FalconVQA Docs

Troubleshooting

The failures that actually happen, what causes them, and how to clear them.

Symptoms

Resetting

Everything core writes lives under core/data/:

rm -rf core/data

This deletes the analysis — records and vectors — not just the cache. Application rows survive in Postgres, so afterwards those videos will show as failed and re-index is the recovery.

To clear only the regenerable part, delete core/data/cache instead. That costs a re-download and nothing else.

Environment notes

These look like things to fix and are not:

  • USE_TF=0 is set before anything imports transformers. Without it, transformers imports TensorFlow — several seconds of startup and a wall of banners, for a backend nothing here uses.
  • The torchcodec warning is filtered, not fixed. Audio is decoded with PyAV and handed to pyannote in memory, so torchcodec is never used.
  • PaddleOCR was tried and removed. It broke cv2, forced a torch import order, had no GPU build for CUDA 13, and needed enable_mkldnn=False to run at all. EasyOCR replaced it.
  • Python 3.13 with torch 2.11.0+cu130 is the tested combination. Newer than most guides assume, and it works.

Verifying against a live server

Verify changes against a running server, not by reading code. Several bugs in this project — broken synthesis, stale ports, wrong media paths — only appeared over HTTP.

On this page