Spaces:
Sleeping
Sleeping
File size: 19,798 Bytes
6bef416 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 | from __future__ import annotations
from collections.abc import Iterable, Sequence
from dataclasses import dataclass
import numpy as np
import pandas as pd
MIN_SLICE_N = 30
MIN_CONFIG_N = 50
# Offline diagnostic boundary: Recall@10 >= 0.80 means most gold evidence is present in the top-10 retrieved chunks.
# It is a review lens for this evaluation corpus, not a production deployment policy.
RETRIEVAL_OK_THRESHOLD = 0.80
MISSING_LABEL = "Missing / Not provided"
# Review-priority weights for offline risk slicing. They intentionally emphasize answer error and
# hallucination exposure ahead of retrieval weakness because the dashboard is meant to drive human
# QA review queues, not tune a live serving policy. Keep these values deterministic for reproducible
# portfolio artifacts; recalibrate them before using the approach on a real production corpus.
RISK_SCORE_WEIGHTS = {"error": 0.45, "hallucination": 0.35, "retrieval": 0.20}
# Evidence-strength weights summarize retrieval-side evidence signals for offline policy review.
# This score is not model confidence or a calibrated probability of answer correctness.
EVIDENCE_STRENGTH_WEIGHTS = {"top1_score": 0.35, "mean_retrieved_score": 0.15, "recall_at_10": 0.30, "mrr_at_10": 0.20}
# Offline objective presets used by the configuration leaderboard. The weights are hand-tuned review
# lenses for this bundled synthetic evaluation set: quality-heavy by default, with alternative views
# for hallucination, latency, and cost sensitivity. They are not learned coefficients or production SLAs.
CONFIG_OBJECTIVE_WEIGHTS = {
"Balanced": {"correct": 0.42, "recall": 0.18, "halluc": 0.22, "latency": 0.10, "cost": 0.08},
"Max quality": {"correct": 0.55, "recall": 0.25, "halluc": 0.15, "latency": 0.03, "cost": 0.02},
"Min hallucination": {"correct": 0.30, "recall": 0.15, "halluc": 0.45, "latency": 0.05, "cost": 0.05},
"Low latency": {"correct": 0.28, "recall": 0.12, "halluc": 0.15, "latency": 0.38, "cost": 0.07},
"Low cost": {"correct": 0.28, "recall": 0.12, "halluc": 0.15, "latency": 0.07, "cost": 0.38},
}
@dataclass(frozen=True)
class DecisionBrief:
posture: str
posture_reason: str
main_driver: str
worst_slice: str
best_config: str
recommended_action: str
def _numeric(series: pd.Series) -> pd.Series:
return pd.to_numeric(series, errors="coerce")
def _p95(series: pd.Series) -> float:
val = _numeric(series).quantile(0.95)
return np.nan if pd.isna(val) else float(val)
def _safe_col(df: pd.DataFrame, col: str, default: float = np.nan) -> pd.Series:
if col in df.columns:
return _numeric(df[col])
return pd.Series(default, index=df.index, dtype="float64")
def safe_mean(df: pd.DataFrame, col: str, default: float = 0.0) -> float:
if col not in df.columns or len(df) == 0:
return default
val = _numeric(df[col]).mean()
return default if pd.isna(val) else float(val)
def safe_p95(df: pd.DataFrame, col: str, default: float = 0.0) -> float:
if col not in df.columns or len(df) == 0:
return default
val = _numeric(df[col]).quantile(0.95)
return default if pd.isna(val) else float(val)
def fmt_pct(x: float) -> str:
if pd.isna(x):
return "n/a"
return f"{x:.1%}"
def fmt_money(x: float) -> str:
if pd.isna(x):
return "n/a"
return f"${x:.4f}"
def overview_metrics(eval_df: pd.DataFrame, docs: pd.DataFrame, chunks: pd.DataFrame, retrieval: pd.DataFrame) -> dict[str, float]:
return {
"evaluations": float(len(eval_df)),
"retrieval_events": float(len(retrieval)),
"documents": float(len(docs)),
"chunks": float(len(chunks)),
"correct_rate": safe_mean(eval_df, "is_correct", default=np.nan),
"hallucination_rate": safe_mean(eval_df, "hallucination_flag", default=np.nan),
"recall_at_10": safe_mean(eval_df, "recall_at_10", default=np.nan),
"mrr_at_10": safe_mean(eval_df, "mrr_at_10", default=np.nan),
"p95_latency_ms": safe_p95(eval_df, "total_latency_ms", default=np.nan),
"avg_cost_usd": safe_mean(eval_df, "total_cost_usd", default=np.nan),
}
def quality_posture(metrics: dict[str, float]) -> tuple[str, str]:
correct = metrics.get("correct_rate", np.nan)
halluc = metrics.get("hallucination_rate", np.nan)
recall = metrics.get("recall_at_10", np.nan)
if pd.isna(correct) or pd.isna(halluc) or pd.isna(recall):
return "Review", "one or more key quality signals are unavailable under the current filters"
if correct >= 0.78 and halluc <= 0.10 and recall >= 0.70:
return "Stable", "correctness, hallucination, and retrieval signals are within a usable operating band"
if correct < 0.62 or halluc > 0.20 or recall < 0.45:
return "High Risk", "one or more quality signals are outside the expected operating band"
return "Watch", "quality is usable for analysis, but risk slices require targeted review"
def risk_slices(
eval_df: pd.DataFrame,
group_cols: Sequence[str] = ("domain", "scenario_type", "difficulty"),
min_n: int = MIN_SLICE_N,
) -> pd.DataFrame:
"""Aggregate risk by slice without silent count-as-metric fallbacks."""
available = [c for c in group_cols if c in eval_df.columns]
if not available or len(eval_df) == 0:
return pd.DataFrame()
src = eval_df.copy()
for col in available:
src[col] = src[col].astype("string").fillna(MISSING_LABEL)
src["__row_count"] = 1
agg_map = {"n": ("__row_count", "size")}
optional_aggs = {
"correct_rate": ("is_correct", "mean"),
"hallucination_rate": ("hallucination_flag", "mean"),
"recall_at_10": ("recall_at_10", "mean"),
"mrr_at_10": ("mrr_at_10", "mean"),
"p95_latency_ms": ("total_latency_ms", _p95),
"avg_cost_usd": ("total_cost_usd", "mean"),
}
for out_col, spec in optional_aggs.items():
if spec[0] in src.columns:
agg_map[out_col] = spec
out = src.groupby(available, dropna=False).agg(**agg_map).reset_index()
out = out[out["n"] >= int(min_n)].copy()
if out.empty:
return out
for col in ["correct_rate", "hallucination_rate", "recall_at_10", "mrr_at_10", "p95_latency_ms", "avg_cost_usd"]:
if col not in out.columns:
out[col] = np.nan
out["error_rate"] = 1 - out["correct_rate"]
error_component = out["error_rate"].fillna(0.0).clip(0, 1)
halluc_component = out["hallucination_rate"].fillna(0.0).clip(0, 1)
retrieval_component = (1 - out["recall_at_10"].fillna(1.0)).clip(0, 1)
out["risk_score"] = (
error_component * RISK_SCORE_WEIGHTS["error"]
+ halluc_component * RISK_SCORE_WEIGHTS["hallucination"]
+ retrieval_component * RISK_SCORE_WEIGHTS["retrieval"]
)
return out.sort_values("risk_score", ascending=False).reset_index(drop=True)
def retrieval_outcomes(eval_df: pd.DataFrame, threshold: float = RETRIEVAL_OK_THRESHOLD) -> pd.DataFrame:
"""Classify rows into retrieval/generation/hallucination modes using one canonical implementation."""
required = {"recall_at_10", "is_correct"}
if len(eval_df) == 0 or not required.issubset(eval_df.columns):
return pd.DataFrame()
src = eval_df.copy()
src["__row_count"] = 1
src["retrieval_state"] = np.where(_numeric(src["recall_at_10"]) >= threshold, "retrieval_ok", "retrieval_weak")
src["answer_state"] = np.where(_numeric(src["is_correct"]).fillna(0.0) >= 0.5, "answer_correct", "answer_incorrect")
halluc = _safe_col(src, "hallucination_flag", 0.0).fillna(0.0) >= 0.5
src["failure_mode"] = np.select(
[
halluc & (src["answer_state"] == "answer_correct"),
halluc & (src["answer_state"] == "answer_incorrect"),
(src["retrieval_state"] == "retrieval_weak") & (src["answer_state"] == "answer_incorrect"),
(src["retrieval_state"] == "retrieval_ok") & (src["answer_state"] == "answer_incorrect"),
(src["retrieval_state"] == "retrieval_weak") & (src["answer_state"] == "answer_correct"),
(src["retrieval_state"] == "retrieval_ok") & (src["answer_state"] == "answer_correct"),
],
[
"hallucination_risk_correct_answer",
"hallucination_failure",
"retrieval_failure",
"generation_failure",
"recovered_by_generation",
"healthy",
],
default=MISSING_LABEL,
)
agg_map = {
"n": ("__row_count", "size"),
"correct_rate": ("is_correct", "mean"),
"recall_at_10": ("recall_at_10", "mean"),
}
if "hallucination_flag" in src.columns:
agg_map["hallucination_rate"] = ("hallucination_flag", "mean")
if "total_latency_ms" in src.columns:
agg_map["p95_latency_ms"] = ("total_latency_ms", _p95)
if "total_cost_usd" in src.columns:
agg_map["avg_cost_usd"] = ("total_cost_usd", "mean")
out = src.groupby("failure_mode", dropna=False).agg(**agg_map).reset_index()
for col in ["hallucination_rate", "p95_latency_ms", "avg_cost_usd"]:
if col not in out.columns:
out[col] = np.nan
out["share"] = out["n"] / max(out["n"].sum(), 1)
return out.sort_values("n", ascending=False).reset_index(drop=True)
def p95_scaled(series: pd.Series) -> pd.Series:
vals = _numeric(series)
denom = vals.quantile(0.95)
if pd.isna(denom) or denom <= 0:
return pd.Series(np.zeros(len(vals)), index=vals.index)
return (vals.fillna(0.0) / denom).clip(0, 2)
def _rate_ci(p: pd.Series, n: pd.Series, z: float = 1.96) -> tuple[pd.Series, pd.Series]:
p = _numeric(p).clip(0, 1)
n = _numeric(n).clip(lower=1)
se = np.sqrt((p * (1 - p)) / n)
return (p - z * se).clip(0, 1), (p + z * se).clip(0, 1)
def config_leaderboard(eval_df: pd.DataFrame, objective: str = "Balanced", min_n: int = MIN_CONFIG_N) -> pd.DataFrame:
required = {"retrieval_strategy", "generator_model", "is_correct", "hallucination_flag", "total_latency_ms", "total_cost_usd"}
if not required.issubset(eval_df.columns) or len(eval_df) == 0:
return pd.DataFrame()
src = eval_df.copy()
src["__row_count"] = 1
group_cols = ["retrieval_strategy", "generator_model"]
if "chunking_strategy" in src.columns:
group_cols.append("chunking_strategy")
for col in group_cols:
src[col] = src[col].astype("string").fillna(MISSING_LABEL)
agg_map = {
"n": ("__row_count", "size"),
"correct_rate": ("is_correct", "mean"),
"hallucination_rate": ("hallucination_flag", "mean"),
"p95_latency_ms": ("total_latency_ms", _p95),
"avg_cost_usd": ("total_cost_usd", "mean"),
}
if "recall_at_10" in src.columns:
agg_map["recall_at_10"] = ("recall_at_10", "mean")
if "mrr_at_10" in src.columns:
agg_map["mrr_at_10"] = ("mrr_at_10", "mean")
out = src.groupby(group_cols, dropna=False).agg(**agg_map).reset_index()
out = out[out["n"] >= int(min_n)].copy()
if out.empty:
return out
for col in ["recall_at_10", "mrr_at_10"]:
if col not in out.columns:
out[col] = np.nan
out["correct_rate_ci_low"], out["correct_rate_ci_high"] = _rate_ci(out["correct_rate"], out["n"])
out["hallucination_rate_ci_low"], out["hallucination_rate_ci_high"] = _rate_ci(out["hallucination_rate"], out["n"])
out["latency_scaled"] = p95_scaled(out["p95_latency_ms"])
out["cost_scaled"] = p95_scaled(out["avg_cost_usd"])
weights = CONFIG_OBJECTIVE_WEIGHTS.get(objective, CONFIG_OBJECTIVE_WEIGHTS["Balanced"])
out["score"] = (
out["correct_rate"].fillna(0) * weights["correct"]
+ out["recall_at_10"].fillna(0) * weights["recall"]
- out["hallucination_rate"].fillna(0) * weights["halluc"]
- out["latency_scaled"].fillna(0) * weights["latency"]
- out["cost_scaled"].fillna(0) * weights["cost"]
)
out["config"] = out[group_cols].astype(str).agg(" / ".join, axis=1)
return out.sort_values("score", ascending=False).reset_index(drop=True)
def demand_coverage(eval_df: pd.DataFrame, docs: pd.DataFrame) -> pd.DataFrame:
if "domain" not in eval_df.columns or "domain" not in docs.columns:
return pd.DataFrame()
demand = eval_df["domain"].astype("string").fillna(MISSING_LABEL).value_counts(normalize=True).rename("eval_demand_share")
corpus = docs["domain"].astype("string").fillna(MISSING_LABEL).value_counts(normalize=True).rename("corpus_document_share")
out = pd.concat([demand, corpus], axis=1).fillna(0)
out.index.name = "domain"
out = out.reset_index()
out["demand_minus_corpus"] = out["eval_demand_share"] - out["corpus_document_share"]
return out.sort_values("eval_demand_share", ascending=False).reset_index(drop=True)
def evidence_strength_proxy(eval_df: pd.DataFrame, reference_df: pd.DataFrame | None = None) -> pd.Series:
"""Offline evidence-strength proxy derived from retrieval-side evaluation signals.
The score is not LLM confidence, a calibrated correctness probability, or a production
approval signal. The optional reference_df fixes normalization anchors so threshold
behavior does not drift when the user changes dashboard filters.
"""
ref = reference_df if reference_df is not None and len(reference_df) else eval_df
parts = []
weights = []
for col, weight in EVIDENCE_STRENGTH_WEIGHTS.items():
if col in eval_df.columns:
s = _numeric(eval_df[col]).fillna(0.0)
ref_s = _numeric(ref[col]).fillna(0.0) if col in ref.columns else s
min_v, max_v = ref_s.min(), ref_s.max()
if max_v > min_v:
s = (s - min_v) / (max_v - min_v)
parts.append(s.clip(0, 1) * weight)
weights.append(weight)
if not parts:
return pd.Series(np.zeros(len(eval_df)), index=eval_df.index)
score = sum(parts) / max(sum(weights), 1e-9)
return score.clip(0, 1)
def policy_curve(
eval_df: pd.DataFrame,
thresholds: Iterable[float] | None = None,
reference_df: pd.DataFrame | None = None,
) -> pd.DataFrame:
if len(eval_df) == 0:
return pd.DataFrame()
if thresholds is None:
thresholds = np.linspace(0.05, 0.95, 19)
src = eval_df.copy()
src["evidence_strength_proxy"] = evidence_strength_proxy(src, reference_df=reference_df)
rows = []
total_hallucinations = _safe_col(src, "hallucination_flag", 0.0).sum()
for thr in thresholds:
auto = src[src["evidence_strength_proxy"] >= thr]
review = src[src["evidence_strength_proxy"] < thr]
review_hallucinations = _safe_col(review, "hallucination_flag", 0.0).sum()
rows.append(
{
"threshold": float(thr),
"auto_approve_rate": len(auto) / max(len(src), 1),
"review_queue_size": int(len(review)),
"auto_correct_rate": safe_mean(auto, "is_correct", default=np.nan) if len(auto) else np.nan,
"auto_hallucination_rate": safe_mean(auto, "hallucination_flag", default=np.nan) if len(auto) else np.nan,
"risk_captured_in_review": (review_hallucinations / total_hallucinations) if total_hallucinations > 0 else np.nan,
}
)
return pd.DataFrame(rows)
def policy_at_threshold(eval_df: pd.DataFrame, threshold: float, reference_df: pd.DataFrame | None = None) -> dict[str, float]:
curve = policy_curve(eval_df, [threshold], reference_df=reference_df)
if curve.empty:
return {}
return curve.iloc[0].to_dict()
def make_decision_brief(
eval_df: pd.DataFrame,
docs: pd.DataFrame,
chunks: pd.DataFrame,
retrieval: pd.DataFrame,
min_slice_n: int = MIN_SLICE_N,
min_config_n: int = MIN_CONFIG_N,
*,
risk_table: pd.DataFrame | None = None,
retrieval_table: pd.DataFrame | None = None,
config_table: pd.DataFrame | None = None,
) -> DecisionBrief:
metrics = overview_metrics(eval_df, docs, chunks, retrieval)
posture, reason = quality_posture(metrics)
retrieval_table = retrieval_table if retrieval_table is not None else retrieval_outcomes(eval_df)
risk = risk_table if risk_table is not None else risk_slices(eval_df, min_n=min_slice_n)
configs = config_table if config_table is not None else config_leaderboard(eval_df, min_n=min_config_n)
main_driver = "Mixed"
if not retrieval_table.empty:
# Pick the dominant non-healthy mode for actionability. A large healthy segment should not hide
# the strongest remaining failure class in the decision strip.
driver_rows = retrieval_table[retrieval_table["failure_mode"].astype(str) != "healthy"]
if driver_rows.empty:
driver_rows = retrieval_table
top_mode = str(driver_rows.iloc[0]["failure_mode"])
if "hallucination" in top_mode:
main_driver = "Hallucination"
elif "retrieval" in top_mode:
main_driver = "Retrieval"
elif "generation" in top_mode:
main_driver = "Generation"
elif top_mode == "healthy":
main_driver = "Healthy majority"
worst_slice = "No high-risk slice above minimum sample size"
if not risk.empty:
row = risk.iloc[0]
parts = [str(row[c]) for c in ["domain", "scenario_type", "difficulty"] if c in risk.columns]
worst_slice = " / ".join(parts) + f" 路 risk={row['risk_score']:.2f}"
best_config = "No eligible configuration"
if not configs.empty:
row = configs.iloc[0]
best_config = f"{row['config']} 路 score={row['score']:.2f} 路 n={int(row['n'])}"
if posture == "High Risk":
action = "Prioritize the top risk slice and inspect retrieval evidence before widening auto-approval."
elif main_driver == "Retrieval":
action = "Start with retrieval diagnostics: recall coverage, chunk ranking, and corpus-demand alignment."
elif main_driver == "Generation":
action = "Inspect answer generation behavior on retrieval-ok but incorrect examples."
elif main_driver == "Hallucination":
action = "Review hallucination-heavy examples even when answer correctness appears acceptable."
else:
action = "Use the policy simulator to choose a review threshold that balances coverage and risk."
return DecisionBrief(
posture=posture,
posture_reason=reason,
main_driver=main_driver,
worst_slice=worst_slice,
best_config=best_config,
recommended_action=action,
)
def top_examples(eval_df: pd.DataFrame, mode: str = "High risk", n: int = 100, reference_df: pd.DataFrame | None = None) -> pd.DataFrame:
if len(eval_df) == 0:
return pd.DataFrame()
src = eval_df.copy()
src["evidence_strength_proxy"] = evidence_strength_proxy(src, reference_df=reference_df)
src["risk_rank_score"] = (
(1 - _safe_col(src, "is_correct", 0.0).fillna(0.0)).clip(0, 1) * RISK_SCORE_WEIGHTS["error"]
+ _safe_col(src, "hallucination_flag", 0.0).fillna(0.0).clip(0, 1) * RISK_SCORE_WEIGHTS["hallucination"]
+ (1 - _safe_col(src, "recall_at_10", 1.0).fillna(1.0)).clip(0, 1) * RISK_SCORE_WEIGHTS["retrieval"]
)
if mode == "Incorrect" and "is_correct" in src.columns:
src = src[_numeric(src["is_correct"]).fillna(0) == 0]
elif mode == "Hallucination" and "hallucination_flag" in src.columns:
src = src[_numeric(src["hallucination_flag"]).fillna(0) == 1]
elif mode == "Low retrieval" and "recall_at_10" in src.columns:
src = src[_numeric(src["recall_at_10"]).fillna(1) < RETRIEVAL_OK_THRESHOLD]
return src.sort_values("risk_rank_score", ascending=False).head(n).reset_index(drop=True)
|