevance=relevance, reason=f"命中关键词: {'/'.join(hits[:3])}", )) matched.sort(key=lambda item: item.relevance, reverse=True) return matched[:5] def _infer_catalyst_type(text: str) -> str: for catalyst_type, keywords in CATALYST_TYPE_KEYWORDS.items(): if any(keyword in text for keyword in keywords): return catalyst_type return "news" def _score_strength(text: str, themes: list[CatalystTheme], catalyst_type: str) -> float: score = 35.0 if themes: score += min(max(theme.relevance for theme in themes) * 0.25, 25) for bonus, keywords in STRENGTH_KEYWORDS.items(): if any(keyword in text for keyword in keywords): score += bonus if catalyst_type == "policy": score += 10 elif catalyst_type == "announcement": score += 6 return round(min(score, 100), 1) def _score_freshness(published_at: datetime | None) -> float: if not published_at: return 70 dt = published_at if dt.tzinfo is None: dt = dt.replace(tzinfo=timezone.utc) hours = max((datetime.now(timezone.utc) - dt.astimezone(timezone.utc)).total_seconds() / 3600, 0) if hours <= 6: return 100 if hours <= 24: return 90 if hours <= 72: return 70 if hours <= 168: return 45 return 20 def _summarize_text(text: str) -> str: value = re.sub(r"\s+", " ", text or "").strip() return value[:120] def _clean(value: str) -> str: return re.sub(r"[\s_\-（）()【】\[\]、，。:：]+", "", value or "") def _extract_json(text: str) -> dict: text = (text or "").strip() if text.startswith("```"): text = re.sub(r"^```(?:json)?", "", text).strip() text = re.sub(r"```$", "", text).strip() try: return json.loads(text) except Exception: pass start = text.find("{") end = text.rfind("}") if start >= 0 and end > start: try: return json.loads(text[start:end + 1]) except Exception: return {} return {} def _normalize_type(value) -> str: text = str(value or "").strip().lower() return text if text in {"policy", "industry", "event", "earnings", "announcement", "news"} else "" def _clamp_float(value, minimum: float, maximum: float, default: float) -> float: try: num = float(value) except (TypeError, ValueError): return default return max(minimum, min(maximum, num))