Files

280 lines
10 KiB
Python

"""
Clawrity — QA Agent
Evaluates Gen Agent responses for faithfulness against data context.
Uses Groq LLM at temperature 0.1 for strict, deterministic evaluation.
Returns JSON: { score, passed, issues }
Threshold from client YAML hallucination_threshold (default 0.75).
"""
import json
import logging
import re
from typing import Optional, List, Dict
import pandas as pd
from config.llm_client import get_llm_client, get_model_name, chat_with_retry
logger = logging.getLogger(__name__)
EVAL_PROMPT = """You are a strict quality assurance evaluator for business intelligence responses.
Your job: verify that the response ONLY contains claims supported by the provided data.
## Data Context (ground truth)
{data_context}
## Response to Evaluate
{response}
## Evaluation Criteria
### 1. Branch Name Validation (CRITICAL)
- Extract ALL branch/city names mentioned in the response
- Compare against the branch names in the Data Context above
- Branch/entity names listed under "Valid Entities from User Question" are VALID even if not listed in query results
- Branch/entity names listed under "Branches/entities filtered in SQL WHERE clause" are VALID even if not in result rows (e.g., if SQL has WHERE branch = 'X', then 'X' is valid context)
- If ANY branch name appears in the response but NOT in the Data Context, the valid-entities list, or the SQL WHERE clause filters, this is a HALLUCINATION
- Deduct 0.3 from score for EACH unrelated branch mentioned
### 2. Numerical Accuracy (CRITICAL)
- ALL revenue, spend, lead, conversion, and ROI figures in the response must match the Data Context EXACTLY
- If a number is mentioned that does not appear in the Data Context, deduct 0.2 from score
- Rounded numbers are acceptable only if clearly approximate (e.g., "~$1.2M")
### 3. Historical Context Relevance
- If the response includes historical context or trends, it is acceptable ONLY if it directly supports the answer about branches/entities present in the Data Context
- Historical context about branches NOT in the current Data Context must be penalized: deduct 0.3 from score
- Example: If Data Context shows Toronto, Vancouver, Dubai but response mentions "Lawton showed 16436% growth" — this is IRRELEVANT historical context and must be penalized
### 4. Completeness
- Does the response address the user's question?
- Are key data points from the Data Context included?
### 5. Appropriate Hedging
- Does the response use uncertain language for inferences?
- Recommendations should be clearly marked as suggestions, not facts
## Scoring
Start at 1.0 and deduct points per the rules above. Minimum score is 0.0.
Return a JSON object with exactly this structure:
{{
"score": <float between 0.0 and 1.0>,
"passed": <true if score >= {threshold}>,
"issues": [<list of specific issues found, empty if none>]
}}
IMPORTANT: If score < {threshold}, include in issues list exactly which branches, figures, or historical data were mentioned that do NOT appear in the Data Context. Format as:
"Mentioned branches/figures not in current query result: [list them]"
Return ONLY the JSON. No other text."""
class QAAgent:
"""Quality assurance agent for validating Gen Agent responses."""
def __init__(self):
self.client = get_llm_client()
self.model = get_model_name()
def evaluate(
self,
response: str,
data_context: Optional[pd.DataFrame] = None,
threshold: float = 0.75,
supplementary_context: Optional[pd.DataFrame] = None,
user_question: str = "",
sql: Optional[str] = None,
) -> Dict:
"""
Evaluate a response for faithfulness.
Args:
response: Gen Agent's response text
data_context: The data the response should be grounded in
threshold: Minimum score to pass (from client YAML)
supplementary_context: Benchmark data (top performers) that is also valid ground truth
user_question: The user's original question (entities mentioned here are valid context)
sql: The SQL query that produced the data context (branch/entity filters are valid context)
Returns:
Dict with score (float), passed (bool), issues (list[str])
"""
data_str = ""
if data_context is not None and len(data_context) > 0:
data_str = data_context.to_markdown(index=False)
else:
data_str = "No structured data available."
# Include the SQL query so QA understands what filters were applied
# (e.g., branch names in WHERE clause are valid context even if not in result rows)
if sql:
data_str += (
f"\n\n### SQL Query (defines the data scope)\n```sql\n{sql}\n```"
)
# Extract branch/entity filters from SQL WHERE clause
where_branches = self._extract_where_entities(sql)
if where_branches:
data_str += (
f"\nBranches/entities filtered in SQL WHERE clause (VALID context): "
f"{', '.join(sorted(where_branches))}"
)
# Include supplementary (benchmark) context as valid ground truth
if supplementary_context is not None and len(supplementary_context) > 0:
data_str += "\n\n### Benchmark Data (also valid ground truth)\n"
data_str += supplementary_context.to_markdown(index=False)
# Include user question so QA knows which entities are valid context
if user_question:
entities = self._extract_entities(user_question)
if entities:
entity_list = ", ".join(sorted(entities))
else:
entity_list = "(none)"
data_str += (
"\n\n### User Question Context\n"
f'The user asked: "{user_question}"\n'
f"Valid Entities from User Question: {entity_list}"
)
prompt = EVAL_PROMPT.format(
data_context=data_str,
response=response,
threshold=threshold,
)
try:
result = chat_with_retry(
self.client,
model=self.model,
messages=[
{
"role": "system",
"content": "You are a strict QA evaluator. Return only valid JSON. Pay special attention to branch names and figures that appear in the response but NOT in the data context — these are hallucinations.",
},
{"role": "user", "content": prompt},
],
temperature=0.1,
max_tokens=512,
)
raw = result.choices[0].message.content.strip()
evaluation = self._parse_response(raw, threshold)
logger.info(
f"QA evaluation: score={evaluation['score']:.2f}, "
f"passed={evaluation['passed']}, issues={len(evaluation['issues'])}"
)
return evaluation
except Exception as e:
logger.error(f"QA evaluation failed: {e}")
# On failure, pass with warning
return {
"score": 0.5,
"passed": True,
"issues": [f"QA evaluation error: {str(e)}"],
}
def _parse_response(self, raw: str, threshold: float) -> Dict:
"""Parse JSON response from QA LLM call."""
try:
# Strip markdown code fences if present
cleaned = raw.strip()
if cleaned.startswith("```"):
cleaned = cleaned.split("\n", 1)[1] if "\n" in cleaned else cleaned[3:]
if cleaned.endswith("```"):
cleaned = cleaned[:-3]
cleaned = cleaned.strip()
data = json.loads(cleaned)
score = float(data.get("score", 0.5))
return {
"score": score,
"passed": score >= threshold,
"issues": data.get("issues", []),
}
except (json.JSONDecodeError, ValueError) as e:
logger.warning(f"Could not parse QA response: {e}. Raw: {raw[:200]}")
return {
"score": 0.5,
"passed": True,
"issues": ["QA response parsing failed"],
}
def _extract_where_entities(self, sql: str) -> List[str]:
"""Extract branch/city entity names from SQL WHERE clause filters."""
if not sql:
return []
entities = set()
# Match patterns like: branch = 'Seattle', city = 'Toronto'
for match in re.finditer(
r"(?:branch|city|country)\s*=\s*'([^']+)'",
sql,
re.IGNORECASE,
):
val = match.group(1).strip()
if val and len(val) > 1:
entities.add(val)
# Also handle IN ('val1', 'val2') patterns
for match in re.finditer(
r"(?:branch|city|country)\s+IN\s*\(([^)]+)\)",
sql,
re.IGNORECASE,
):
for val in re.findall(r"'([^']+)'", match.group(1)):
if val and len(val) > 1:
entities.add(val)
return list(entities)
def _extract_entities(self, text: str) -> List[str]:
"""Extract likely branch/city entities from a user question."""
if not text:
return []
lowered = text.lower()
patterns = [
r"\bbranch\s+([a-z][a-z\s\-']{1,60})",
r"\bin\s+([a-z][a-z\s\-']{1,60})",
r"\bfor\s+the\s+([a-z][a-z\s\-']{1,60})\s+branch",
]
stops = {
"the",
"a",
"an",
"my",
"our",
"this",
"that",
"these",
"those",
"branch",
"branches",
"revenue",
"sales",
"roi",
"profit",
"performance",
}
entities = set()
for pattern in patterns:
for match in re.findall(pattern, lowered):
candidate = match.strip(" .,!?:;\"'")
candidate = " ".join(candidate.split())
if not candidate:
continue
if candidate in stops:
continue
if any(word in stops for word in candidate.split()):
candidate = " ".join(w for w in candidate.split() if w not in stops)
candidate = candidate.strip()
if len(candidate) < 2:
continue
entities.add(candidate.title())
return list(entities)