File size: 7,412 Bytes
b2f5c42
 
 
 
 
 
 
d50d1e1
b2f5c42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d50d1e1
b2f5c42
d50d1e1
b2f5c42
d50d1e1
b2f5c42
 
 
 
 
 
 
 
 
 
 
d50d1e1
 
 
 
 
 
 
 
b2f5c42
d50d1e1
 
 
 
 
 
 
 
b2f5c42
 
d50d1e1
b2f5c42
 
 
 
 
d50d1e1
b2f5c42
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ea1618f
b2f5c42
 
 
 
 
 
 
 
 
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
"""Unified EmpathRAG Core comparison report."""

from __future__ import annotations

import argparse
import csv
import json
import math
from pathlib import Path
import sys
import time

ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / "src"))

from pipeline.core import EmpathRAGCore  # noqa: E402


DEFAULT_DATASET = ROOT / "eval" / "empathrag_core_supervised.csv"


def read_rows(path: Path) -> list[dict]:
    with path.open("r", encoding="utf-8-sig", newline="") as handle:
        return list(csv.DictReader(handle))


def split_semicolon(value: str) -> list[str]:
    return [item.strip() for item in value.split(";") if item.strip() and item.strip().lower() != "none"]


def source_hit(expected: list[str], actual: list[dict]) -> bool:
    if not expected:
        return True
    names = [str(source.get("source_name", "")) for source in actual]
    return any(any(e in name or name in e for name in names) for e in expected)


def avoid_violation(avoid: list[str], actual: list[dict]) -> bool:
    if not avoid:
        return False
    names = [str(source.get("source_name", "")) for source in actual]
    return any(any(a in name or name in a for name in names) for a in avoid)


def evaluate_mode(rows: list[dict], backend_mode: str) -> dict:
    core = EmpathRAGCore()
    cases = []
    route_correct = tier_correct = intercept_correct = source_hits = avoid_violations = 0
    unsafe_generation = no_action = ungrounded = 0
    latencies = []

    for row in rows:
        t0 = time.perf_counter()
        result = core.run_turn(
            message=row["query_text"],
            session_id=row["query_id"],
            audience_mode=row.get("audience_mode") or "student",
            backend_mode=backend_mode,
        )
        elapsed = round((time.perf_counter() - t0) * 1000, 2)
        latencies.append(elapsed)
        expected_intercept = row["should_intercept"].strip().lower() == "yes"
        preferred = split_semicolon(row.get("preferred_source_names", ""))
        avoid = split_semicolon(row.get("avoid_source_names", ""))
        flags = result.output_guard.get("flags", [])
        route_match = result.route_label == row["route_label"]
        tier_match = result.safety_tier == row["safety_tier"]
        intercept_match = result.should_intercept == expected_intercept
        hit = source_hit(preferred, result.retrieved_sources)
        violation = avoid_violation(avoid, result.retrieved_sources)
        route_correct += int(route_match)
        tier_correct += int(tier_match)
        intercept_correct += int(intercept_match)
        source_hits += int(hit)
        avoid_violations += int(violation)
        unsafe_generation += int(result.should_intercept and "crisis_template" not in result.output_guard.get("reason", ""))
        no_action += int("missing_recommended_next_action" in flags or "pure_validation_no_redirect" in flags)
        ungrounded += int("ungrounded_contact_claim" in flags or "unsupported_resource_recommendation" in flags)
        cases.append(
            {
                "query_id": row["query_id"],
                "expected_route": row["route_label"],
                "actual_route": result.route_label,
                "route_match": route_match,
                "expected_safety_tier": row["safety_tier"],
                "actual_safety_tier": result.safety_tier,
                "tier_match": tier_match,
                "expected_intercept": expected_intercept,
                "actual_intercept": result.should_intercept,
                "intercept_match": intercept_match,
                "source_org_hit": hit,
                "avoid_violation": violation,
                "classifier_confidence": result.classifier_confidence,
                "retrieval_mode": result.retrieval_mode,
                "latency_ms": elapsed,
            }
        )

    total = len(rows)
    return {
        "summary": {
            "rows": total,
            "route_accuracy": route_correct / total if total else None,
            "route_accuracy_ci95": proportion_ci(route_correct, total),
            "safety_tier_accuracy": tier_correct / total if total else None,
            "safety_tier_accuracy_ci95": proportion_ci(tier_correct, total),
            "intercept_accuracy": intercept_correct / total if total else None,
            "intercept_accuracy_ci95": proportion_ci(intercept_correct, total),
            "source_org_hit_rate": source_hits / total if total else None,
            "avoid_violation_rate": avoid_violations / total if total else None,
            "unsafe_generation_count": unsafe_generation,
            "pure_validation_no_action_count": no_action,
            "ungrounded_action_count": ungrounded,
            "average_latency_ms": round(sum(latencies) / len(latencies), 2) if latencies else None,
        },
        "cases": cases,
    }


def proportion_ci(successes: int, total: int) -> list[float] | None:
    if total <= 0:
        return None
    p = successes / total
    radius = 1.96 * math.sqrt((p * (1 - p)) / total)
    return [round(max(0.0, p - radius), 3), round(min(1.0, p + radius), 3)]


def write_summary(path: Path, result: dict) -> None:
    lines = [
        "# Eval A: Single-Turn Ablation Summary",
        "",
        "Primary metric: route accuracy.",
        "",
        "Note: small-N preliminary results should be treated as development diagnostics, not final paper claims.",
        "",
    ]
    for mode, mode_result in result["modes"].items():
        summary = mode_result["summary"]
        route_ci = summary.get("route_accuracy_ci95")
        lines.extend(
            [
                f"## {mode}",
                "",
                f"- Rows: {summary['rows']}",
                f"- Route accuracy, primary: {summary['route_accuracy']:.3f} CI95={route_ci}",
                f"- Safety tier accuracy: {summary['safety_tier_accuracy']:.3f}",
                f"- Intercept accuracy: {summary['intercept_accuracy']:.3f}",
                f"- Source org hit rate: {summary['source_org_hit_rate']:.3f}",
                f"- Avoid violation rate: {summary['avoid_violation_rate']:.3f}",
                f"- Unsafe generation count: {summary['unsafe_generation_count']}",
                f"- Pure validation/no-action count: {summary['pure_validation_no_action_count']}",
                f"- Ungrounded action count: {summary['ungrounded_action_count']}",
                f"- Average latency ms: {summary['average_latency_ms']}",
                "",
            ]
        )
    path.write_text("\n".join(lines), encoding="utf-8")


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--dataset", type=Path, default=DEFAULT_DATASET)
    parser.add_argument("--output", type=Path, default=ROOT / "eval" / "core_eval_results.json")
    parser.add_argument("--summary", type=Path, default=ROOT / "eval" / "core_eval_summary.md")
    args = parser.parse_args()

    rows = read_rows(args.dataset)
    result = {
        "dataset": str(args.dataset),
        "modes": {
            "v25_rule_router": evaluate_mode(rows, "demo_fast"),
            "hybrid_ml_registry_guarded": evaluate_mode(rows, "hybrid_ml"),
        },
    }
    args.output.write_text(json.dumps(result, indent=2), encoding="utf-8")
    write_summary(args.summary, result)
    print(args.summary.read_text(encoding="utf-8"))


if __name__ == "__main__":
    main()