Blame for tools/audit_coverage.py

#!/usr/bin/env python3
"""Measure answer-string coverage and question-query retrieval recall.

These are noisy diagnostics, not answer accuracy: aliases and paraphrases can
be valid when the normalized gold string is absent, while common strings can
occur in irrelevant documents.
"""

from __future__ import annotations

import argparse
import json
import re
import unicodedata
from collections import Counter, defaultdict
from functools import lru_cache
from pathlib import Path
from typing import Any

try:
    from .common import read_jsonl, write_jsonl
    from .corpus_db import connect, fts_counts, fts_exists, search_docids
except ImportError:  # direct script execution
    from common import read_jsonl, write_jsonl
    from corpus_db import connect, fts_counts, fts_exists, search_docids


def normalize_for_match(text: str) -> str:
    decomposed = unicodedata.normalize("NFKD", text or "")
    without_marks = "".join(character for character in decomposed if not unicodedata.combining(character))
    return re.sub(r"[^\w]+", " ", without_marks.casefold(), flags=re.UNICODE).strip()


def is_lexically_evaluable(answer: str) -> bool:
    normalized = normalize_for_match(answer)
    return normalized not in {"yes", "no", "true", "false"} and len(normalized.replace(" ", "")) >= 3


def contains_answer(text: str, answer: str) -> bool:
    needle = normalize_for_match(answer)
    return bool(needle) and needle in normalize_for_match(text)


def provenance_documents(connection, qid: str) -> list[dict[str, Any]]:
    return [
        dict(row)
        for row in connection.execute(
            """
            SELECT d.url_id AS docid, d.canonical_url AS url
            FROM document_qids AS q JOIN documents AS d USING(url_id)
            WHERE q.qid=? AND d.status='success'
            """,
            (qid,),
        )
    ]


def failure_category(source: str, error: str) -> str:
    value = f"{source} {error}".casefold()
    if "robots.txt disallows" in value:
        return "robots_disallowed"
    if "robots.txt" in value:
        return "robots_unavailable_or_invalid"
    if "challenge" in value or "captcha" in value or "security verification" in value:
        return "challenge_or_captcha"
    status = re.search(r"http\s+([45]\d\d)", value)
    if status:
        return f"http_{status.group(1)}"
    if "timeout" in value:
        return "timeout"
    if "too large" in value or "exceeded byte limit" in value:
        return "response_too_large"
    if "unsupported content" in value or "non-text content" in value:
        return "unsupported_or_non_text"
    if "pdf" in value:
        return "pdf_extraction"
    if "extracted only" in value:
        return "too_little_text"
    if "connect" in value or "name or service not known" in value:
        return "connection_or_dns"
    return "other"


def corpus_statistics(connection) -> dict[str, Any]:
    status = dict(connection.execute("SELECT status, count(*) FROM documents GROUP BY status").fetchall())
    sources = dict(
        connection.execute(
            "SELECT source, count(*) FROM documents WHERE status='success' GROUP BY source ORDER BY count(*) DESC"
        ).fetchall()
    )
    content_types = dict(
        connection.execute(
            """
            SELECT content_type, count(*)
            FROM documents
            WHERE status='success'
            GROUP BY content_type
            ORDER BY count(*) DESC
            LIMIT 20
            """
        ).fetchall()
    )
    text = connection.execute(
        """
        SELECT count(*) AS documents, coalesce(sum(length(text)), 0) AS characters,
               count(DISTINCT content_hash) AS unique_hashes
        FROM documents WHERE status='success'
        """
    ).fetchone()
    failure_categories: Counter[str] = Counter()
    failure_details = []
    for row in connection.execute(
        """
        SELECT source, error, count(*) AS count
        FROM documents WHERE status='failed'
        GROUP BY source, error
        ORDER BY count(*) DESC
        """
    ):
        failure_categories[failure_category(row["source"], row["error"])] += row["count"]
        if len(failure_details) < 30:
            failure_details.append(dict(row))
    index_counts = None
    if fts_exists(connection):
        index_counts = fts_counts(connection)
    return {
        "status": status,
        "successful_sources": sources,
        "successful_content_types_top_20": content_types,
        "successful_text_characters": int(text["characters"]),
        "successful_unique_content_hashes": int(text["unique_hashes"]),
        "exact_duplicate_successes_excluded_from_fts": int(text["documents"] - text["unique_hashes"]),
        "fts_documents": index_counts["documents"] if index_counts else None,
        "fts_chunks": index_counts["chunks"] if index_counts else None,
        "failure_categories": dict(failure_categories.most_common()),
        "top_failure_details": failure_details,
    }


def audit(
    database: Path,
    data_paths: list[Path],
    *,
    k: int,
    limit: int = 0,
    only_with_fetched_provenance: bool = False,
) -> tuple[dict, list[dict]]:
    samples = [record for path in data_paths for record in read_jsonl(path)]
    connection = connect(database, readonly=True)

    @lru_cache(maxsize=32768)
    def normalized_document(docid: str) -> str:
        row = connection.execute(
            "SELECT text FROM documents WHERE url_id=? AND status='success'",
            (docid,),
        ).fetchone()
        return normalize_for_match(row["text"]) if row else ""

    provenance_by_qid: defaultdict[str, list[dict[str, Any]]] = defaultdict(list)
    provenance_status_by_qid: defaultdict[str, Counter[str]] = defaultdict(Counter)
    for row in connection.execute(
        """
        SELECT q.qid, d.url_id AS docid, d.canonical_url AS url, d.status
        FROM document_qids AS q JOIN documents AS d USING(url_id)
        """
    ):
        provenance_status_by_qid[row["qid"]][row["status"]] += 1
        if row["status"] == "success":
            provenance_by_qid[row["qid"]].append({"docid": row["docid"], "url": row["url"]})

    input_samples = len(samples)
    if only_with_fetched_provenance:
        fetched_qids = set(provenance_by_qid)
        samples = [
            sample
            for sample in samples
            if (sample["metadata"].get("manifest_qid") or sample["metadata"]["qid"]) in fetched_qids
        ]
    if limit:
        samples = samples[:limit]
    results: list[dict[str, Any]] = []
    aggregate: Counter[str] = Counter()
    by_difficulty: defaultdict[str, Counter[str]] = defaultdict(Counter)

    for sample in samples:
        metadata = sample["metadata"]
        qid = metadata.get("manifest_qid") or metadata["qid"]
        question = metadata["question"]
        answer = metadata.get("answer") or sample.get("label", "")
        evaluable = is_lexically_evaluable(answer)
        normalized_answer = normalize_for_match(answer)
        provenance = provenance_by_qid[qid]
        provenance_status = provenance_status_by_qid[qid]
        provenance_hit = evaluable and any(
            normalized_answer in normalized_document(document["docid"]) for document in provenance
        )
        try:
            retrieved_docids = search_docids(connection, question, k)
        except ValueError:
            retrieved_docids = []
        retrieval_hit = evaluable and any(
            normalized_answer in normalized_document(docid) for docid in retrieved_docids
        )
        result = {
            "qid": metadata["qid"],
            "manifest_qid": qid,
            "difficulty": int(metadata["difficulty"]),
            "lexically_evaluable": evaluable,
            "manifest_provenance_documents": sum(provenance_status.values()),
            "fetched_provenance_documents": len(provenance),
            "failed_provenance_documents": provenance_status["failed"],
            "pending_provenance_documents": provenance_status["pending"],
            "answer_in_fetched_provenance": bool(provenance_hit),
            f"answer_in_question_search_top_{k}": bool(retrieval_hit),
            "retrieved_documents": len(retrieved_docids),
        }
        results.append(result)
        counters = (aggregate, by_difficulty[str(metadata["difficulty"])])
        for counter in counters:
            counter["questions"] += 1
            counter["lexically_evaluable"] += evaluable
            counter["has_manifest_provenance"] += bool(provenance_status)
            counter["has_fetched_provenance"] += bool(provenance)
            counter["answer_in_fetched_provenance"] += provenance_hit
            counter[f"answer_in_question_search_top_{k}"] += retrieval_hit

    def summarize(counter: Counter[str]) -> dict[str, Any]:
        evaluable = counter["lexically_evaluable"]
        return {
            **dict(counter),
            "provenance_answer_coverage": counter["answer_in_fetched_provenance"] / evaluable if evaluable else None,
            f"question_search_recall_at_{k}": counter[f"answer_in_question_search_top_{k}"] / evaluable
            if evaluable
            else None,
        }

    summary = {
        "database": str(database),
        "top_k": k,
        "input_questions": input_samples,
        "selected_questions": len(samples),
        "only_with_fetched_provenance": only_with_fetched_provenance,
        "corpus": corpus_statistics(connection),
        "normalized_document_cache": {
            "hits": normalized_document.cache_info().hits,
            "misses": normalized_document.cache_info().misses,
            "maxsize": normalized_document.cache_info().maxsize,
            "currsize": normalized_document.cache_info().currsize,
        },
        "overall": summarize(aggregate),
        "by_difficulty": {difficulty: summarize(counter) for difficulty, counter in sorted(by_difficulty.items())},
        "interpretation": (
            "Normalized answer-string matching is a noisy proxy, not a lower or upper bound. "
            "It excludes yes/no answers and misses aliases or paraphrases, while common answers can occur incidentally. "
            "Provenance coverage asks whether a fetched teacher-visited page contains the answer; retrieval recall asks whether "
            "the original question retrieves a document containing the answer string. When only_with_fetched_provenance is "
            "true, the report "
            "is a fetch/search smoke test and is selection-biased, not a corpus-wide coverage estimate."
        ),
    }
    connection.close()
    return summary, results


def main() -> None:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--database", type=Path, required=True)
    parser.add_argument("--data", type=Path, nargs="+", required=True, help="prepared train/test JSONL")
    parser.add_argument("--top-k", type=int, default=20)
    parser.add_argument("--limit", type=int, default=0)
    parser.add_argument(
        "--only-with-fetched-provenance",
        action="store_true",
        help="biased smoke mode: audit only qids having at least one successfully fetched visited page",
    )
    parser.add_argument("--output-dir", type=Path, required=True)
    args = parser.parse_args()
    if args.top_k < 1 or args.limit < 0:
        parser.error("--top-k must be >= 1 and --limit must be non-negative")
    summary, per_question = audit(
        args.database,
        args.data,
        k=args.top_k,
        limit=args.limit,
        only_with_fetched_provenance=args.only_with_fetched_provenance,
    )
    args.output_dir.mkdir(parents=True, exist_ok=True)
    (args.output_dir / "summary.json").write_text(
        json.dumps(summary, indent=2, sort_keys=True) + "\n", encoding="utf-8"
    )
    write_jsonl(args.output_dir / "per_question.jsonl", per_question)
    print(json.dumps(summary, indent=2, sort_keys=True))


if __name__ == "__main__":
    main()