occ-stack / benchmarks /benchmark_retrieval_qa.py
narcolepticchicken's picture
Upload benchmarks/benchmark_retrieval_qa.py
b4d00e5 verified
"""
Benchmark 2: Retrieval QA / Legal-Factual QA
Compares:
A. direct answer
B. RAG baseline
C. RAG + verifier
D. RAG + abstention rule
E. OCC resource allocation
F. OCC + verifier + abstention reward
Uses synthetic grounded QA with adversarial evidence.
"""
import json
import random
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
import numpy as np
import sys
sys.path.insert(0, str(Path(__file__).parent.parent))
from oracle.oracle import ImpactOracle, OracleResult
from ledger.ledger import CreditLedger
from broker.broker import ResourceBroker, Decision
from rl.reward import RewardHook
@dataclass
class Question:
question: str
answer: Optional[str] # None = unanswerable
evidence: List[str]
adversarial: List[str] # misleading evidence
is_unanswerable: bool = False
class SimulatedRetrievalAgent:
"""
Simulates a RAG agent with configurable accuracy, hallucination, and calibration.
"""
def __init__(
self,
agent_id: str,
accuracy: float = 0.6,
hallucination_rate: float = 0.15,
calibration_error: float = 0.2, # ECE-like
abstention_rate: float = 0.1,
cost_per_retrieval: float = 10.0,
cost_per_answer: float = 5.0,
gaming_mode: bool = False,
):
self.agent_id = agent_id
self.accuracy = accuracy
self.hallucination_rate = hallucination_rate
self.calibration_error = calibration_error
self.abstention_rate = abstention_rate
self.cost_per_retrieval = cost_per_retrieval
self.cost_per_answer = cost_per_answer
self.gaming_mode = gaming_mode
self.retrieval_calls = 0
self.answers_given = 0
def answer(
self,
question: Question,
oracle: ImpactOracle,
max_retrievals: int = 3,
use_occ: bool = False,
broker: Optional[ResourceBroker] = None,
ledger: Optional[CreditLedger] = None,
) -> Dict:
"""Answer a question, optionally with OCC-managed retrievals."""
retrieved = []
compute_cost = 0.0
# Retrieve evidence
for i in range(max_retrievals):
if use_occ and broker and ledger:
balance = ledger.balance(self.agent_id, "retrieval", "global")
dec = broker.request(
"retrieval_call",
self.agent_id,
balance,
task_state={"progress": len(retrieved) / max_retrievals},
)
if dec.decision == Decision.DENY:
break
self.retrieval_calls += 1
compute_cost += self.cost_per_retrieval
# Mix genuine and adversarial evidence
if i == 0:
retrieved.extend(question.evidence)
else:
if random.random() < 0.3:
retrieved.extend(question.adversarial)
else:
retrieved.extend(question.evidence)
# OCC: smart stopping — if we already have good evidence, stop retrieving
if use_occ and i >= 1:
has_strong_evidence = any(
"legal text" in ev or "According to" in ev for ev in retrieved
)
has_contradiction = any(
"unknown" in ev or "blog" in ev for ev in retrieved
)
# If strong evidence and no contradiction, stop early (save compute)
if has_strong_evidence and not has_contradiction:
break
# If too much adversarial evidence, stop to avoid confusion
if has_contradiction and i >= 1:
break
# If broker denied after first retrieval, stop
if use_occ and broker and ledger:
balance = ledger.balance(self.agent_id, "retrieval", "global")
dec = broker.request(
"retrieval_call",
self.agent_id,
balance,
task_state={"progress": len(retrieved) / max_retrievals},
)
if dec.decision == Decision.DENY:
break
# Decide whether to abstain
abstained = False
if question.is_unanswerable:
abstained = random.random() < (self.abstention_rate + 0.3)
else:
abstained = random.random() < self.abstention_rate
if abstained:
self.answers_given += 1
compute_cost += self.cost_per_answer
confidence = 0.5 + random.uniform(-self.calibration_error, self.calibration_error)
confidence = max(0.0, min(1.0, confidence))
# Evidence NLI simulation
evidence = {
"entailment_score": 0.0,
"contradiction_score": 0.0,
}
oracle_res = oracle.score(
mode="retrieval_qa",
action={"abstained": True},
context={"gold_answer": question.answer},
result={
"answer": None,
"confidence": confidence,
"evidence": evidence,
"compute_cost": compute_cost,
},
agent_id=self.agent_id,
)
return {
"answer": None,
"abstained": True,
"correct": question.is_unanswerable,
"confidence": confidence,
"oracle_score": oracle_res.raw_score,
"reward": oracle_res.reward_value,
"compute_cost": compute_cost,
"retrieval_calls": len(retrieved),
}
# Generate answer
self.answers_given += 1
compute_cost += self.cost_per_answer
if question.is_unanswerable:
# Should have abstained
correct = False
answer_text = self._generate_fake_answer(question)
else:
# Evidence-quality-aware accuracy
base_accuracy = self.accuracy
strong_evidence = any("legal text" in ev or "According to" in ev for ev in retrieved)
adversarial_evidence = any("unknown" in ev or "blog" in ev for ev in retrieved)
if strong_evidence and not adversarial_evidence:
effective_accuracy = min(0.95, base_accuracy + 0.25)
elif adversarial_evidence:
effective_accuracy = max(0.3, base_accuracy - 0.15)
else:
effective_accuracy = base_accuracy
correct = random.random() < effective_accuracy
if not correct and random.random() < self.hallucination_rate:
answer_text = self._generate_hallucinated_answer(question)
correct = False
else:
answer_text = question.answer if correct else self._generate_wrong_answer(question)
confidence = self._calibrate_confidence(correct)
# Evidence NLI simulation
if correct:
entailment = 0.8 + random.random() * 0.2
contradiction = 0.0
else:
if random.random() < self.hallucination_rate:
entailment = 0.2
contradiction = 0.7 + random.random() * 0.3
else:
entailment = 0.4
contradiction = 0.1
evidence = {
"entailment_score": entailment,
"contradiction_score": contradiction,
}
oracle_res = oracle.score(
mode="retrieval_qa",
action={"abstained": False},
context={"gold_answer": question.answer},
result={
"answer": answer_text,
"confidence": confidence,
"evidence": evidence,
"compute_cost": compute_cost,
},
agent_id=self.agent_id,
)
return {
"answer": answer_text,
"abstained": False,
"correct": correct,
"confidence": confidence,
"oracle_score": oracle_res.raw_score,
"reward": oracle_res.reward_value,
"compute_cost": compute_cost,
"retrieval_calls": len(retrieved),
"hallucination": contradiction > 0.5,
}
def _calibrate_confidence(self, correct: bool) -> float:
"""Generate confidence with controlled miscalibration."""
if correct:
base = 0.8 + random.random() * 0.2
else:
base = 0.3 + random.random() * 0.5
# Inject calibration error
error = random.uniform(-self.calibration_error, self.calibration_error)
return max(0.0, min(1.0, base + error))
def _generate_fake_answer(self, question: Question) -> str:
return f"I cannot answer based on the available evidence."
def _generate_hallucinated_answer(self, question: Question) -> str:
return f"The answer is {question.answer} according to source X." if question.answer else "Unknown."
def _generate_wrong_answer(self, question: Question) -> str:
return "42" # generic wrong
class RetrievalQABenchmark:
"""
Benchmark retrieval QA with abstention and calibration under budgets.
"""
def __init__(
self,
n_questions: int = 100,
unanswerable_ratio: float = 0.2,
adversarial_ratio: float = 0.3,
seed: int = 42,
):
self.n_questions = n_questions
self.unanswerable_ratio = unanswerable_ratio
self.adversarial_ratio = adversarial_ratio
self.seed = seed
self.questions: List[Question] = []
self.oracle = ImpactOracle(compute_budget=1e4)
def generate_questions(self):
random.seed(self.seed)
np.random.seed(self.seed)
topics = [
("What is the statute of limitations for contract disputes?", "6 years"),
("Who authored the Copyright Act of 1976?", "United States Congress"),
("What is the maximum penalty under GDPR Article 83?", "20 million EUR"),
("Which amendment protects against unreasonable search and seizure?", "Fourth Amendment"),
("What is the burden of proof in criminal cases?", "beyond reasonable doubt"),
("What is the definition of negligence?", "breach of duty causing harm"),
("When was the Paris Agreement signed?", "2015"),
("What is the legal drinking age in the US?", "21"),
("Which court handles patent appeals?", "Federal Circuit"),
("What is the Dodd-Frank Act primarily about?", "financial regulation"),
]
for i in range(self.n_questions):
if i < int(self.n_questions * self.unanswerable_ratio):
q = Question(
question=f"Unanswerable question {i}: What is the secret code of Atlantis?",
answer=None,
evidence=["No reliable source mentions Atlantis codes."],
adversarial=["Some blogs claim Atlantis code is 1234."],
is_unanswerable=True,
)
else:
topic = topics[i % len(topics)]
has_adv = random.random() < self.adversarial_ratio
q = Question(
question=topic[0],
answer=topic[1],
evidence=[f"According to legal text X, {topic[1]}."],
adversarial=[f"Some sources claim the answer is 'unknown' for {topic[0]}."] if has_adv else [],
is_unanswerable=False,
)
self.questions.append(q)
def run_direct_answer(self, agent: SimulatedRetrievalAgent) -> Dict:
"""Baseline A: direct answer, no retrieval."""
results = []
for q in self.questions:
# Force 0 retrievals
agent.retrieval_calls = 0
r = agent.answer(q, self.oracle, max_retrievals=0)
results.append(r)
return self._summarize(results, "direct_answer")
def run_rag_baseline(self, agent: SimulatedRetrievalAgent) -> Dict:
"""Baseline B: RAG with fixed retrievals."""
results = []
for q in self.questions:
r = agent.answer(q, self.oracle, max_retrievals=2, use_occ=False)
results.append(r)
return self._summarize(results, "rag_baseline")
def run_rag_verifier(self, agent: SimulatedRetrievalAgent) -> Dict:
"""Baseline C: RAG + verifier (extra check)."""
results = []
for q in self.questions:
r = agent.answer(q, self.oracle, max_retrievals=2, use_occ=False)
# Simulate verifier: if hallucination detected, retry once
if r.get("hallucination", False):
r2 = agent.answer(q, self.oracle, max_retrievals=1, use_occ=False)
r2["compute_cost"] += r["compute_cost"]
r2["retrieval_calls"] += r["retrieval_calls"]
r = r2
results.append(r)
return self._summarize(results, "rag_verifier")
def run_occ(self, agent: SimulatedRetrievalAgent) -> Dict:
"""Baseline E/F: OCC resource allocation for retrievals."""
ledger = CreditLedger(decay_lambda=0.05)
broker = ResourceBroker()
results = []
# Seed initial trial credits for the agent
ledger.earn(
agent_id=agent.agent_id,
task_id="seed",
action_id="seed",
amount=10.0,
oracle_score=0.0,
compute_cost=0.0,
reason="initial_trial_credit",
capability_scope="retrieval",
)
for q in self.questions:
r = agent.answer(q, self.oracle, max_retrievals=5, use_occ=True, broker=broker, ledger=ledger)
# Update ledger based on outcome
earn_amount = max(0.0, r["reward"] * 3.0)
if earn_amount > 0:
ledger.earn(
agent_id=agent.agent_id,
task_id=f"q_{q.question[:30]}",
action_id="answer",
amount=earn_amount,
oracle_score=r["oracle_score"],
compute_cost=r["compute_cost"],
reason="correct_answer",
capability_scope="retrieval",
)
else:
# Penalty for wrong / low-reward answers (capped so we don't over-spend)
bal = ledger.balance(agent.agent_id, "retrieval", "global")
penalty = min(bal, max(0.5, abs(r["reward"])))
if penalty > 0:
ledger.spend(
agent_id=agent.agent_id,
task_id=f"q_{q.question[:30]}",
action_id="answer",
amount=penalty,
capability_scope="retrieval",
reason="wrong_answer_penalty",
)
results.append(r)
return self._summarize(results, "occ_allocation")
def _summarize(self, results: List[Dict], label: str) -> Dict:
n = len(results)
correct = sum(1 for r in results if r["correct"])
abstained = sum(1 for r in results if r.get("abstained", False))
# Count abstentions properly
unanswerable_qs = [i for i, r in enumerate(results) if self.questions[i].is_unanswerable]
correct_abstentions = sum(
1 for i in unanswerable_qs if results[i].get("abstained", False)
)
wrong_abstentions = sum(
1 for i, r in enumerate(results)
if not self.questions[i].is_unanswerable and r.get("abstained", False)
)
hallucinations = sum(1 for r in results if r.get("hallucination", False))
confidences = [r["confidence"] for r in results]
correct_flags = [r["correct"] for r in results]
# ECE approximation
ece = self.oracle.compute_ece(confidences, correct_flags, n_bins=5)
total_compute = sum(r["compute_cost"] for r in results)
total_retrievals = sum(r["retrieval_calls"] for r in results)
return {
"label": label,
"n": n,
"accuracy": correct / n if n else 0.0,
"abstention_rate": abstained / n if n else 0.0,
"correct_abstentions": correct_abstentions,
"wrong_abstentions": wrong_abstentions,
"hallucination_rate": hallucinations / n if n else 0.0,
"confident_wrong_rate": sum(
1 for r in results if not r["correct"] and r["confidence"] > 0.8
) / n if n else 0.0,
"ece": float(ece),
"total_compute": float(total_compute),
"total_retrievals": total_retrievals,
"results": results,
}
def _make_agent(self, agent_id: str = "rag_agent") -> SimulatedRetrievalAgent:
"""Create a fresh agent for fair comparison."""
return SimulatedRetrievalAgent(
agent_id=agent_id,
accuracy=0.65,
hallucination_rate=0.12,
calibration_error=0.15,
abstention_rate=0.1,
)
def run_all(self) -> Dict[str, Dict]:
if not self.questions:
self.generate_questions()
return {
"direct_answer": self.run_direct_answer(self._make_agent("direct_agent")),
"rag_baseline": self.run_rag_baseline(self._make_agent("rag_agent")),
"rag_verifier": self.run_rag_verifier(self._make_agent("verifier_agent")),
"occ_allocation": self.run_occ(self._make_agent("occ_agent")),
}
def main():
bench = RetrievalQABenchmark(n_questions=100, seed=42)
bench.generate_questions()
results = bench.run_all()
print("=" * 60)
print("RETRIEVAL QA BENCHMARK")
print("=" * 60)
for label, res in results.items():
print(f"\n{label}")
print(f" accuracy: {res['accuracy']:.3f}")
print(f" abstention_rate: {res['abstention_rate']:.3f}")
print(f" correct_abstentions: {res['correct_abstentions']}")
print(f" wrong_abstentions: {res['wrong_abstentions']}")
print(f" hallucination_rate: {res['hallucination_rate']:.3f}")
print(f" confident_wrong_rate: {res['confident_wrong_rate']:.3f}")
print(f" ECE: {res['ece']:.3f}")
print(f" total_compute: {res['total_compute']:.0f}")
print(f" total_retrievals: {res['total_retrievals']}")
Path("/app/occ/reports").mkdir(parents=True, exist_ok=True)
with open("/app/occ/reports/benchmark_retrieval_qa_results.json", "w") as f:
json.dump(results, f, indent=2, default=str)
print("\nSaved to reports/benchmark_retrieval_qa_results.json")
if __name__ == "__main__":
main()