Skip to content

Open-source AI podcast clipper. Generate vertical clips with face tracking and burned-in captions. CLI, MCP server, web app.

4.4 KB
Raw Blame
"""Attribution (clip ↔ uploaded video) and metric sync.

podcli renders but never publishes, so performance is reconstructed after the
fact: match each rendered clip to its uploaded video, then pull metrics onto
the clip. Matching is proposed, never silent. A wrong link poisons the signal.
"""
from __future__ import annotations

from datetime import datetime, timezone
from difflib import SequenceMatcher
from typing import Any

import sys

from services.clips_history import load_clips_history, save_clips_history, update_clip
from . import client


def _ratio(a: str, b: str) -> float:
    return SequenceMatcher(None, (a or "").lower(), (b or "").lower()).ratio()


def _now() -> str:
    return datetime.now(timezone.utc).isoformat()


def propose_links(limit: int = 50) -> list[dict[str, Any]]:
    """Best upload match per unlinked clip: duration proximity + title + recency."""
    uploads = client.list_uploads(limit=limit)
    proposals = []
    for clip in load_clips_history():
        if clip.get("youtube_video_id"):
            continue
        best, best_score = None, 0.0
        for up in uploads:
            dur_gap = abs((up.get("duration") or 0) - (clip.get("duration") or 0))
            dur_score = max(0.0, 1.0 - dur_gap / 5.0)  # within ~5s
            score = 0.6 * _ratio(clip.get("title", ""), up["title"]) + 0.4 * dur_score
            if score > best_score:
                best, best_score = up, score
        if best and best_score >= 0.4:
            proposals.append({
                "clip_id": clip["id"], "clip_title": clip.get("title"),
                "video_id": best["video_id"], "video_title": best["title"],
                "score": round(best_score, 2),
            })
    return proposals


def set_link(clip_id: str, video_id: str) -> bool:
    return update_clip(clip_id, youtube_video_id=video_id) is not None


def sync_metrics() -> int:
    """Pull live metrics onto every linked clip. Returns the number updated.

    Mutates the loaded list and saves once (not once-per-clip) to shrink the
    window where a concurrent writer could clobber the file. A single clip's
    fetch failure is isolated so it can't abort the whole sync.
    """
    entries = load_clips_history()
    count, failed = 0, 0
    for clip in entries:
        vid = clip.get("youtube_video_id")
        if not vid:
            continue
        try:
            metrics = client.fetch_metrics(vid)
        except Exception as e:
            failed += 1
            print(f"  ! metrics fetch failed for {vid}: {e}", file=sys.stderr)
            continue
        metrics["fetched_at"] = _now()
        clip["metrics"] = metrics
        count += 1
    if count:
        save_clips_history(entries)
    if failed:
        print(f"  ! {failed} clip(s) skipped due to fetch errors", file=sys.stderr)
    _refresh_learnings()
    return count


def _refresh_learnings() -> None:
    """Regenerate the learnings digest; isolated so its failure can't discard saved metrics."""
    try:
        from . import learnings
        learnings.write_learnings()
    except Exception as e:
        print(f"  ! learnings refresh skipped: {e}", file=sys.stderr)


def sync_from_csv(path: str, threshold: float = 0.6) -> dict[str, Any]:
    """Match a YouTube Studio CSV to clips by title and write metrics. No auth."""
    rows = client.parse_analytics_csv(path)
    entries = load_clips_history()
    links: list[dict[str, Any]] = []
    unmatched = []
    for clip in entries:
        best, best_score = None, 0.0
        for row in rows:
            r = _ratio(clip.get("title", ""), row["title"])
            if r > best_score:
                best, best_score = row, r
        if best and best_score >= threshold:
            metrics = {k: best[k] for k in ("views", "retention", "ctr", "impressions") if k in best}
            metrics["fetched_at"] = _now()
            clip["metrics"] = metrics
            # Title-only attribution is fuzzy; surface every pairing + score so a
            # wrong match is visible rather than silently poisoning the signal.
            links.append({
                "clip_title": clip.get("title"), "row_title": best["title"],
                "score": round(best_score, 2),
            })
        else:
            unmatched.append(clip.get("title"))
    if links:
        save_clips_history(entries)
    _refresh_learnings()
    return {"matched": len(links), "links": links, "unmatched": unmatched, "rows": len(rows)}