flowmarg / agent /verifier.py
akash
Deploy to Hugging Face (source code only, no binary files)
2379054
Raw History Blame Contribute Delete
4.03 kB
import os
import sys
import json
from pathlib import Path
from dotenv import load_dotenv
# Force load .env from project root
env_path = Path(__file__).parent.parent / '.env'
load_dotenv(dotenv_path=env_path)
import google.generativeai as genai
from core.pydantic_models import VerificationReport
class SeniorVerifier:
def __init__(self):
self.model_name = os.environ.get("GEMINI_VERIFIER_MODEL", "gemini-3.6-flash")
self._configure_gemini()
def _configure_gemini(self):
# Reload key if not present
if not os.environ.get("GEMINI_API_KEY"):
env_path = Path(__file__).parent.parent / '.env'
load_dotenv(dotenv_path=env_path)
api_key = os.environ.get("GEMINI_API_KEY")
if not api_key:
raise ValueError("GEMINI_API_KEY not set in environment or .env")
genai.configure(api_key=api_key)
def _call_gemini(self, prompt: str) -> str:
print("⏳ Verifier calling Gemini...")
model = genai.GenerativeModel(self.model_name)
response = model.generate_content(prompt, generation_config={"temperature": 0.2})
print("✅ Verifier got response.")
return response.text.strip()
def verify(self, query: str, plan: dict, technical: dict, draft_answer: str) -> VerificationReport:
prompt = f"""
You are a PRINCIPAL TRACK DESIGN ENGINEER with 20 years of experience in UK railway engineering.
You have reviewed thousands of design reports and know exactly what can go wrong.
Your task is to verify the draft answer against the deterministic technical data provided.
You must think like a senior reviewer who:
- Knows that lift is measured in mm, not m, and that values over 50 mm are unusual for plain line tamping.
- Knows that cant is limited by track category (A/B/C/D) and that excess cant deficiency is a safety issue.
- Knows that twist is rate of change over 3m or 5m, and that it must be within limits.
- Understands that S&C has different limits from plain line.
- Will question any number that seems implausible, even if it matches the data (e.g., a lift of 1000 mm).
- Will check that the cited standard clause actually applies to the context.
USER QUERY: {query}
PLAN: {plan}
TECHNICAL DATA: {technical}
DRAFT ANSWER: {draft_answer}
Now perform the following checks and provide a corrected answer if needed.
Your output must be **ONLY** valid JSON (no markdown, no extra text) with these fields:
- "verdict": "PASS" or "FAIL_CORRECTED" or "QUESTIONABLE"
- "issues_found": list of strings (each describing a specific issue)
- "corrected_answer": the final answer, corrected if needed
- "verification_summary": a concise summary of what you checked and what you changed
If you find that the technical data itself seems implausible (e.g., a radius of 10 m for a main line), you must flag it as "QUESTIONABLE" and suggest the user verify the source data.
"""
response = self._call_gemini(prompt)
# Strip markdown code fences
import re
json_match = re.search(r'```(?:json)?\s*([\s\S]*?)\s*```', response, re.DOTALL)
if json_match:
response = json_match.group(1).strip()
try:
data = json.loads(response)
return VerificationReport(
verdict=data.get("verdict", "QUESTIONABLE"),
issues_found=data.get("issues_found", []),
corrected_answer=data.get("corrected_answer", draft_answer),
verification_summary=data.get("verification_summary", "No summary provided.")
)
except json.JSONDecodeError as e:
print(f"Verifier JSON parse error: {e}")
return VerificationReport(
verdict="QUESTIONABLE",
issues_found=["Verifier response could not be parsed."],
corrected_answer=draft_answer + "\n\n⚠️ Verification system had a hiccup – please manually confirm the data.",
verification_summary="Manual check recommended."
)