Files
llm/scripts/report_repeated_failures.py

109 lines
3.9 KiB
Python

from __future__ import annotations
import argparse
import json
from collections import defaultdict
from pathlib import Path
from typing import Any
ROOT = Path(__file__).resolve().parents[1]
DEFAULT_REPORT_DIR = ROOT / "reports" / "observability"
def iter_turn_audit(report_dir: Path) -> list[dict[str, Any]]:
records: list[dict[str, Any]] = []
if not report_dir.exists():
return records
for path in sorted(report_dir.rglob("turn_audit/*.jsonl")):
with path.open("r", encoding="utf-8") as handle:
for line_number, line in enumerate(handle, start=1):
line = line.strip()
if not line:
continue
try:
record = json.loads(line)
except json.JSONDecodeError as exc:
raise ValueError(f"{path}:{line_number}: invalid JSONL: {exc}") from exc
if isinstance(record, dict):
records.append(record)
return records
def normalize_prompt(text: str) -> str:
return " ".join(str(text or "").split())
def group_repeated_failures(records: list[dict[str, Any]], *, min_count: int = 2, limit: int = 50) -> list[dict[str, Any]]:
groups: dict[tuple[str, str], dict[str, Any]] = defaultdict(
lambda: {
"count": 0,
"failure_type": "",
"normalized_user_text": "",
"examples": [],
"_timestamps": [],
}
)
for record in records:
if str(record.get("outcome") or "") == "success":
continue
failure_type = str(record.get("failure_type") or "unknown")
normalized_user_text = normalize_prompt(str(record.get("user_text") or ""))
if not normalized_user_text:
continue
key = (failure_type, normalized_user_text)
item = groups[key]
item["count"] += 1
item["failure_type"] = failure_type
item["normalized_user_text"] = normalized_user_text
item["_timestamps"].append(str(record.get("timestamp") or ""))
if len(item["examples"]) < 3:
item["examples"].append(
{
"timestamp": record.get("timestamp"),
"turn_id": record.get("turn_id"),
"error_code": record.get("error_code"),
"error_message": record.get("error_message"),
"request_id": record.get("request_id"),
}
)
rows = []
for item in groups.values():
if item["count"] < min_count:
continue
timestamps = sorted(ts for ts in item.pop("_timestamps") if ts)
item["first_seen"] = timestamps[0] if timestamps else ""
item["last_seen"] = timestamps[-1] if timestamps else ""
rows.append(item)
rows.sort(key=lambda row: (-int(row["count"]), str(row["failure_type"]), str(row["normalized_user_text"])))
return rows[:limit]
def write_report(path: Path, payload: dict[str, Any]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
def main() -> int:
parser = argparse.ArgumentParser(description="Report repeated failed turns grouped by failure type and normalized prompt.")
parser.add_argument("--report-dir", type=Path, default=DEFAULT_REPORT_DIR)
parser.add_argument("--min-count", type=int, default=2)
parser.add_argument("--limit", type=int, default=50)
parser.add_argument("--output", type=Path, help="Optional JSON output path.")
args = parser.parse_args()
records = iter_turn_audit(args.report_dir)
groups = group_repeated_failures(records, min_count=args.min_count, limit=args.limit)
payload = {"records": len(records), "repeated_failures": groups}
print(json.dumps(payload, ensure_ascii=False, indent=2))
if args.output:
write_report(args.output, payload)
return 0
if __name__ == "__main__":
raise SystemExit(main())