File size: 8,589 Bytes
e335526 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 | """
Intent Parser - Layer 1: User Intent Understanding
This module implements the first layer of the three-layer architecture.
It uses LLM to understand user's natural language goals and converts them
to structured AnalysisIntent objects.
CRITICAL BOUNDARIES:
- ✅ LLM is used ONLY for semantic understanding
- ❌ NO numerical calculations
- ❌ NO regulatory decisions
- ❌ NO data sufficiency checks
"""
import json
import re
from typing import Dict, Any, Optional, Tuple, List
from dataclasses import asdict
from schemas.analysis_intent import (
AnalysisIntent,
AnalysisType,
AnalysisPurpose,
UserPreferences,
HardConstraints,
ExtractedDataSummary,
)
from prompts.intent_prompts import IntentPrompts
from layers.model_invoker import ModelInvoker
class IntentParser:
"""
Layer 1: User Intent Understanding.
Converts natural language analysis goals into structured AnalysisIntent.
Uses LLM purely for semantic understanding - no calculations or decisions.
"""
def __init__(self, model_invoker: Optional[ModelInvoker] = None):
"""
Initialize the intent parser.
Args:
model_invoker: LLM invoker instance. Creates new one if not provided.
"""
self.model_invoker = model_invoker or ModelInvoker()
def parse(
self,
user_goal: str,
data_summary: Optional[ExtractedDataSummary] = None
) -> AnalysisIntent:
"""
Parse user's natural language goal into structured intent.
Args:
user_goal: User's raw analysis goal text
data_summary: Optional summary of extracted data
Returns:
AnalysisIntent structure for Layer 2
"""
if data_summary is None:
data_summary = ExtractedDataSummary()
# Get prompts
system_prompt, user_prompt = IntentPrompts.get_intent_prompt(
user_goal=user_goal,
n_batches=len(data_summary.batch_ids),
n_conditions=len(data_summary.conditions),
cqa_list=", ".join(data_summary.cqa_list) if data_summary.cqa_list else "未知",
max_timepoint=max(data_summary.available_timepoints) if data_summary.available_timepoints else 0
)
# Call LLM
response = self.model_invoker.invoke(
system_prompt=system_prompt,
user_prompt=user_prompt,
temperature=0.1 # Low temperature for consistent parsing
)
if not response.success:
# Fallback to default intent
return self._create_default_intent(user_goal, data_summary)
# Parse LLM response
try:
parsed = self._extract_json(response.content)
return self._build_intent(user_goal, parsed, data_summary)
except Exception as e:
# Fallback on parse error
return self._create_default_intent(user_goal, data_summary, str(e))
def _extract_json(self, text: str) -> Dict[str, Any]:
"""Extract JSON from LLM response (handles markdown code blocks)."""
# Try to find JSON in markdown code block
json_match = re.search(r'```(?:json)?\s*([\s\S]*?)\s*```', text)
if json_match:
json_str = json_match.group(1)
else:
# Try to find raw JSON
json_str = text.strip()
# Clean up common issues
json_str = json_str.strip()
if json_str.startswith('{') and json_str.endswith('}'):
return json.loads(json_str)
raise ValueError(f"Could not extract valid JSON from: {text[:200]}")
def _build_intent(
self,
raw_goal: str,
parsed: Dict[str, Any],
data_summary: ExtractedDataSummary
) -> AnalysisIntent:
"""Build AnalysisIntent from parsed JSON."""
# Map analysis type
analysis_type_str = parsed.get("analysis_type", "trend_assessment")
try:
analysis_type = AnalysisType(analysis_type_str)
except ValueError:
analysis_type = AnalysisType.TREND_ASSESSMENT
# Map purpose
purpose_str = parsed.get("purpose", "rd_reference")
try:
purpose = AnalysisPurpose(purpose_str)
except ValueError:
purpose = AnalysisPurpose.RD_REFERENCE
# Build preferences
preferences = UserPreferences(
allow_extrapolation=parsed.get("allow_extrapolation", True),
target_timepoints=parsed.get("target_timepoints", [24, 36]),
required_confidence=parsed.get("required_confidence", 0.95)
)
# Build constraints
spec_limit = parsed.get("specification_limit")
if spec_limit is None or spec_limit <= 0:
spec_limit = 0.5 # Default
constraints = HardConstraints(
purpose=purpose,
primary_cqa=parsed.get("primary_cqa", "总杂质"),
specification_limit=spec_limit
)
# Build intent
intent = AnalysisIntent(
raw_goal=raw_goal,
analysis_type=analysis_type,
preferences=preferences,
constraints=constraints,
data_summary=data_summary,
parse_confidence=parsed.get("parse_confidence", 0.5),
ambiguities=parsed.get("ambiguities", [])
)
return intent
def _create_default_intent(
self,
raw_goal: str,
data_summary: ExtractedDataSummary,
error_msg: str = None
) -> AnalysisIntent:
"""Create a default intent when parsing fails."""
# Try simple keyword matching for analysis type
goal_lower = raw_goal.lower()
if any(kw in goal_lower for kw in ["预测", "货架期", "shelf"]):
analysis_type = AnalysisType.SHELF_LIFE_PREDICTION
elif any(kw in goal_lower for kw in ["对比", "筛选", "最优", "比较"]):
analysis_type = AnalysisType.BATCH_COMPARISON
elif any(kw in goal_lower for kw in ["风险", "超标", "合规"]):
analysis_type = AnalysisType.RISK_EVALUATION
else:
analysis_type = AnalysisType.TREND_ASSESSMENT
# Extract timepoints if mentioned
timepoints = [24, 36] # Default
tp_match = re.findall(r'(\d+)\s*[Mm月]', raw_goal)
if tp_match:
timepoints = [int(t) for t in tp_match]
ambiguities = []
if error_msg:
ambiguities.append(f"LLM解析失败: {error_msg}")
return AnalysisIntent(
raw_goal=raw_goal,
analysis_type=analysis_type,
preferences=UserPreferences(
target_timepoints=timepoints
),
constraints=HardConstraints(),
data_summary=data_summary,
parse_confidence=0.3, # Low confidence for fallback
ambiguities=ambiguities
)
def parse_with_confirmation(
self,
user_goal: str,
data_summary: Optional[ExtractedDataSummary] = None
) -> Tuple[AnalysisIntent, bool, str]:
"""
Parse intent and determine if user confirmation is needed.
Returns:
Tuple of (intent, needs_confirmation, confirmation_prompt)
"""
intent = self.parse(user_goal, data_summary)
if intent.is_high_confidence():
return intent, False, ""
# Build confirmation prompt
confirmation_parts = [
"请确认以下解析结果：",
f"- 分析类型: {intent.analysis_type.value}",
f"- 目标时间点: {intent.preferences.target_timepoints}",
f"- 主要CQA: {intent.constraints.primary_cqa}",
f"- 分析目的: {intent.constraints.purpose.value}",
]
if intent.ambiguities:
confirmation_parts.append("\n⚠️ 存在以下不确定项：")
for amb in intent.ambiguities:
confirmation_parts.append(f" - {amb}")
confirmation_prompt = "\n".join(confirmation_parts)
return intent, True, confirmation_prompt
|