File size: 6,422 Bytes
914e970 757beb1 914e970 | 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 | """Self-evaluation system for the DRIFT companion bot.
Tracks confidence, auto-critiques responses, detects hallucinations,
and learns from user corrections.
"""
import re
import sqlite3
from datetime import datetime
from pathlib import Path
from typing import Any, Dict, List, Optional
from infj_bot.core.config import DATA_DIR
EVAL_DB = DATA_DIR / "self_eval.db"
# Hallucination markers — heuristic patterns that suggest uncertain claims
HALLUCINATION_MARKERS = [
re.compile(r"\b(I know|I am certain|definitely|absolutely|certainly)\b", re.I),
re.compile(r"\b(always|never|everyone|no one)\b", re.I),
re.compile(r"\b(research shows|studies prove|it is well known)\b", re.I),
]
# Uncertainty markers — these are actually good! They show calibration.
UNCERTAINTY_MARKERS = [
re.compile(r"\b(I think|I believe|probably|likely|maybe|perhaps|it seems)\b", re.I),
re.compile(r"\b(not sure|uncertain|could be|might|may)\b", re.I),
]
class SelfEvaluator:
def __init__(self, db_path: Optional[Path] = None):
self.db_path = str(db_path or EVAL_DB)
self._init_db()
def _init_db(self):
with sqlite3.connect(self.db_path) as conn:
conn.execute(
"""
CREATE TABLE IF NOT EXISTS evaluations (
id INTEGER PRIMARY KEY AUTOINCREMENT,
timestamp TEXT NOT NULL,
prompt TEXT NOT NULL,
response TEXT NOT NULL,
confidence REAL NOT NULL DEFAULT 0.5,
uncertainty_score REAL NOT NULL DEFAULT 0.0,
hallucination_score REAL NOT NULL DEFAULT 0.0,
correction TEXT,
user_rating INTEGER
)
"""
)
conn.commit()
def evaluate(self, prompt: str, response: str) -> Dict[str, float]:
"""Heuristic evaluation of a response. Returns scores."""
# Count hallucination markers (bad)
hall_count = sum(1 for p in HALLUCINATION_MARKERS if p.search(response))
hallucination_score = min(1.0, hall_count * 0.25)
# Count uncertainty markers (good — shows calibration)
unc_count = sum(1 for p in UNCERTAINTY_MARKERS if p.search(response))
uncertainty_score = min(1.0, unc_count * 0.15)
# Confidence heuristic: high certainty words minus uncertainty words, clamped
confidence = 1.0 - hallucination_score + (uncertainty_score * 0.3)
# Dual-channel confidence adjustment: if prompt signals low epistemic / communicative confidence,
# adjust the confidence score downwards.
if "[Communicative Confidence: LOW" in prompt or "Communicative Confidence: LOW" in prompt:
confidence = min(0.35, confidence - 0.4)
confidence = max(0.1, min(0.95, confidence))
return {
"confidence": round(confidence, 2),
"uncertainty_score": round(uncertainty_score, 2),
"hallucination_score": round(hallucination_score, 2),
}
def record(
self, prompt: str, response: str, scores: Dict[str, float], correction: str = ""
) -> int:
"""Store an evaluation record. Returns the record id."""
with sqlite3.connect(self.db_path) as conn:
cur = conn.execute(
"""
INSERT INTO evaluations (timestamp, prompt, response, confidence, uncertainty_score, hallucination_score, correction)
VALUES (?, ?, ?, ?, ?, ?, ?)
""",
(
datetime.now().isoformat(),
prompt[:2000],
response[:4000],
scores["confidence"],
scores["uncertainty_score"],
scores["hallucination_score"],
correction,
),
)
conn.commit()
return cur.lastrowid or 0
def recent_stats(self, limit: int = 50) -> Dict[str, Any]:
"""Return aggregate stats over recent evaluations."""
with sqlite3.connect(self.db_path) as conn:
rows = conn.execute(
"""
SELECT confidence, uncertainty_score, hallucination_score, correction
FROM evaluations
ORDER BY timestamp DESC
LIMIT ?
""",
(limit,),
).fetchall()
if not rows:
return {"count": 0}
confs = [r[0] for r in rows]
halls = [r[2] for r in rows]
corrections = [r[3] for r in rows if r[3]]
return {
"count": len(rows),
"avg_confidence": round(sum(confs) / len(confs), 2),
"avg_hallucination": round(sum(halls) / len(halls), 2),
"high_confidence_pct": round(
sum(1 for c in confs if c > 0.7) / len(confs) * 100, 1
),
"corrections_count": len(corrections),
}
def recent_evaluations(self, limit: int = 10) -> List[Dict]:
with sqlite3.connect(self.db_path) as conn:
conn.row_factory = sqlite3.Row
rows = conn.execute(
"SELECT * FROM evaluations ORDER BY timestamp DESC LIMIT ?",
(limit,),
).fetchall()
return [dict(r) for r in rows]
def suggest_uncertainty_labels(self, response: str) -> List[str]:
"""Suggest uncertainty labels to prepend to a response."""
scores = self.evaluate("", response)
labels = []
if scores["hallucination_score"] > 0.5:
labels.append("[high confidence claim — verify if important]")
elif scores["hallucination_score"] > 0.25:
labels.append("[speculative]")
if scores["uncertainty_score"] >= 0.3:
labels.append("[well-calibrated uncertainty]")
if not labels and scores["confidence"] < 0.5:
labels.append("[uncertain — treat as tentative]")
return labels
def clear_old(self, max_age_days: int = 90) -> int:
cutoff = (
datetime.now() - __import__("datetime").timedelta(days=max_age_days)
).isoformat()
with sqlite3.connect(self.db_path) as conn:
cur = conn.execute("DELETE FROM evaluations WHERE timestamp < ?", (cutoff,))
conn.commit()
return cur.rowcount
|