x
This commit is contained in:
@@ -0,0 +1,334 @@
|
||||
"""
|
||||
LLM 增强分析服务
|
||||
|
||||
提供两个核心能力：
|
||||
1. generate_design_report — 将分析 JSON 转换为结构化评审报告
|
||||
2. recommend_parting_direction — 基于几何 + 制造约束推荐最优分型方向
|
||||
|
||||
适配层：OpenAI 兼容 API（支持 OpenAI / DeepSeek / vLLM / Ollama 等）
|
||||
未配置 LLM 时静默降级，不影响主流程。
|
||||
"""
|
||||
|
||||
import json
|
||||
import re
|
||||
from typing import Optional, Dict, Any, List
|
||||
|
||||
import httpx
|
||||
|
||||
from config.settings import settings
|
||||
from utils.logger import get_logger
|
||||
|
||||
logger = get_logger(__name__)
|
||||
|
||||
|
||||
_DESIGN_REPORT_SYSTEM = """你是一位资深注塑模具设计工程师，拥有 20 年模具 DFM 评审经验。
|
||||
请根据提供的模具分析数据，生成一份专业的模具设计评审报告。
|
||||
|
||||
要求：
|
||||
1. 使用中文
|
||||
2. 按 "问题摘要 → 关键风险 → 分模方案推荐 → 制造可行性 → 修改建议" 结构组织
|
||||
3. 技术术语准确（如：锁模力、投影面积、分型面、滑块、斜顶、拔模角、缩痕、熔接痕）
|
||||
4. 每个建议标注优先级（高/中/低）和预计工时
|
||||
5. 报告末尾给出一个总体评分（1-10分）
|
||||
6. 如果数据不足以判断某项，明确标注"数据不足，需人工确认"
|
||||
|
||||
直接输出 Markdown 格式报告，不要输出 JSON。"""
|
||||
|
||||
_DESIGN_REPORT_USER = """请根据以下模具分析数据生成评审报告：
|
||||
|
||||
## 产品信息
|
||||
- 文件：{filename}
|
||||
- 材料：{material}
|
||||
- 体积：{volume}
|
||||
- 表面积：{surface_area}
|
||||
- 边界框：{bbox}
|
||||
|
||||
## 检测特征
|
||||
{features}
|
||||
|
||||
## 质量指标
|
||||
{quality_metrics}
|
||||
|
||||
## 分模方案
|
||||
{schemes}
|
||||
|
||||
## 制造参数
|
||||
- 推荐模具材料：{mold_material}
|
||||
- 推荐模具硬度：{mold_hardness}
|
||||
- 预估锁模力：{clamping_force}
|
||||
- 模具尺寸（长×宽×高）：{mold_size}
|
||||
- 预估成型周期：{cycle_time}
|
||||
- 拔模角：{draft_angle}
|
||||
- 收缩率：{shrinkage_rate}
|
||||
|
||||
## 原始设计建议
|
||||
{recommendations}"""
|
||||
|
||||
_PARTING_SYSTEM = """你是一位注塑模具分模专家。
|
||||
根据产品几何特征和多个候选分模方向的评分数据，推荐最优分模方向。
|
||||
|
||||
输出要求：严格输出 JSON，不要输出其他内容。
|
||||
JSON 格式：
|
||||
{
|
||||
"recommended_axis": "Z",
|
||||
"confidence": 0.85,
|
||||
"reasoning": "详细的中文推理过程...",
|
||||
"risk_notes": ["风险1", "风险2"],
|
||||
"rankings": [{"axis":"Z","rank":1,"score":92,"note":"..."},{"axis":"X","rank":2,"score":78,"note":"..."}]
|
||||
}"""
|
||||
|
||||
_PARTING_USER = """请评估以下候选分模方向并推荐最优方案：
|
||||
|
||||
产品几何：
|
||||
- 边界框 (mm)：{bbox}
|
||||
- 面法向分布：{normal_stats}
|
||||
- 惯性矩：{inertia}
|
||||
|
||||
约束条件：
|
||||
- 材料：{material}
|
||||
- 型腔数：{cavity_count}
|
||||
- 最大锁模力 (吨)：{max_clamping_force}
|
||||
- 泡沫材料：{is_foam}
|
||||
|
||||
候选方案：
|
||||
{schemes}
|
||||
|
||||
请综合评估制造可行性、成本和风险，给出推荐。"""
|
||||
|
||||
|
||||
class LLMService:
|
||||
"""LLM 增强分析服务（单例）"""
|
||||
|
||||
def __init__(self):
|
||||
self._enabled = settings.LLM_ENABLED
|
||||
self._api_url = settings.LLM_API_URL.rstrip("/")
|
||||
self._api_key = settings.LLM_API_KEY
|
||||
self._model = settings.LLM_MODEL
|
||||
self._timeout = settings.LLM_TIMEOUT
|
||||
self._max_tokens = settings.LLM_MAX_TOKENS
|
||||
|
||||
if self._enabled:
|
||||
logger.info(
|
||||
"LLM 增强分析已启用: model=%s endpoint=%s",
|
||||
self._model, self._api_url,
|
||||
)
|
||||
else:
|
||||
logger.info("LLM 增强分析未启用（设置 LLM_ENABLED=true 启用）")
|
||||
|
||||
async def generate_design_report(
|
||||
self,
|
||||
analysis_result: Dict[str, Any],
|
||||
detailed_cavity_json: Optional[Dict[str, Any]] = None,
|
||||
) -> Optional[str]:
|
||||
"""生成模具设计评审报告 (Markdown)"""
|
||||
if not self._enabled:
|
||||
return None
|
||||
|
||||
try:
|
||||
prompt = self._build_design_report_prompt(analysis_result, detailed_cavity_json)
|
||||
response = await self._chat(
|
||||
system=_DESIGN_REPORT_SYSTEM,
|
||||
user=prompt,
|
||||
max_tokens=self._max_tokens,
|
||||
)
|
||||
if response:
|
||||
logger.info("LLM 设计报告生成成功 (%d 字符)", len(response))
|
||||
return response
|
||||
except Exception as e:
|
||||
logger.warning("LLM 设计报告生成失败（不影响主流程）: %s", e)
|
||||
return None
|
||||
|
||||
async def recommend_parting_direction(
|
||||
self,
|
||||
geometry_data: Dict[str, Any],
|
||||
candidate_schemes: List[Dict[str, Any]],
|
||||
material: Dict[str, Any],
|
||||
cavity_count: int = 1,
|
||||
) -> Optional[Dict[str, Any]]:
|
||||
"""推荐最优分型方向"""
|
||||
if not self._enabled:
|
||||
return None
|
||||
|
||||
try:
|
||||
prompt = self._build_parting_prompt(
|
||||
geometry_data, candidate_schemes, material, cavity_count,
|
||||
)
|
||||
response = await self._chat(
|
||||
system=_PARTING_SYSTEM,
|
||||
user=prompt,
|
||||
max_tokens=min(self._max_tokens, 1200),
|
||||
expect_json=True,
|
||||
)
|
||||
if response:
|
||||
result = self._parse_json_response(response)
|
||||
if result:
|
||||
logger.info(
|
||||
"LLM 分型推荐: %s (置信度 %.2f)",
|
||||
result.get("recommended_axis", "?"),
|
||||
result.get("confidence", 0),
|
||||
)
|
||||
return result
|
||||
return None
|
||||
except Exception as e:
|
||||
logger.warning("LLM 分型推荐失败（不影响主流程）: %s", e)
|
||||
return None
|
||||
|
||||
def _build_design_report_prompt(
|
||||
self,
|
||||
analysis_result: Dict[str, Any],
|
||||
detailed_cavity_json: Optional[Dict[str, Any]],
|
||||
) -> str:
|
||||
detected_features = analysis_result.get("detected_features", [])
|
||||
quality_metrics = analysis_result.get("quality_metrics", {})
|
||||
recommendations = analysis_result.get("design_recommendations", [])
|
||||
|
||||
feature_text = json.dumps(detected_features, ensure_ascii=False, indent=2)
|
||||
if len(feature_text) > 4000:
|
||||
feature_text = feature_text[:4000] + "\n... (已截断)"
|
||||
|
||||
schemes_text = ""
|
||||
if detailed_cavity_json:
|
||||
schemes = detailed_cavity_json.get("candidate_schemes", [])
|
||||
if schemes:
|
||||
schemes_text = json.dumps(
|
||||
[
|
||||
{
|
||||
"scheme_id": s.get("scheme_id"),
|
||||
"rank": s.get("rank"),
|
||||
"title": s.get("title"),
|
||||
"score": s.get("score"),
|
||||
"confidence_score": s.get("confidence_score"),
|
||||
"summary": s.get("summary"),
|
||||
"parting_axis": s.get("parting", {}).get("axis"),
|
||||
"mold_structure_type": s.get("mold_structure_type"),
|
||||
"dfm_violations": s.get("dfm_violations", []),
|
||||
}
|
||||
for s in schemes
|
||||
],
|
||||
ensure_ascii=False,
|
||||
indent=2,
|
||||
)
|
||||
|
||||
best_scheme = (
|
||||
detailed_cavity_json.get("candidate_schemes", [{}])[0]
|
||||
if detailed_cavity_json
|
||||
else {}
|
||||
)
|
||||
cavity_data = best_scheme.get("cavity_data", {}) if isinstance(best_scheme, dict) else {}
|
||||
mfg_info = cavity_data.get("manufacturing_info", {})
|
||||
metadata = cavity_data.get("metadata", {})
|
||||
|
||||
return _DESIGN_REPORT_USER.format(
|
||||
filename=metadata.get("file_name", "unknown.stp"),
|
||||
material=metadata.get("selected_material", "ABS"),
|
||||
volume=f"{analysis_result.get('geometry_data', {}).get('volume', 0):.1f} mm³",
|
||||
surface_area=f"{analysis_result.get('geometry_data', {}).get('surface_area', 0):.1f} mm²",
|
||||
bbox=json.dumps(analysis_result.get("geometry_data", {}).get("bounding_box", {}), ensure_ascii=False),
|
||||
features=feature_text or "无特征检测数据",
|
||||
quality_metrics=json.dumps(quality_metrics, ensure_ascii=False, indent=2),
|
||||
schemes=schemes_text or "无分模方案数据",
|
||||
mold_material=mfg_info.get("mold_material", "自动选择"),
|
||||
mold_hardness=mfg_info.get("mold_hardness", "自动选择"),
|
||||
clamping_force=mfg_info.get("estimated_clamping_force", "自动计算"),
|
||||
mold_size=json.dumps(mfg_info.get("estimated_mold_size", {}), ensure_ascii=False),
|
||||
cycle_time=mfg_info.get("estimated_cycle_time", "自动计算"),
|
||||
draft_angle=f"{metadata.get('draft_angle', 2.0)}°",
|
||||
shrinkage_rate=f"{metadata.get('shrinkage_rate', 0.0) * 100:.2f}%"
|
||||
if isinstance(metadata.get("shrinkage_rate"), (int, float))
|
||||
else "自动计算",
|
||||
recommendations=json.dumps(recommendations, ensure_ascii=False, indent=2) if recommendations else "无",
|
||||
)
|
||||
|
||||
def _build_parting_prompt(
|
||||
self,
|
||||
geometry_data: Dict[str, Any],
|
||||
candidate_schemes: List[Dict[str, Any]],
|
||||
material: Dict[str, Any],
|
||||
cavity_count: int,
|
||||
) -> str:
|
||||
bbox = geometry_data.get("bounding_box", {})
|
||||
axis_normal_stats = geometry_data.get("axis_normal_stats", {})
|
||||
inertia = geometry_data.get("inertia_matrix", [])
|
||||
inertia_diag = [
|
||||
inertia[i][i] if i < len(inertia) and i < len(inertia[i]) else 0.0
|
||||
for i in range(3)
|
||||
]
|
||||
|
||||
schemes_text = json.dumps(
|
||||
[
|
||||
{
|
||||
"axis": s.get("parting", {}).get("axis") or s.get("axis"),
|
||||
"score": s.get("score"),
|
||||
"confidence_score": s.get("confidence_score"),
|
||||
"summary": s.get("summary"),
|
||||
"mold_structure_type": s.get("mold_structure_type"),
|
||||
"core_required": s.get("core_required"),
|
||||
"dfm_violations": s.get("dfm_violations", []),
|
||||
"undercut_regions_count": len(s.get("undercut_regions", [])),
|
||||
"score_breakdown": s.get("score_breakdown", {}),
|
||||
}
|
||||
for s in candidate_schemes
|
||||
],
|
||||
ensure_ascii=False,
|
||||
indent=2,
|
||||
)
|
||||
|
||||
return _PARTING_USER.format(
|
||||
bbox=json.dumps(bbox, ensure_ascii=False),
|
||||
normal_stats=json.dumps(axis_normal_stats, ensure_ascii=False),
|
||||
inertia=json.dumps(inertia_diag, ensure_ascii=False),
|
||||
material=material.get("name", "ABS"),
|
||||
cavity_count=cavity_count,
|
||||
max_clamping_force="3000 吨（最大）",
|
||||
is_foam="是" if material.get("is_foam") else "否",
|
||||
schemes=schemes_text,
|
||||
)
|
||||
|
||||
async def _chat(
|
||||
self,
|
||||
system: str,
|
||||
user: str,
|
||||
max_tokens: int = 2000,
|
||||
expect_json: bool = False,
|
||||
temperature: float = 0.3,
|
||||
) -> Optional[str]:
|
||||
url = f"{self._api_url}/chat/completions"
|
||||
headers = {
|
||||
"Authorization": f"Bearer {self._api_key}",
|
||||
"Content-Type": "application/json",
|
||||
}
|
||||
payload = {
|
||||
"model": self._model,
|
||||
"messages": [
|
||||
{"role": "system", "content": system},
|
||||
{"role": "user", "content": user},
|
||||
],
|
||||
"max_tokens": max_tokens,
|
||||
"temperature": temperature,
|
||||
}
|
||||
if expect_json:
|
||||
payload["response_format"] = {"type": "json_object"}
|
||||
|
||||
async with httpx.AsyncClient(timeout=self._timeout) as client:
|
||||
resp = await client.post(url, json=payload, headers=headers)
|
||||
resp.raise_for_status()
|
||||
data = resp.json()
|
||||
content = data["choices"][0]["message"]["content"]
|
||||
return content.strip() if content else None
|
||||
|
||||
@staticmethod
|
||||
def _parse_json_response(raw: str) -> Optional[Dict[str, Any]]:
|
||||
try:
|
||||
return json.loads(raw)
|
||||
except json.JSONDecodeError:
|
||||
match = re.search(r"\{[\s\S]*\}", raw)
|
||||
if match:
|
||||
try:
|
||||
return json.loads(match.group())
|
||||
except json.JSONDecodeError:
|
||||
pass
|
||||
logger.warning("LLM JSON 解析失败: %s...", raw[:200])
|
||||
return None
|
||||
|
||||
|
||||
llm_service = LLMService()
|
||||
Reference in New Issue
Block a user