379 lines
14 KiB
Python
379 lines
14 KiB
Python
"""LLM 候选预筛 + 逐股深度分析
|
||
|
||
先做轻量预筛，控制深度裁决成本；
|
||
再对重点股票单独调用 LLM 做深度分析。
|
||
"""
|
||
|
||
import asyncio
|
||
import json
|
||
import logging
|
||
import re
|
||
|
||
from app.config import settings
|
||
from app.db.error_logger import log_error
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
async def prefilter_single_stock(candidate: dict, market_summary: str) -> dict:
|
||
"""对单只候选股票做轻量 LLM 预筛。"""
|
||
from app.llm.prompts import STOCK_PREFILTER_PROMPT
|
||
from app.llm.strategy_config import get_prompt_content
|
||
from app.llm.client import get_client
|
||
|
||
stock_text = f"""\
|
||
股票: {candidate['name']}({candidate['ts_code']})
|
||
主题: {candidate.get('sector', '未知')}
|
||
主线主题匹配: {"是，匹配 " + candidate.get("hot_theme_name", "") if candidate.get("hot_theme_matched") else "否"}
|
||
主题别名: {", ".join(candidate.get("hot_theme_aliases", []) or ["无"])}
|
||
召回来源: {', '.join(candidate.get('recall_tags', []) or ['未标注'])}
|
||
规则参考分: {candidate.get('quant_score', 0)}/100
|
||
资金顺势分: {candidate.get('flow_momentum_score', 0)}/100
|
||
位置安全: {candidate.get('position_score', 50)}/100
|
||
当前价: {candidate.get('current_price', '未知')}
|
||
主题阶段: {candidate.get('sector_stage', '未知')}
|
||
个股角色线索: {candidate.get('stock_role_hint', '待判断')}"""
|
||
|
||
if candidate.get("kline_summary"):
|
||
stock_text += f"\n\n## 技术结构摘要\n{candidate['kline_summary']}"
|
||
|
||
if candidate.get("capital_flow_summary"):
|
||
stock_text += f"\n\n## 资金与活跃度摘要\n{candidate['capital_flow_summary']}"
|
||
|
||
if candidate.get("intraday_volume"):
|
||
stock_text += f"\n\n## 分时量能摘要\n{candidate['intraday_volume']}"
|
||
|
||
prompt = await get_prompt_content("stock_prefilter", STOCK_PREFILTER_PROMPT)
|
||
user_msg = f"{prompt}\n\n## 市场环境\n{market_summary}\n\n{stock_text}\n\n请输出 JSON。"
|
||
|
||
try:
|
||
client = get_client()
|
||
response = await client.chat.completions.create(
|
||
model=settings.deepseek_model,
|
||
messages=[
|
||
{
|
||
"role": "system",
|
||
"content": (
|
||
"你是A股候选池预审官。"
|
||
"你只负责决定资源分配优先级，不直接下最终交易结论。"
|
||
"必须返回合法JSON。"
|
||
),
|
||
},
|
||
{"role": "user", "content": user_msg},
|
||
],
|
||
max_tokens=400,
|
||
temperature=0.2,
|
||
)
|
||
content = response.choices[0].message.content.strip()
|
||
return _parse_prefilter_response(content)
|
||
except Exception as e:
|
||
logger.error(f"LLM 预筛 {candidate.get('ts_code')} 失败: {e}")
|
||
await log_error(
|
||
"llm_prefilter",
|
||
f"LLM 预筛 {candidate.get('ts_code')} 失败: {e}",
|
||
detail=f"candidate={candidate.get('ts_code')}|{candidate.get('name', '')}",
|
||
)
|
||
return {
|
||
"decision": "watch",
|
||
"confidence": 5,
|
||
"reason": "AI 预筛暂不可用，保留观察",
|
||
"focus_points": [],
|
||
}
|
||
|
||
|
||
async def analyze_single_stock(candidate: dict, market_summary: str) -> dict:
|
||
"""对单只股票做 LLM 深度分析
|
||
|
||
candidate: 包含 ts_code, name, sector, quant_score, position_score,
|
||
kline_summary, capital_flow_summary
|
||
market_summary: 市场环境摘要
|
||
|
||
返回: {
|
||
"verdict": "execute"/"watch"/"skip",
|
||
"action_plan": "可操作"/"重点关注"/"观察",
|
||
"conviction": int,
|
||
"timing": str,
|
||
"entry_price": float or None,
|
||
"target_price": float or None,
|
||
"stop_loss": float or None,
|
||
"trigger_condition": str,
|
||
"invalidation_condition": str,
|
||
"position_pct": int,
|
||
"risk_flag": str,
|
||
"analysis": str,
|
||
}
|
||
"""
|
||
from app.llm.prompts import SINGLE_STOCK_ANALYSIS_PROMPT
|
||
from app.llm.strategy_config import get_prompt_content
|
||
from app.llm.client import get_client
|
||
|
||
# 构建 prompt — 不传 signal_type，让 LLM 独立判断
|
||
stock_text = f"""\
|
||
股票: {candidate['name']}({candidate['ts_code']})
|
||
主题: {candidate.get('sector', '未知')}
|
||
主线主题匹配: {"是，匹配 " + candidate.get("hot_theme_name", "") if candidate.get("hot_theme_matched") else "否"}
|
||
规则参考分: {candidate.get('quant_score', 0)}/100
|
||
资金顺势分: {candidate.get('flow_momentum_score', 0)}/100
|
||
位置安全: {candidate.get('position_score', 50)}/100
|
||
当前价: {candidate.get('current_price', '未知')}"""
|
||
|
||
if candidate.get("kline_summary"):
|
||
stock_text += f"\n\n## 技术分析结论\n{candidate['kline_summary']}"
|
||
|
||
if candidate.get("capital_flow_summary"):
|
||
stock_text += f"\n\n## 资金流向\n{candidate['capital_flow_summary']}"
|
||
|
||
prompt = await get_prompt_content("single_stock_analysis", SINGLE_STOCK_ANALYSIS_PROMPT)
|
||
user_msg = f"{prompt}\n\n## 市场环境\n{market_summary}\n\n{stock_text}\n\n请给出你的分析。"
|
||
|
||
try:
|
||
client = get_client()
|
||
response = await client.chat.completions.create(
|
||
model=settings.deepseek_model,
|
||
messages=[
|
||
{
|
||
"role": "system",
|
||
"content": (
|
||
"你是一位A股短线交易裁决员。"
|
||
"你的任务是决定这只股票今天是否该进入推荐前列，以及应该归入可操作、重点关注还是观察。"
|
||
"若候选股没有匹配今日主线主题，除非它是全市场极强异动且触发条件清晰，否则不要给可操作或重点关注。"
|
||
"不要复述数据，不要写成长文，不要被规则参考分绑架。"
|
||
"必须返回合法JSON。"
|
||
),
|
||
},
|
||
{"role": "user", "content": user_msg},
|
||
],
|
||
max_tokens=800,
|
||
temperature=0.3,
|
||
)
|
||
content = response.choices[0].message.content.strip()
|
||
return _parse_single_response(content)
|
||
|
||
except Exception as e:
|
||
logger.error(f"LLM 分析 {candidate.get('ts_code')} 失败: {e}")
|
||
await log_error(
|
||
"llm_final",
|
||
f"LLM 分析 {candidate.get('ts_code')} 失败: {e}",
|
||
detail=f"candidate={candidate.get('ts_code')}|{candidate.get('name', '')}",
|
||
)
|
||
return {
|
||
"verdict": "watch",
|
||
"action_plan": "重点关注",
|
||
"conviction": 4,
|
||
"timing": "",
|
||
"entry_price": None,
|
||
"target_price": None,
|
||
"stop_loss": None,
|
||
"trigger_condition": "",
|
||
"invalidation_condition": "",
|
||
"position_pct": 0,
|
||
"risk_flag": "AI 裁决暂不可用",
|
||
"analysis": "AI分析暂不可用",
|
||
}
|
||
|
||
|
||
def _parse_prefilter_response(text: str) -> dict:
|
||
data = _extract_json_object(text)
|
||
if not data:
|
||
return {
|
||
"decision": "watch",
|
||
"confidence": 5,
|
||
"reason": "预筛输出不可解析，默认保留观察",
|
||
"focus_points": [],
|
||
}
|
||
|
||
decision = str(data.get("decision", "watch")).strip().lower()
|
||
if decision not in {"priority", "watch", "ignore"}:
|
||
decision = "watch"
|
||
|
||
focus_points = data.get("focus_points") or []
|
||
if not isinstance(focus_points, list):
|
||
focus_points = []
|
||
|
||
return {
|
||
"decision": decision,
|
||
"confidence": _clamp_int(data.get("confidence"), minimum=1, maximum=10, default=5),
|
||
"reason": str(data.get("reason", "")).strip() or "暂无说明",
|
||
"focus_points": [str(item).strip() for item in focus_points[:3] if str(item).strip()],
|
||
}
|
||
|
||
|
||
def _parse_single_response(text: str) -> dict:
|
||
"""解析单只股票的 LLM 返回"""
|
||
data = _extract_json_object(text)
|
||
if data:
|
||
verdict = str(data.get("verdict", "watch")).strip().lower()
|
||
if verdict not in {"execute", "watch", "skip"}:
|
||
verdict = "watch"
|
||
|
||
action_plan = str(data.get("action_plan", "")).strip()
|
||
if action_plan not in {"可操作", "重点关注", "观察"}:
|
||
action_plan = {"execute": "可操作", "watch": "重点关注", "skip": "观察"}[verdict]
|
||
|
||
conviction = _clamp_int(data.get("conviction"), minimum=1, maximum=10, default=6)
|
||
position_pct = _clamp_int(data.get("position_pct"), minimum=0, maximum=35, default=0)
|
||
|
||
return {
|
||
"verdict": verdict,
|
||
"action_plan": action_plan,
|
||
"conviction": conviction,
|
||
"timing": str(data.get("timing", "")).strip(),
|
||
"entry_price": _parse_float(data.get("entry_price")),
|
||
"target_price": _parse_float(data.get("target_price")),
|
||
"stop_loss": _parse_float(data.get("stop_loss")),
|
||
"trigger_condition": str(data.get("trigger_condition", "")).strip(),
|
||
"invalidation_condition": str(data.get("invalidation_condition", "")).strip(),
|
||
"position_pct": position_pct,
|
||
"analysis": str(data.get("analysis", "")).strip() or "暂无分析",
|
||
"risk_flag": str(data.get("risk_flag", "")).strip(),
|
||
}
|
||
|
||
# 兼容旧格式
|
||
signal = "HOLD"
|
||
signal_match = re.search(r"信号[:\s]*(BUY|HOLD|SKIP)", text)
|
||
if signal_match:
|
||
signal = signal_match.group(1)
|
||
|
||
verdict = "execute" if signal == "BUY" else "skip" if signal == "SKIP" else "watch"
|
||
strength = "中"
|
||
strength_match = re.search(r"信号强度[:\s]*(强|中|弱)", text)
|
||
if strength_match:
|
||
strength = strength_match.group(1)
|
||
conviction = {"强": 8, "中": 6, "弱": 4}.get(strength, 6)
|
||
|
||
entry_price = None
|
||
entry_match = re.search(r"买入价[:\s]*(\d+(?:\.\d+)?)", text)
|
||
if entry_match:
|
||
entry_price = float(entry_match.group(1))
|
||
target_price = None
|
||
target_match = re.search(r"止盈价[:\s]*(\d+(?:\.\d+)?)", text)
|
||
if target_match:
|
||
target_price = float(target_match.group(1))
|
||
stop_loss = None
|
||
stop_match = re.search(r"止损价[:\s]*(\d+(?:\.\d+)?)", text)
|
||
if stop_match:
|
||
stop_loss = float(stop_match.group(1))
|
||
analysis = ""
|
||
analysis_match = re.search(r"分析[:\s]*(.+)", text, re.DOTALL)
|
||
if analysis_match:
|
||
analysis = analysis_match.group(1).strip()
|
||
|
||
return {
|
||
"verdict": verdict,
|
||
"action_plan": {"execute": "可操作", "watch": "重点关注", "skip": "观察"}[verdict],
|
||
"conviction": conviction,
|
||
"timing": "",
|
||
"entry_price": entry_price,
|
||
"target_price": target_price,
|
||
"stop_loss": stop_loss,
|
||
"trigger_condition": "",
|
||
"invalidation_condition": "",
|
||
"position_pct": 20 if verdict == "execute" else 0,
|
||
"analysis": analysis or "暂无分析",
|
||
"risk_flag": "",
|
||
}
|
||
|
||
|
||
def _extract_json_object(text: str) -> dict | None:
|
||
match = re.search(r"\{[\s\S]*\}", text)
|
||
if not match:
|
||
return None
|
||
try:
|
||
parsed = json.loads(match.group(0))
|
||
return parsed if isinstance(parsed, dict) else None
|
||
except Exception:
|
||
return None
|
||
|
||
|
||
def _parse_float(value) -> float | None:
|
||
try:
|
||
if value in (None, "", 0, "0"):
|
||
return None
|
||
return float(value)
|
||
except Exception:
|
||
return None
|
||
|
||
|
||
def _clamp_int(value, minimum: int, maximum: int, default: int) -> int:
|
||
try:
|
||
parsed = int(round(float(value)))
|
||
except Exception:
|
||
return default
|
||
return max(minimum, min(maximum, parsed))
|
||
|
||
|
||
async def analyze_candidates_individually(
|
||
candidates: list[dict], market_summary: str, max_concurrent: int = 3
|
||
) -> dict[str, dict]:
|
||
"""对候选股票逐个做 LLM 分析（控制并发数）
|
||
|
||
返回: {ts_code: {"verdict", "action_plan", "conviction", "entry_price", ...}}
|
||
"""
|
||
if not settings.deepseek_api_key or not candidates:
|
||
return {}
|
||
|
||
results = {}
|
||
semaphore = asyncio.Semaphore(max_concurrent)
|
||
|
||
async def _analyze_with_semaphore(c: dict):
|
||
async with semaphore:
|
||
ts_code = c["ts_code"]
|
||
logger.info(f"LLM 分析: {c.get('name', ts_code)}")
|
||
result = await analyze_single_stock(c, market_summary)
|
||
logger.info(
|
||
f"LLM 结果: {c.get('name', ts_code)} → "
|
||
f"裁决={result['verdict']} 计划={result['action_plan']} 置信度={result['conviction']} "
|
||
f"买入={result.get('entry_price')} 止盈={result.get('target_price')} "
|
||
f"止损={result.get('stop_loss')}"
|
||
)
|
||
return ts_code, result
|
||
|
||
tasks = [_analyze_with_semaphore(c) for c in candidates]
|
||
|
||
completed = await asyncio.gather(*tasks, return_exceptions=True)
|
||
|
||
for item in completed:
|
||
if isinstance(item, Exception):
|
||
logger.error(f"LLM 分析任务异常: {item}")
|
||
continue
|
||
ts_code, result = item
|
||
results[ts_code] = result
|
||
|
||
logger.info(f"LLM 逐股分析完成: {len(results)}/{len(candidates)} 只")
|
||
return results
|
||
|
||
|
||
async def prefilter_candidates_individually(
|
||
candidates: list[dict], market_summary: str, max_concurrent: int = 6
|
||
) -> dict[str, dict]:
|
||
"""对候选股票逐个做 LLM 预筛。"""
|
||
if not settings.deepseek_api_key or not candidates:
|
||
return {}
|
||
|
||
results = {}
|
||
semaphore = asyncio.Semaphore(max_concurrent)
|
||
|
||
async def _prefilter_with_semaphore(c: dict):
|
||
async with semaphore:
|
||
ts_code = c["ts_code"]
|
||
logger.info(f"LLM 预筛: {c.get('name', ts_code)}")
|
||
result = await prefilter_single_stock(c, market_summary)
|
||
logger.info(
|
||
f"LLM 预筛结果: {c.get('name', ts_code)} → "
|
||
f"decision={result['decision']} confidence={result['confidence']}"
|
||
)
|
||
return ts_code, result
|
||
|
||
tasks = [_prefilter_with_semaphore(c) for c in candidates]
|
||
completed = await asyncio.gather(*tasks, return_exceptions=True)
|
||
|
||
for item in completed:
|
||
if isinstance(item, Exception):
|
||
logger.error(f"LLM 预筛任务异常: {item}")
|
||
continue
|
||
ts_code, result = item
|
||
results[ts_code] = result
|
||
|
||
logger.info(f"LLM 预筛完成: {len(results)}/{len(candidates)} 只")
|
||
return results
|