scripts/score_predictions.py

#!/usr/bin/env python3
"""Score arbitrary Builder outputs without requiring a canonical surface schema.

Answer and Efficiency reuse the official WorkSurface-Bench implementations.
Evidence is projected back to raw Workspace-Bench files, so a Builder is not
penalized for choosing different chunks, table names, schemas, or graph IDs.
"""

from __future__ import annotations

import argparse
import json
import re
import sys
from pathlib import Path
from typing import Any


HERE = Path(__file__).resolve().parent
PROJECT = HERE.parent
DEFAULT_OFFICIAL_REPO = PROJECT / "official_worksurface_bench"
DEFAULT_WSB_LOCK = DEFAULT_OFFICIAL_REPO / "data" / "wsb_lock.json"


def read_jsonl(path: Path) -> list[dict[str, Any]]:
    return [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines() if line]


def norm_filename(value: str) -> str:
    return re.sub(r"[^a-z0-9]+", "", Path(value).name.lower())


def strip_canonical_prefix(value: str, source_id: str) -> str:
    name = Path(value).name
    name = re.sub(rf"^t{re.escape(source_id)}__", "", name)
    if name.endswith(".md"):
        name = name[:-3]
    return name


def match_raw_name(candidate: str, raw_names: list[str]) -> str | None:
    key = norm_filename(candidate)
    exact = {norm_filename(name): name for name in raw_names}
    if key in exact:
        return exact[key]
    matches = [
        name for name in raw_names
        if key.startswith(norm_filename(Path(name).stem))
        or norm_filename(Path(name).stem).startswith(key)
    ]
    if not matches:
        return None
    return max(matches, key=lambda name: len(norm_filename(Path(name).stem)))


def table_to_raw(table: str, source_id: str, raw_names: list[str]) -> str | None:
    candidate = re.sub(rf"^t{re.escape(source_id)}__", "", table)
    table_key = norm_filename(candidate)
    matches = [
        name for name in raw_names
        if table_key.startswith(norm_filename(Path(name).stem))
    ]
    return max(matches, key=lambda name: len(norm_filename(Path(name).stem))) if matches else None


def gold_raw_files(evidence: dict[str, Any], source_id: str, raw_names: list[str]) -> tuple[list[str], bool]:
    """Return evaluator raw files and whether every file is required."""
    complete = evidence.get("verified_complete_set") or evidence.get("verified_required_tabular_inputs")
    if complete:
        mapped = [match_raw_name(str(item).split("::", 1)[-1], raw_names) for item in complete]
        return [item for item in mapped if item], True
    if evidence.get("source_file"):
        mapped = match_raw_name(str(evidence["source_file"]), raw_names)
        return ([mapped] if mapped else []), False
    if evidence.get("surface") == "rag" and evidence.get("file"):
        mapped = match_raw_name(strip_canonical_prefix(str(evidence["file"]), source_id), raw_names)
        return ([mapped] if mapped else []), False
    path = evidence.get("graph_path") or []
    if path:
        terminal = str(path[-1]).split("::", 1)[-1]
        mapped = match_raw_name(terminal, raw_names)
        if mapped:
            return [mapped], False
    if evidence.get("table"):
        mapped = table_to_raw(str(evidence["table"]), source_id, raw_names)
        return ([mapped] if mapped else []), False
    return [], False


def predicted_raw_files(trace: dict[str, Any], source_id: str, raw_names: list[str]) -> set[str]:
    candidates: list[str] = []
    for ref in trace.get("evidence_refs", []):
        if ref.get("source_file"):
            candidates.append(str(ref["source_file"]))
    candidates.extend(map(str, trace.get("rag_files", [])))
    for node in trace.get("graph_nodes", []):
        candidates.append(str(node).split("::", 1)[-1])
    for ref in trace.get("table_sources", []):
        if isinstance(ref, dict) and ref.get("source_file"):
            candidates.append(str(ref["source_file"]))
        elif isinstance(ref, str):
            candidates.append(ref)
    mapped = {
        match_raw_name(strip_canonical_prefix(candidate, source_id), raw_names)
        for candidate in candidates
    }
    return {item for item in mapped if item}


def score_source_evidence(task: dict[str, Any], trace: dict[str, Any], lock: dict[str, Any]) -> dict[str, Any]:
    source_id = str(task["source"]["task_id"])
    raw_names = list(lock["task_id_to_file_hashes"][source_id])
    predicted = predicted_raw_files(trace, source_id, raw_names)
    outcomes: list[tuple[str, bool]] = []
    unscorable = 0
    for evidence in task.get("gold_evidence", []):
        expected, require_all = gold_raw_files(evidence, source_id, raw_names)
        if not expected:
            unscorable += 1
            continue
        hit = all(item in predicted for item in expected) if require_all else any(item in predicted for item in expected)
        outcomes.append((evidence["surface"], hit))
    per_surface: dict[str, float] = {}
    for surface in sorted({surface for surface, _ in outcomes}):
        values = [hit for current, hit in outcomes if current == surface]
        per_surface[surface] = round(sum(values) / len(values), 4)
    score = sum(hit for _, hit in outcomes) / len(outcomes) if outcomes else None
    total = len(outcomes) + unscorable
    return {
        "score": round(score, 4) if score is not None else None,
        "per_surface": per_surface,
        "scorable_items": len(outcomes),
        "unscorable_items": unscorable,
        "coverage": round(len(outcomes) / total, 4) if total else 0.0,
        "predicted_raw_files": sorted(predicted),
    }


def mean(values: list[float | None]) -> float | None:
    present = [value for value in values if value is not None]
    return round(sum(present) / len(present), 4) if present else None


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--gold", type=Path, required=True)
    parser.add_argument("--predictions", type=Path, required=True)
    parser.add_argument("--output", type=Path, required=True)
    parser.add_argument("--official-repo", type=Path, default=DEFAULT_OFFICIAL_REPO)
    parser.add_argument("--wsb-lock", type=Path, default=DEFAULT_WSB_LOCK)
    args = parser.parse_args()
    sys.path.insert(0, str(args.official_repo))
    from scoring.answer import score_answer  # type: ignore
    from scoring.efficiency_safety import score_efficiency  # type: ignore

    tasks = read_jsonl(args.gold)
    traces = {row["id"]: row for row in read_jsonl(args.predictions)}
    lock = json.loads(args.wsb_lock.read_text(encoding="utf-8"))
    rows: list[dict[str, Any]] = []
    for task in tasks:
        trace = traces.get(task["id"])
        if trace is None:
            continue
        answer = score_answer(task, trace.get("answer"), anchors=task.get("qualitative_anchors"))
        evidence = score_source_evidence(task, trace, lock)
        efficiency = score_efficiency(trace.get("total_tokens", 0), task.get("efficiency_budget_tokens"))
        if evidence["score"] is None:
            # A small set of graph-only tasks has no raw-file-addressable gold
            # evidence.  Do not force a canonical graph schema merely to score
            # them; renormalize Answer and Efficiency for those tasks.
            utility = (0.55 * answer.score + 0.10 * efficiency) / 0.65
        else:
            utility = 0.55 * answer.score + 0.35 * evidence["score"] + 0.10 * efficiency
        rows.append({
            "id": task["id"],
            "source_task_id": str(task["source"]["task_id"]),
            "task_type": task["task_type"],
            "answer": {"score": answer.score, "detail": answer.detail},
            "source_evidence": evidence,
            "efficiency": efficiency,
            "query_utility": round(utility, 4),
        })

    overall = {
        "n": len(rows),
        "answer": mean([row["answer"]["score"] for row in rows]),
        "source_evidence": mean([row["source_evidence"]["score"] for row in rows]),
        "source_evidence_coverage": mean([row["source_evidence"]["coverage"] for row in rows]),
        "efficiency": mean([row["efficiency"] for row in rows]),
        "query_utility": mean([row["query_utility"] for row in rows]),
        "missing_predictions": len(tasks) - len(rows),
    }
    by_type = {}
    for task_type in sorted({row["task_type"] for row in rows}):
        subset = [row for row in rows if row["task_type"] == task_type]
        by_type[task_type] = {
            "n": len(subset),
            "answer": mean([row["answer"]["score"] for row in subset]),
            "source_evidence": mean([row["source_evidence"]["score"] for row in subset]),
            "query_utility": mean([row["query_utility"] for row in subset]),
        }
    report = {
        "scorer": "worksurface_build_source_grounded_v0.1",
        "weights": {"answer": 0.55, "source_evidence": 0.35, "efficiency": 0.10},
        "overall": overall,
        "by_task_type": by_type,
        "per_task": rows,
    }
    args.output.parent.mkdir(parents=True, exist_ok=True)
    args.output.write_text(json.dumps(report, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
    print(json.dumps(overall, ensure_ascii=False, indent=2))


if __name__ == "__main__":
    main()