Function _is_noisy_pdf_evidence_text

agents/reading-agent/main.py:674–704

(text: Any, *, min_chars: int = 40)

def _is_noisy_pdf_evidence_text(text: Any, *, min_chars: int = 40) -> bool:
    cleaned = _clean_text(text)
    if len(cleaned) < min_chars:
        return True
    alpha_count = len(re.findall(r"[A-Za-z]", cleaned))
    alpha_ratio = alpha_count / max(1, len(cleaned))
    if alpha_ratio < 0.45:
        return True

    lower = cleaned.lower()
    hard_noise_patterns = (
        "corresponding author",
        "this work was supported",
        "all rights reserved",
        "published as a conference paper",
        "arxiv preprint arxiv:",
        "references",
    )
    if any(pattern in lower for pattern in hard_noise_patterns):
        return True

    citation_like = len(re.findall(r"\[\d+\]|\bet al\.\b|\bpages?\s+\d+\b", lower))
    figure_like = len(re.findall(r"\bfig(?:ure)?\.?\s*\d+|\btable\s+\d+", lower))
    if citation_like + figure_like >= 4:
        return True

    symbol_count = len(re.findall(r"[=<>±∑√∞≈≠≤≥_{}^$\\]", cleaned))
    if symbol_count > 12 and symbol_count / max(1, len(cleaned)) > 0.03:
        return True

    return False


def _prefer_candidate_abstract(
    current_abstract: Any,