Function build_learning_signals

agents/feedback-agent/main.py:436–475

Generate lightweight feedback insights aligned with the PDF interaction.

(selected: Set[int], skipped: Set[int], papers: List[Dict])

def build_learning_signals(selected: Set[int], skipped: Set[int], papers: List[Dict]) -> List[str]:
    """Generate lightweight feedback insights aligned with the PDF interaction."""
    selected_categories = {}
    skipped_categories = {}

    for paper_num in selected:
        paper = papers[paper_num - 1]
        category = paper.get("category", "unknown")
        selected_categories[category] = selected_categories.get(category, 0) + 1

    for paper_num in skipped:
        paper = papers[paper_num - 1]
        category = paper.get("category", "unknown")
        skipped_categories[category] = skipped_categories.get(category, 0) + 1

    signals = []

    lower_bucket_selected = selected_categories.get("maybe_interested", 0) + selected_categories.get("edge_relevant", 0)
    if lower_bucket_selected:
        signals.append(
            f"✓ 你选中了 {lower_bucket_selected} 篇原本靠后的候选，对应主题会被视为更强正信号"
        )

    top_bucket_skipped = skipped_categories.get("must_read", 0) + skipped_categories.get("high_relevant", 0)
    if top_bucket_skipped:
        signals.append(
            f"✓ 你这次跳过了 {top_bucket_skipped} 篇 🔒/🔴 候选，我会先按弱负信号处理，不会立刻重罚"
        )

    if not selected:
        signals.append("✓ 今天的“都不看”会被记为一次弱负信号，只有连续多天跳过同类论文才会明显降权")

    if selected_categories:
        dominant_category = max(selected_categories.items(), key=lambda item: item[1])[0]
        dominant_label = CATEGORY_LABELS.get(dominant_category, dominant_category)
        signals.append(f"✓ 本次选择主要集中在“{dominant_label}”分组，后续我会优先沿这条线细化推荐")

    signals.extend(_build_contrastive_signals(selected, skipped, papers))

    return signals[:4]


def format_selection_summary(
    selected: Set[int],
    total_papers: int,