Fix video count (use disk scan not stale manifest); add embedding progress bars and per-section API timing

Author: nodeeeeee

electron/main.js

@@ -188,21 +188,13 @@ function getFilesStatus(courseId) {
   const outDir   = getOutputDir();
   const courseDir = path.join(outDir, String(courseId));
 
-  // Videos from manifest (prefer OUTPUT_DIR, then DATA_DIR)
+  // Videos: count mp4 files actually on disk (manifest can have stale/duplicate entries)
+  const videosDir = path.join(courseDir, 'videos');
   let videoDone = 0, videoTotal = 0;
-  for (const mp of [
-    path.join(outDir,    'manifest.json'),
-    path.join(DATA_DIR,  'manifest.json'),
-  ]) {
-    if (fs.existsSync(mp)) {
-      try {
-        const m = JSON.parse(fs.readFileSync(mp, 'utf8'));
-        const items = Object.values(m).filter(v => v.path && v.path.includes(String(courseId)));
-        videoTotal = items.length;
-        videoDone  = items.filter(v => v.status === 'done').length;
-        if (videoTotal > 0) break;
-      } catch {}
-    }
+  if (fs.existsSync(videosDir)) {
+    const mp4s = fs.readdirSync(videosDir).filter(f => f.toLowerCase().endsWith('.mp4'));
+    videoTotal = mp4s.length;
+    videoDone  = mp4s.length;  // present on disk = downloaded
   }
 
   const countJson = (dir, excludePattern) => {

note_generation.py

@@ -808,9 +808,10 @@ def generate_section(
         return sec_file.read_text(encoding="utf-8")
 
     chunk_title = _chunk_title(chunk)
+    fi_tag = f"F{ld.file_idx} " if ld.file_idx > 1 else ""
     if bar:
-        fi = f"F{ld.file_idx} " if ld.file_idx > 1 else ""
-        bar.set_postfix_str(f"L{lec_num}{fi}§{ci}/{chunk_title[:22]} generating")
+        bar.set_postfix_str(f"L{lec_num}{fi_tag}§{ci}/{chunk_title[:22]} generating")
+    tqdm.write(f"  → L{lec_num}{fi_tag}§{ci} [{chunk_title[:40]}]  slides {chunk[0].index+1}–{chunk[-1].index+1}  calling {NOTE_MODEL}…")
 
     # Render only this chunk's slide images (lazy, avoids OOM)
     img_map = ld.render_chunk_images([s.index for s in chunk])
@@ -830,7 +831,10 @@ def generate_section(
         has_transcript=has_transcript,
     )
 
+    import time as _time
+    _t0 = _time.monotonic()
     draft = _call(NOTE_MODEL, _P("system"), user, _max_tokens(detail))
+    tqdm.write(f"     ✓ {len(draft):,} chars  ({_time.monotonic()-_t0:.0f}s)")
 
     if not draft:
         tqdm.write(f"  [warn] Empty draft for L{lec_num} §{ci} — skipping")

semantic_alignment.py

@@ -398,14 +398,15 @@ def get_embedder():
     return model
 
 
-def embed_texts(model, texts: list[str]) -> np.ndarray:
+def embed_texts(model, texts: list[str], desc: str = "  embedding") -> np.ndarray:
     """Return L2-normalised float32 embeddings, shape (N, D)."""
     vecs = model.encode(
         texts,
         batch_size=BATCH_SIZE,
-        show_progress_bar=False,
+        show_progress_bar=True,
         normalize_embeddings=True,   # cosine sim → inner product on unit sphere
         convert_to_numpy=True,
+        tqdm_kwargs={"desc": desc, "unit": "text", "leave": False},
     )
     return vecs.astype(np.float32)
 
@@ -663,15 +664,15 @@ def align(caption_path: Path, slide_path: Path,
 
     slide_texts = _enrich_sparse_slides([s.text if s.text.strip() else s.label for s in slides])
     print(f"  Embedding {len(slides)} slides...")
-    slide_embs = embed_texts(embedder, slide_texts)
+    slide_embs = embed_texts(embedder, slide_texts, desc="  slides")
     index      = build_faiss_index(slide_embs)
 
     # ── Embed transcript segments (with context window) ───────────────────────
     print(f"  Building context windows (±{CONTEXT_SEC:.0f}s)...")
     window_texts = build_window_texts(segments, CONTEXT_SEC)
 
     print(f"  Embedding {len(window_texts)} transcript windows...")
-    seg_embs = embed_texts(embedder, window_texts)   # (T, D)
+    seg_embs = embed_texts(embedder, window_texts, desc="  transcript")   # (T, D)
 
     # ── Raw similarity scores: each segment vs all slides ─────────────────────
     # faiss.search returns (scores, indices); with IndexFlatIP + normalised
@@ -849,7 +850,7 @@ def align_multi_slides(
     print(f"  Building context windows (±{CONTEXT_SEC:.0f}s)...")
     window_texts = build_window_texts(segments, CONTEXT_SEC)
     print(f"  Embedding {len(window_texts)} transcript windows...")
-    seg_embs = embed_texts(embedder, window_texts)
+    seg_embs = embed_texts(embedder, window_texts, desc="  transcript")
 
     # ── FAISS search → log-likelihood matrix ──────────────────────────────────
     print("  Querying FAISS index...")