Spaces:
Runtime error
Runtime error
File size: 9,497 Bytes
f9a9b47 | 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 | """Offline golden-scenario evaluation for the P1 elder-paperwork demo.
The evaluator is intentionally local and transparent: it indexes the bundled
markdown manuals into SQLite, queries the same lightweight token retrieval path
used by the app, and reports the actual retrieved sections instead of fabricating
hits. In offline demo mode, no external model calls are made.
"""
from __future__ import annotations
import json
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Any
from .demo_pack import ingest_demo_pack
from .demo_packs import load_demo_pack
from .storage import SQLiteStore, init_db
SAFE_TERMS = (
"safety",
"shutdown",
"meter",
"isolate",
"energized",
"lockout",
"disconnect",
"emergency",
)
@dataclass
class EvalResult:
scenario_id: str
query: str
top_sections: list[dict[str, Any]]
expected_section_ids: list[int]
expected_section_titles: list[str]
hit_top3: bool
safety_present: bool
sufficient: bool
@dataclass(frozen=True)
class IndexedSection:
title: str
text: str
source_file: str
manual_title: str
section_index: int
def load_scenarios(pack_dir: str | Path) -> list[dict[str, Any]]:
pack_dir = Path(pack_dir)
with open(pack_dir / "golden_scenarios.json", "r", encoding="utf-8") as f:
payload = json.load(f)
if isinstance(payload, dict):
scenarios = payload.get("scenarios", [])
return list(scenarios) if isinstance(scenarios, list) else []
return list(payload)
def _norm(text: str) -> str:
return " ".join((text or "").lower().split())
def _manuals_root(pack_dir: Path) -> Path:
manuals_dir = pack_dir / "manuals"
if manuals_dir.exists():
return manuals_dir
return pack_dir
def _parse_manual_sections(manual_path: Path) -> list[IndexedSection]:
text = manual_path.read_text(encoding="utf-8")
lines = text.splitlines()
doc_title = manual_path.stem.replace("_", " ").title()
for line in lines:
if line.startswith("# "):
doc_title = line[2:].strip()
break
sections: list[IndexedSection] = []
current_title = "Overview"
current_lines: list[str] = []
seen_heading = False
def flush() -> None:
nonlocal current_lines, current_title
section_text = "\n".join(line.rstrip() for line in current_lines).strip()
if section_text:
sections.append(
IndexedSection(
title=current_title,
text=section_text,
source_file=manual_path.name,
manual_title=doc_title,
section_index=len(sections) + 1,
)
)
current_lines = []
for line in lines:
if line.startswith("# "):
continue
if line.startswith("## "):
if seen_heading or current_lines:
flush()
current_title = line[3:].strip() or "Untitled section"
seen_heading = True
continue
current_lines.append(line)
flush()
if not sections:
sections.append(
IndexedSection(
title=doc_title,
text=text.strip(),
source_file=manual_path.name,
manual_title=doc_title,
section_index=1,
)
)
return sections
def _index_manual_sections(store: SQLiteStore, pack_dir: Path, project: str) -> list[dict[str, Any]]:
indexed: list[dict[str, Any]] = []
for manual_path in sorted(_manuals_root(pack_dir).glob("*.md")):
for section in _parse_manual_sections(manual_path):
payload = {
"manual_title": section.manual_title,
"manual_file": section.source_file,
"section_title": section.title,
"section_index": section.section_index,
}
record_id = store.store_record(
project,
pack_dir.name,
f"{section.manual_title} :: {section.title}",
section.text,
payload,
)
store.store_embedding(
record_id,
project,
f"{section.manual_title} {section.title} {section.text}",
metadata={"manual_file": section.source_file, "section_title": section.title},
)
indexed.append({"record_id": record_id, **payload, "primary_text": section.text})
return indexed
def _matches_expected(title: str, expected_titles: list[str]) -> bool:
normalized = _norm(title)
for expected in expected_titles:
expected_norm = _norm(expected)
if expected_norm and (expected_norm == normalized or expected_norm in normalized or normalized in expected_norm):
return True
return False
def _safety_observed(title: str, text: str) -> bool:
haystack = f"{title}\n{text}".lower()
return any(term in haystack for term in SAFE_TERMS)
def _search_ranked_sections(store: SQLiteStore, project: str, query: str, limit: int = 5) -> list[dict[str, Any]]:
index = store._embedding_index(project)
scored = index.search(query, limit=limit)
ranked: list[dict[str, Any]] = []
for rank, (record_id, score) in enumerate(scored, start=1):
record = store.get_record(record_id)
if not record:
continue
payload = json.loads(record["json_blob"])
ranked.append(
{
"rank": rank,
"record_id": record_id,
"score": round(float(score), 3),
"title": payload.get("section_title") or record["title"],
"citation": f'{payload.get("manual_file", "manual")} :: {payload.get("section_title") or record["title"]}',
"excerpt": record["primary_text"][:220],
"manual_title": payload.get("manual_title", ""),
"section_index": payload.get("section_index"),
}
)
return ranked
def evaluate_pack(pack_dir: str | Path, db_path: str | Path | None = None) -> dict[str, Any]:
pack_dir = Path(pack_dir)
db_path = Path(db_path or Path("app_data.sqlite3"))
init_db(db_path)
ingest_demo_pack(pack_dir, db_path=db_path, reset=True)
pack = load_demo_pack(pack_dir)
scenarios = load_scenarios(pack_dir)
store = SQLiteStore(db_path, db_path.parent / "artifacts")
try:
retrieval_project = f"{pack.project}_eval"
_index_manual_sections(store, pack_dir, project=retrieval_project)
results: list[EvalResult] = []
for scenario in scenarios:
query_parts = [scenario.get("symptom", "")]
if scenario.get("equipment_type"):
query_parts.append(str(scenario["equipment_type"]))
if scenario.get("notes"):
query_parts.append(str(scenario["notes"]))
query = " ".join(part for part in query_parts if part).strip()
top_sections = _search_ranked_sections(store, retrieval_project, query, limit=5)
expected_titles = [str(title) for title in scenario.get("expected_section_titles", [])]
top_three = top_sections[:3]
matched_titles = [section["title"] for section in top_three if _matches_expected(section["title"], expected_titles)]
hit_top3 = bool(matched_titles)
safety_present = any(_safety_observed(section["title"], section["excerpt"]) for section in top_sections)
sufficient = not bool(scenario.get("requires_insufficient", False))
expected_section_ids = [section["rank"] for section in top_three if _matches_expected(section["title"], expected_titles)]
results.append(
EvalResult(
scenario_id=str(scenario["scenario_id"]),
query=query,
top_sections=top_sections,
expected_section_ids=expected_section_ids,
expected_section_titles=expected_titles,
hit_top3=hit_top3,
safety_present=safety_present,
sufficient=sufficient,
)
)
total = len(results)
top3_hits = sum(1 for result in results if result.hit_top3)
safety_hits = sum(1 for result in results if result.safety_present)
insufficient_cases = sum(1 for result in results if not result.sufficient)
return {
"pack": str(pack_dir),
"pack_id": pack.pack_id,
"scenario_count": total,
"top3_hit_rate": round(top3_hits / total if total else 0.0, 3),
"safety_presence_rate": round(safety_hits / total if total else 0.0, 3),
"insufficient_cases": insufficient_cases,
"retrieval_project": retrieval_project,
"results": [asdict(result) for result in results],
}
finally:
store.close()
def main() -> None:
import argparse
parser = argparse.ArgumentParser(description="Evaluate P1 elder-paperwork golden scenarios")
parser.add_argument("--pack", required=True, help="Path to demo pack")
parser.add_argument("--db", default=None, help="SQLite database path")
args = parser.parse_args()
report = evaluate_pack(args.pack, db_path=args.db)
print(json.dumps(report, indent=2))
if __name__ == "__main__":
main()
|