ReCall

Demo

What is ReCall?

ReCall turns hours of CCTV or camera footage into searchable events. Instead of scrubbing through an 8-hour recording to find “what happened at 3:17 PM,” you ask in natural language - “person in a white shirt near the counter”, “someone entered the restricted area”, “what happened before the machine stopped?” - and jump straight to the moment with timestamp, description, and video evidence.

The key idea is event over time, not just labels per frame. Individual detections (“Person at 14:31:04”) don’t tell you what happened; a tracked event (“a person approaches the machine, opens the panel, warning light appears”) does. ReCall is built around that.

It runs local-first by default - YOLO, ByteTrack, SmolVLM2 and BGE all run on your machine via Python workers, with FFmpeg and PostgreSQL + pgvector under the hood. No footage needs to leave your machine; Gemini is optional if you want a cloud VLM.

How it works

• Ingest & sample - video is segmented and sampled (1 frame / 2s) to keep costs down.
• Detect & track - YOLO detections + ByteTrack/IoU tracking build persistent tracks.
• Describe & embed - tracks and frames are turned into temporal descriptions (SmolVLM2 or Gemini) and embedded with BGE bge-small-en-v1.5.
• Search & verify - descriptions are stored in pgvector; a query is embedded and similarity-searched, then enriched with detections, tracks, events and frames. Every result links back to the original video so you can verify it.

The current system is intentionally simple - a Go (Fiber) API + React (Vite) web app with a Go polling worker that claims videos via Postgres row locking and resumes from video_processing_checkpoints. No Kafka/Redis yet.

For more

This is a short summary. For the full pipeline, architecture, tech stack, and roadmap, visit the GitHub repo: github.com/berzz26/ReCall.