hyungi/hyungi_document_server
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(ask): Phase 3.5b guardrails — verifier + telemetry + grounding 강화

Browse Source 
Phase 3.5a(classifier+refusal gate+grounding) 위에 4개 Item 추가:

Item 0: ask_events telemetry 배선
- AskEvent ORM 모델 + record_ask_event() — ask_events INSERT 완성
- defense_layers에 input_snapshot(query, chunks, answer) 저장
- refused/normal 두 경로 모두 telemetry 호출

Item 3: evidence 간 numeric conflict detection
- 동일 단위 다른 숫자 → weak flag
- "이상/이하/초과/미만" threshold 표현 → skip (FP 방지)

Item 4: fabricated_number normalization 개선
- 단위 접미사 건/원 추가, 범위 표현(10~20%) 양쪽 추출
- bare number 2자리 이상만 (1자리 FP 제거)

Item 1: exaone semantic verifier (판단권 잠금 배선)
- verifier_service.py — 3s timeout, circuit breaker, severity 3단계
- direct_negation만 strong, numeric/intent→medium, 나머지→weak
- verifier strong 단독 refuse 금지 — grounding과 교차 필수
- 6-tier re-gate (4라운드 리뷰 확정)
- grounding strong 2+ OR max_score<0.2 → verifier skip

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Hyungi Ahn
2026-04-10 09:49:56 +09:00
parent a0e1717206
commit b2306c3afd
9 changed files with 533 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

127
app/api/search.py
View File

@@ -28,7 +28,8 @@ from services.search.grounding_check import check as grounding_check
from services.search.refusal_gate import RefusalDecision, decide as refusal_decide
from services.search.search_pipeline import PipelineResult, run_search
from services.search.synthesis_service import SynthesisResult, synthesize
from services.search_telemetry import record_search_event
from services.search.verifier_service import VerifierResult, verify
from services.search_telemetry import record_ask_event, record_search_event
# logs/search.log + stdout 동시 출력 (Phase 0.4)
logger = setup_logger("search")
@@ -451,11 +452,29 @@ async def ask(
q[:80], decision.rule_triggered,
max(all_rerank_scores) if all_rerank_scores else 0.0, total_ms,
)
# telemetry
# telemetry — search + ask_events 두 경로 동시
background_tasks.add_task(
record_search_event, q, user.id, pr.results, "hybrid",
pr.confidence_signal, pr.analyzer_confidence,
)
# input_snapshot (디버깅/재현용)
defense_log["input_snapshot"] = {
"query": q,
"top_chunks_preview": [
{"title": c.get("title", ""), "snippet": c.get("snippet", "")[:100]}
for c in top_chunks[:3]
],
"answer_preview": None,
}
background_tasks.add_task(
record_ask_event,
q, user.id, "insufficient", "skipped", None,
True, classifier_result.verdict,
max(all_rerank_scores) if all_rerank_scores else 0.0,
sum(sorted(all_rerank_scores, reverse=True)[:3]),
[], len(evidence), 0,
defense_log, int(total_ms),
)
debug_obj = None
if debug:
debug_obj = AskDebug(
@@ -491,36 +510,102 @@ async def ask(
sr = await synthesize(q, evidence, debug=debug)
synth_ms = (time.perf_counter() - t_synth) * 1000
# 5. Grounding check (post-synthesis) + re-gate
# 5. Grounding check + Verifier (조건부 병렬) + re-gate (Phase 3.5b)
grounding = grounding_check(q, sr.answer or "", evidence)
# verifier skip: grounding strong 2+ OR retrieval 자체가 망함
grounding_only_strong = [
f for f in grounding.strong_flags if not f.startswith("verifier_")
]
max_rerank = max(all_rerank_scores, default=0.0)
if len(grounding_only_strong) >= 2 or max_rerank < 0.2:
verifier_result = VerifierResult("skipped", [], 0.0)
else:
verifier_task = asyncio.create_task(
verify(q, sr.answer or "", evidence)
)
try:
verifier_result = await asyncio.wait_for(verifier_task, timeout=4.0)
except (asyncio.TimeoutError, Exception):
verifier_result = VerifierResult("timeout", [], 0.0)
# Verifier contradictions → grounding flags 머지 (prefix 로 구분, severity 3단계)
for c in verifier_result.contradictions:
if c.severity == "strong":
grounding.strong_flags.append(f"verifier_{c.type}:{c.claim[:30]}")
elif c.severity == "medium":
grounding.weak_flags.append(f"verifier_{c.type}_medium:{c.claim[:30]}")
else:
grounding.weak_flags.append(f"verifier_{c.type}:{c.claim[:30]}")
defense_log["grounding"] = {
"strong": grounding.strong_flags,
"weak": grounding.weak_flags,
}
defense_log["verifier"] = {
"status": verifier_result.status,
"contradictions_count": len(verifier_result.contradictions),
"strong_count": sum(1 for c in verifier_result.contradictions if c.severity == "strong"),
"medium_count": sum(1 for c in verifier_result.contradictions if c.severity == "medium"),
"elapsed_ms": verifier_result.elapsed_ms,
}
# Completeness 결정: grounding 기반 (classifier 는 binary gate 만)
# ── Re-gate: 6-tier completeness 결정 (Phase 3.5b 4차 리뷰 확정) ──
completeness: Literal["full", "partial", "insufficient"] = "full"
covered_aspects = classifier_result.covered_aspects or None
missing_aspects = classifier_result.missing_aspects or None
confirmed_items: list[ConfirmedItem] | None = None
if len(grounding.strong_flags) >= 2:
# Re-gate: multiple strong → refuse
# verifier/grounding strong 구분
g_strong = [f for f in grounding.strong_flags if not f.startswith("verifier_")]
v_strong = [f for f in grounding.strong_flags if f.startswith("verifier_")]
v_medium = [f for f in grounding.weak_flags if f.startswith("verifier_") and "_medium:" in f]
has_direct_negation = any("direct_negation" in f for f in v_strong)
if len(g_strong) >= 2:
# Tier 1: grounding strong 2+ → refuse
completeness = "insufficient"
sr.answer = None
sr.refused = True
sr.confidence = None
defense_log["re_gate"] = "refuse(2+strong)"
elif grounding.strong_flags:
# Single strong → partial downgrade
defense_log["re_gate"] = "refuse(grounding_2+strong)"
elif g_strong and has_direct_negation:
# Tier 2: grounding strong + verifier direct_negation → refuse
completeness = "insufficient"
sr.answer = None
sr.refused = True
sr.confidence = None
defense_log["re_gate"] = "refuse(grounding+direct_negation)"
elif g_strong and sr.confidence == "low" and max_rerank < 0.25:
# Tier 3: grounding strong 1 + (low confidence AND weak evidence) → refuse
completeness = "insufficient"
sr.answer = None
sr.refused = True
sr.confidence = None
defense_log["re_gate"] = "refuse(grounding+low_conf+weak_ev)"
elif g_strong or has_direct_negation:
# Tier 4: grounding strong 1 또는 verifier direct_negation 단독 → partial
completeness = "partial"
sr.confidence = "low"
defense_log["re_gate"] = "partial(1strong)"
defense_log["re_gate"] = "partial(strong_or_negation)"
elif v_medium:
# Tier 5: verifier medium 누적 → count 기반 confidence 하향
medium_count = len(v_medium)
if medium_count >= 3:
sr.confidence = "low"
defense_log["re_gate"] = f"conf_low(medium_x{medium_count})"
elif medium_count == 2 and sr.confidence == "high":
sr.confidence = "medium"
defense_log["re_gate"] = "conf_cap_medium(medium_x2)"
else:
defense_log["re_gate"] = f"medium_x{medium_count}(no_action)"
elif grounding.weak_flags:
# Weak → confidence lower only
# Tier 6: weak → confidence 한 단계 하향
if sr.confidence == "high":
sr.confidence = "medium"
defense_log["re_gate"] = "conf_lower(weak)"
else:
defense_log["re_gate"] = "clean"
# Confidence cap from refusal gate (classifier 부재 시 conservative)
if decision.confidence_cap and sr.confidence:
@@ -554,11 +639,29 @@ async def ask(
ev_ms, synth_ms, total_ms,
)
# 7. telemetry
# 7. telemetry — search + ask_events 두 경로 동시
background_tasks.add_task(
record_search_event, q, user.id, pr.results, "hybrid",
pr.confidence_signal, pr.analyzer_confidence,
)
# input_snapshot (디버깅/재현용)
defense_log["input_snapshot"] = {
"query": q,
"top_chunks_preview": [
{"title": (r.title or "")[:50], "snippet": (r.snippet or "")[:100]}
for r in pr.results[:3]
],
"answer_preview": (sr.answer or "")[:200],
}
background_tasks.add_task(
record_ask_event,
q, user.id, completeness, sr.status, sr.confidence,
sr.refused, classifier_result.verdict,
max(all_rerank_scores) if all_rerank_scores else 0.0,
sum(sorted(all_rerank_scores, reverse=True)[:3]),
sr.hallucination_flags, len(evidence), len(citations),
defense_log, int(total_ms),
)
debug_obj = None
if debug: