| name | categorization-debug |
| description | Debug why a PlaylistMiner video was categorized incorrectly or not categorized, by tracing through all 3 pipeline layers. |
Debug Video Categorization
Investigate unexpected categorization results for a video.
Step 1: Get Video Details
Find the video in the database and check its current tags, then check which keyword rules match its title/description.
Step 2: Trace the Pipeline
-
Layer 1 (Keywords): Check tag_rules table for keyword matches against the video's title and description. Sum weights per tag — if total < 0.7 threshold, no suggestion from this layer.
-
Layer 2 (TF-IDF): Check if the TF-IDF corpus has been built (requires manually-tagged videos). If corpus is empty, this layer produces nothing.
-
Layer 3 (Ollama): Only runs if Layers 1+2 produce zero suggestions AND Ollama container is running.
Step 3: Common Issues
- No keyword rules match: Add manual tag → triggers self-learning → creates rules
- Weight too low: Rules exist but total weight < threshold (0.7 default)
- TF-IDF corpus empty: Need more manually-tagged videos to build meaningful vectors
- Ollama offline: Check if pm-ollama container is running
Step 4: Fix
- Tag manually (preferred — trains the engine)
- Or add keyword rule directly in Tags management UI
- Or lower categorization thresholds in Settings
