Spaces:
Sleeping
Sleeping
File size: 11,819 Bytes
2aeb55a 92dc348 2aeb55a | 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 | """Empirical threshold calibration for VERIFIED_DENSE_THRESHOLD / VERIFIED_HYBRID_THRESHOLD.
Methodology
-----------
For each labeled query in the golden retrieval cases we retrieve a wide candidate pool
(match_threshold=0.20, match_count=20) from the real Supabase vector store, then label
every returned chunk as either True Positive (TP) or True Negative (TN):
TP = source name matches the expected source AND
at least one expected keyword is found in the chunk content
Everything else is TN.
We then build two separate score distributions:
Dense path β all TP vs all TN cosine similarity scores
Cross-modal β same, but restricted to chunks also found by FTS
For each candidate threshold t β [0.45, 0.80] we compute Youden's J statistic:
J(t) = TPR(t) β FPR(t)
= (TP above t / total TP) β (TN above t / total TN)
The threshold that maximises J is the operating point with minimum TP/TN overlap.
Requirements
------------
- Real Supabase connection (SUPABASE_URL + SUPABASE_KEY env vars or .env file)
- BGE-M3 model accessible (fastembed downloads on first run)
Usage
-----
cd "AI Chatbot"
python scripts/calibrate_threshold.py
"""
from __future__ import annotations
import asyncio
import json
import sys
from pathlib import Path
from typing import Any
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from app.retrieval_eval import load_golden_retrieval_cases
from app.vector_store import (
VERIFIED_DENSE_THRESHOLD,
VERIFIED_HYBRID_THRESHOLD,
search_knowledge,
search_knowledge_fts,
)
# ββ constants ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
CALIB_THRESHOLD = 0.20 # retrieval floor for data collection (wide net)
CALIB_COUNT = 20 # chunks per query (wider than production default of 7)
HISTOGRAM_BINS = 15 # score buckets for the ASCII plot
HISTOGRAM_LOW = 0.40 # left edge of histogram x-axis
HISTOGRAM_HIGH = 0.82 # right edge
CANDIDATE_RANGE = range(45, 81) # thresholds to evaluate (0.45 β 0.80 in 0.01 steps)
ScoreRecord = tuple[float, bool, bool, str] # (similarity, is_tp, is_cross_modal, case_id)
# ββ helpers ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _is_tp(
chunk: dict[str, Any],
expected_source: str,
expected_keywords: list[str],
expected_filename: str | None = None,
expected_page: int | None = None,
) -> bool:
source = chunk.get("source", "").strip()
content = (chunk.get("content") or "").lower()
source_match = source == expected_source.strip()
if not source_match:
return False
# Optionally narrow to exact page (when available, prefer it but don't require it exclusively)
if expected_page is not None and chunk.get("page_number") is not None:
if chunk.get("page_number") == expected_page:
return True # exact page + source = definite TP regardless of keywords
# wrong page β fall back to keyword check (might still be TP from adjacent page)
if not expected_keywords:
return True
return any(kw.lower() in content for kw in expected_keywords)
async def _collect_case_scores(
query: str,
expected_source: str,
expected_keywords: list[str],
case_id: str,
expected_filename: str | None = None,
expected_page: int | None = None,
) -> list[ScoreRecord]:
dense_chunks = await search_knowledge(
query=query,
match_threshold=CALIB_THRESHOLD,
match_count=CALIB_COUNT,
query_label=f"calib:{case_id}",
)
fts_chunks = await search_knowledge_fts(
query=query,
match_count=CALIB_COUNT,
)
fts_keys: set[tuple] = {
(c.get("filename"), c.get("page_number")) for c in fts_chunks
}
records: list[ScoreRecord] = []
for chunk in dense_chunks:
sim = chunk.get("similarity", 0.0)
key = (chunk.get("filename"), chunk.get("page_number"))
is_cross_modal = key in fts_keys
tp = _is_tp(chunk, expected_source, expected_keywords, expected_filename, expected_page)
records.append((sim, tp, is_cross_modal, case_id))
tp_count = sum(1 for _, tp, _, _ in records if tp)
print(f" β {len(records)} chunks returned, {tp_count} TP")
return records
async def collect_all_scores() -> list[ScoreRecord]:
cases = load_golden_retrieval_cases()
rag_cases = [
c for c in cases
if c.get("expected_mode") == "vector_rag" and c.get("expected_found") and c.get("expected_source")
]
print(f"\nCalibrating over {len(rag_cases)} labeled cases (vector_rag + calibration):\n")
all_records: list[ScoreRecord] = []
for case in rag_cases:
query = str(case["query"])
source = str(case["expected_source"])
keywords: list[str] = case.get("expected_content_keywords") or []
case_id = str(case["case_id"])
filename: str | None = case.get("expected_filename")
page: int | None = case.get("expected_page")
print(f" [{case_id}] \"{query[:70]}\"")
records = await _collect_case_scores(query, source, keywords, case_id, filename, page)
all_records.extend(records)
return all_records
# ββ analysis βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def compute_optimal_threshold(
records: list[ScoreRecord],
cross_modal_only: bool = False,
) -> tuple[float, float]:
"""Return (optimal_threshold, youden_j) maximising Youden's J over candidate range."""
subset = [
(sim, tp)
for sim, tp, is_cm, _ in records
if (not cross_modal_only or is_cm)
]
if not subset:
return 0.0, 0.0
tp_scores = [s for s, tp in subset if tp]
tn_scores = [s for s, tp in subset if not tp]
if not tp_scores or not tn_scores:
return 0.0, 0.0
best_t, best_j = 0.0, -99.0
for ti in CANDIDATE_RANGE:
t = ti / 100.0
tpr = sum(1 for s in tp_scores if s >= t) / len(tp_scores)
fpr = sum(1 for s in tn_scores if s >= t) / len(tn_scores)
j = tpr - fpr
if j > best_j:
best_j = j
best_t = t
return best_t, best_j
def print_histogram(
records: list[ScoreRecord],
cross_modal_only: bool = False,
) -> None:
subset = [
(sim, tp)
for sim, tp, is_cm, _ in records
if (not cross_modal_only or is_cm)
]
if not subset:
print(" (no data)\n")
return
tp_scores = [s for s, tp in subset if tp]
tn_scores = [s for s, tp in subset if not tp]
label = "Cross-modal chunks (dense AND fts)" if cross_modal_only else "All dense chunks"
print(f"\n{label} (n={len(subset)}, TP={len(tp_scores)}, TN={len(tn_scores)})")
print(f"{'Bucket':>14} {'TP':>4} {'TN':>4} {'TP (β)':25} {'TN (β)':25}")
print("β" * 75)
bin_width = (HISTOGRAM_HIGH - HISTOGRAM_LOW) / HISTOGRAM_BINS
max_count = 1
for i in range(HISTOGRAM_BINS):
lo = HISTOGRAM_LOW + i * bin_width
hi = lo + bin_width
max_count = max(
max_count,
sum(1 for s in tp_scores if lo <= s < hi),
sum(1 for s in tn_scores if lo <= s < hi),
)
for i in range(HISTOGRAM_BINS):
lo = HISTOGRAM_LOW + i * bin_width
hi = lo + bin_width
tp_n = sum(1 for s in tp_scores if lo <= s < hi)
tn_n = sum(1 for s in tn_scores if lo <= s < hi)
tp_bar = "β" * int(tp_n / max_count * 24)
tn_bar = "β" * int(tn_n / max_count * 24)
print(f" {lo:.2f}β{hi:.2f} {tp_n:>4} {tn_n:>4} {tp_bar:<25} {tn_bar}")
def print_precision_recall_table(
records: list[ScoreRecord],
cross_modal_only: bool = False,
) -> None:
"""Print precision / recall / F1 across the interesting threshold range."""
subset = [
(sim, tp)
for sim, tp, is_cm, _ in records
if (not cross_modal_only or is_cm)
]
tp_scores = [s for s, tp in subset if tp]
tn_scores = [s for s, tp in subset if not tp]
if not tp_scores:
return
print(f"\n{'Threshold':>12} {'TPR':>7} {'FPR':>7} {'J':>7} {'Prec':>7} {'F1':>7}")
print("β" * 55)
for ti in range(55, 78, 2):
t = ti / 100.0
tp_above = sum(1 for s in tp_scores if s >= t)
tn_above = sum(1 for s in tn_scores if s >= t)
tpr = tp_above / len(tp_scores) if tp_scores else 0
fpr = tn_above / len(tn_scores) if tn_scores else 0
j = tpr - fpr
prec = tp_above / (tp_above + tn_above) if (tp_above + tn_above) else 0
f1 = 2 * prec * tpr / (prec + tpr) if (prec + tpr) else 0
print(f" t={t:.2f} {tpr:>6.1%} {fpr:>6.1%} {j:>+7.3f} {prec:>6.1%} {f1:>6.1%}")
# ββ main βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
async def main() -> None:
records = await collect_all_scores()
total_tp = sum(1 for _, tp, _, _ in records if tp)
total_tn = sum(1 for _, tp, _, _ in records if not tp)
cross_modal_total = sum(1 for _, _, is_cm, _ in records if is_cm)
print(f"\nTotal data points: {len(records)} (TP={total_tp}, TN={total_tn}, cross-modal={cross_modal_total})\n")
# ββ Dense path ββ
print("=" * 75)
print("DENSE PATH CALIBRATION")
print("=" * 75)
print_histogram(records, cross_modal_only=False)
print_precision_recall_table(records, cross_modal_only=False)
dense_t, dense_j = compute_optimal_threshold(records, cross_modal_only=False)
# ββ Cross-modal path ββ
print("\n" + "=" * 75)
print("CROSS-MODAL PATH CALIBRATION (dense AND fts confirmed chunks only)")
print("=" * 75)
print_histogram(records, cross_modal_only=True)
print_precision_recall_table(records, cross_modal_only=True)
hybrid_t, hybrid_j = compute_optimal_threshold(records, cross_modal_only=True)
# ββ Recommendation ββ
print("\n" + "=" * 75)
print("RECOMMENDED THRESHOLDS (argmax Youden's J)")
print("=" * 75)
print(f" VERIFIED_DENSE_THRESHOLD = {dense_t:.2f} (J = {dense_j:+.3f})")
if hybrid_t:
print(f" VERIFIED_HYBRID_THRESHOLD = {hybrid_t:.2f} (J = {hybrid_j:+.3f})")
else:
print(" VERIFIED_HYBRID_THRESHOLD = (insufficient cross-modal data)")
print()
print(" Currently set:")
print(f" VERIFIED_DENSE_THRESHOLD = {VERIFIED_DENSE_THRESHOLD}")
print(f" VERIFIED_HYBRID_THRESHOLD = {VERIFIED_HYBRID_THRESHOLD}")
print()
if abs(dense_t - VERIFIED_DENSE_THRESHOLD) < 0.01:
print(" β
Dense threshold looks well-calibrated.")
else:
direction = "β raise" if dense_t > VERIFIED_DENSE_THRESHOLD else "β lower"
print(f" β οΈ Dense threshold should change: {VERIFIED_DENSE_THRESHOLD} β {dense_t:.2f} ({direction})")
if hybrid_t and abs(hybrid_t - VERIFIED_HYBRID_THRESHOLD) < 0.01:
print(" β
Hybrid threshold looks well-calibrated.")
elif hybrid_t:
direction = "β raise" if hybrid_t > VERIFIED_HYBRID_THRESHOLD else "β lower"
print(f" β οΈ Hybrid threshold should change: {VERIFIED_HYBRID_THRESHOLD} β {hybrid_t:.2f} ({direction})")
if __name__ == "__main__":
asyncio.run(main())
|