File size: 12,644 Bytes
3be54c6 | 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 | """retrieval_signals.py — Universal signals over retrieval-time score
distributions.
Completes the third leg of the τψφξΩ trio:
1. Training-time signals (next_token_trainer.compute_tau_signals)
2. Inference-time signals (outcome_signals.compute_outcome_signals)
3. **Retrieval-time signals** ← this module
Why this matters:
A query whose top-k retrieval has uniform low scores (everyone scores
0.20) means the corpus has nothing relevant — the user's question is
either too narrow or too off-topic. A query whose top-k has one big
score and a long tail has good signal. We can MEASURE this gap.
With these signals exposed:
• UI can show "אמינות אחזור: 72%" next to the answer
• System can detect retrieval drift over time (Ω falling = corpus
degrading or query distribution shifting)
• Pipeline-health endpoint can alert on sustained low Ω
────────────────────────────────────────────────────────────────────────────
Signal definitions (all clipped to [0, 1])
────────────────────────────────────────────────────────────────────────────
τ (tau, "top-hit strength"):
How strong is the best result relative to a baseline?
τ = clip01( top_score / strong_threshold )
With strong_threshold ≈ 0.6 (calibrated for HebrewEncoder hybrid
scores). τ = 1.0 means the top hit is unambiguously a strong match.
ψ (psi, "score concentration"):
Is the score mass concentrated in a few top results, or spread
thin? Use coefficient-of-variation on top-k:
ψ = clip01( std(top-k) / (mean(top-k) + ε) )
High ψ means the top-k differ a lot — that's actually GOOD: the
retriever is discriminating well between the top hit and the rest.
Low ψ (uniform scores) means the retriever can't tell them apart.
φ (phi, "top-k topical agreement"):
Do the top-k results agree on a topic? Computed via the centroid:
φ = mean cosine of top-k vectors with their own centroid
If 5/5 results are about "good faith violation", φ ≈ 0.95. If
5/5 are scattered topics, φ ≈ 0.50.
ξ (xi, "score-gap anomaly"):
Is there a discontinuity in the score curve, or smooth decay?
A "shelf" pattern (scores: 0.85, 0.80, 0.20, 0.15, 0.10) means
the top-2 are clearly different — high confidence. A smooth
decay (0.85, 0.83, 0.81, 0.79, 0.77) means the boundary between
relevant and irrelevant is fuzzy.
ξ = 1 - clip01( max_gap / (top_score - bottom_score + ε) )
Low ξ = clear shelf (good), high ξ = smooth decay (uncertain).
Ω (omega):
Standard geometric mean.
Ω = (τ^α · φ^β · ψ^γ · (1−ξ)^δ)^(1/Σ)
"""
from __future__ import annotations
import math
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple
def _clip01(x: float) -> float:
return float(max(0.0, min(1.0, x)))
@dataclass
class RetrievalSignals:
"""Universal signals computed over a single retrieval result list."""
tau: float # top-hit strength
psi: float # score concentration (CV-based)
phi: float # topical agreement of top-k
xi: float # score-gap anomaly
omega: float # combined retrieval health
n_results: int = 0
top_score: float = 0.0
score_range: float = 0.0
debug: Dict[str, Any] = field(default_factory=dict)
def to_dict(self) -> Dict[str, Any]:
return {
"tau": round(self.tau, 4),
"psi": round(self.psi, 4),
"phi": round(self.phi, 4),
"xi": round(self.xi, 4),
"omega": round(self.omega, 4),
"retrieval_health": round(self.omega, 4),
"n_results": self.n_results,
"top_score": round(self.top_score, 4),
"score_range": round(self.score_range, 4),
"interpretation": self._interpret(),
"debug": self.debug,
}
def _interpret(self) -> Dict[str, str]:
out = {}
out["overall"] = (
f"אמינות אחזור: {self.omega*100:.0f}% — " + (
"אחזור איכותי" if self.omega >= 0.65 else
"אחזור בינוני" if self.omega >= 0.45 else
"אחזור חלש — שקול לנסח את השאילתה שוב"
)
)
out["tau"] = (
f"חוזק תוצאה ראשונה (τ={self.tau:.2f}): " + (
"התאמה ברורה" if self.tau >= 0.65 else
"התאמה חלקית" if self.tau >= 0.45 else
"התאמה חלשה — אין במאגר תוצאה דומה במובהק"
)
)
out["psi"] = (
f"הבחנה בין תוצאות (ψ={self.psi:.2f}): " + (
"המנוע מבחין היטב בין רלוונטי ולא" if self.psi >= 0.50 else
"ניקוד אחיד — מנוע מתקשה לדרג"
)
)
out["phi"] = (
f"לכידות נושאית (φ={self.phi:.2f}): " + (
"התוצאות עוסקות באותו נושא" if self.phi >= 0.65 else
"התוצאות מפוזרות על פני נושאים שונים" if self.phi >= 0.45 else
"פיזור גבוה — שאילתה רב-משמעית"
)
)
out["xi"] = (
f"גבול ברור (ξ={self.xi:.2f}, נמוך=טוב): " + (
"יש 'מדף' ברור בין רלוונטי ללא" if self.xi <= 0.35 else
"גבול מטושטש בין תוצאות"
)
)
return out
def compute_retrieval_signals(
hits: List[Any],
strong_threshold: float = 0.6,
omega_weights: Tuple[float, float, float, float] = (1.0, 1.0, 1.0, 1.0),
eps: float = 1e-6,
) -> RetrievalSignals:
"""Compute τψφξΩ over a retrieval result list.
Args:
hits: list of `Retrieved` (or any object with `.score` and
optionally `.chunk.text` for φ computation). Order is
assumed descending by score.
strong_threshold: τ = top_score / strong_threshold, clipped to 1.
Calibrated to the HebrewEncoder hybrid score distribution.
omega_weights: (α_τ, β_φ, γ_ψ, δ_ξ).
eps: numerical-stability constant.
Returns: RetrievalSignals with all 5 signals + debug breakdown.
"""
if not hits:
return RetrievalSignals(
tau=0.0, psi=0.0, phi=0.0, xi=1.0, omega=0.0,
n_results=0, top_score=0.0, score_range=0.0,
debug={"reason": "no hits"},
)
scores = [float(getattr(h, "score", 0.0)) for h in hits]
top_score = scores[0]
bottom_score = scores[-1]
score_range = top_score - bottom_score
# ─────────────────────────────────────────────────────────────────
# τ — top-hit strength
# ─────────────────────────────────────────────────────────────────
tau = _clip01(top_score / max(strong_threshold, eps))
# ─────────────────────────────────────────────────────────────────
# ψ — score concentration via CV
# ─────────────────────────────────────────────────────────────────
if len(scores) >= 2:
mean_s = sum(scores) / len(scores)
var_s = sum((s - mean_s) ** 2 for s in scores) / len(scores)
std_s = math.sqrt(var_s)
cv = std_s / (abs(mean_s) + eps)
# We WANT high CV → map directly to ψ (capping at 1.0)
psi = _clip01(cv)
else:
psi = 0.5
# ─────────────────────────────────────────────────────────────────
# φ — topical agreement among top-k
# ─────────────────────────────────────────────────────────────────
# Compute pairwise lexical overlap (Jaccard) as a cheap, dependency-
# free topic-proxy. If the encoder is available we could use cosine,
# but we already know it gives ≥0.93 for any legal-Hebrew pair —
# not useful for discrimination here. Token Jaccard works better.
import re as _re
HEB = _re.compile(r"[א-ת]+")
top_k = min(len(hits), 10)
token_sets = []
for h in hits[:top_k]:
text = getattr(getattr(h, "chunk", None), "text", "") or ""
toks = set(t for t in HEB.findall(text) if len(t) >= 3)
token_sets.append(toks)
if len(token_sets) >= 2:
sims = []
for i in range(len(token_sets)):
for j in range(i + 1, len(token_sets)):
a, b = token_sets[i], token_sets[j]
u = a | b
if u:
sims.append(len(a & b) / len(u))
else:
sims.append(0.0)
phi = _clip01(sum(sims) / max(len(sims), 1))
# Boost: token Jaccard tends to underestimate topical agreement
# on short legal texts; rescale so 0.30 Jaccard ≈ 0.65 φ
phi = _clip01(phi * 2.2)
else:
phi = 0.5
# ─────────────────────────────────────────────────────────────────
# ξ — score-gap anomaly (LOW ξ = clear shelf, HIGH ξ = smooth decay)
# ─────────────────────────────────────────────────────────────────
if len(scores) >= 3 and score_range > eps:
# max consecutive gap divided by total range
gaps = [scores[i] - scores[i + 1] for i in range(len(scores) - 1)]
max_gap = max(gaps) if gaps else 0.0
# If max_gap is large relative to total range → clear shelf → low ξ
# If max_gap is small (smooth decay) → high ξ
gap_ratio = max_gap / (score_range + eps)
# gap_ratio of 1/(N-1) means perfectly uniform → max ξ
# gap_ratio of >0.5 means strong shelf → ξ ≈ 0
xi = _clip01(1.0 - gap_ratio)
else:
xi = 0.5
# ─────────────────────────────────────────────────────────────────
# Ω — geometric mean
# ─────────────────────────────────────────────────────────────────
α, β, γ, δ = omega_weights
a = max(tau, eps) ** α
b = max(phi, eps) ** β
c = max(psi, eps) ** γ
d = max(1.0 - xi, eps) ** δ
total_weight = α + β + γ + δ
omega = _clip01((a * b * c * d) ** (1.0 / total_weight))
return RetrievalSignals(
tau=round(tau, 4),
psi=round(psi, 4),
phi=round(phi, 4),
xi=round(xi, 4),
omega=round(omega, 4),
n_results=len(hits),
top_score=top_score,
score_range=score_range,
debug={
"scores_first_5": [round(s, 3) for s in scores[:5]],
"scores_last_5": [round(s, 3) for s in scores[-5:]],
"weights": list(omega_weights),
"omega_components": {
"tau_pow": round(a, 4),
"phi_pow": round(b, 4),
"psi_pow": round(c, 4),
"1-xi_pow": round(d, 4),
},
},
)
__all__ = ["RetrievalSignals", "compute_retrieval_signals"]
|