seedeer
A Node.js vision-model toolkit — the vision counterpart to
@jsilvanus/embedeer (text
embeddings) and @jsilvanus/chattydeer
(LLM chat). Import it directly with no HTTP server in the loop; local or
remote model execution is a configuration detail, not an API difference.
Status: all roadmap phases complete. Image embeddings (
JointEmbedder,VisualEmbedder), captioning (Captioner), VQA (VqaAssistant, local and remote backends), and detect/track/zone-trigger (Detector,Tracker,ZoneTrigger,TrackingSession) are all real and working, across all four execution modes (process/thread/socket/grpc) where applicable, with per-pillar benchmark scripts (npm run bench) and multi-server load balancing for the network modes. Seedocs/ROADMAP.mdfor what's coming and in what order, anddocs/ARCHITECTURE.mdfor the design principles behind it.
Planned functionality
| Pillar | What it does | Doc |
|---|---|---|
| Image embeddings | Joint image-text (CLIP-class) and vision-only (DINOv3-class) embeddings | docs/features/embeddings.md |
| Captioning | Cheap, fast, generic image descriptions | docs/features/captioning.md |
| Visual Question Answering | Question-driven answers about an image, local or delegated to a remote OpenAI-compatible vision endpoint | docs/features/vqa.md |
| Detect + Track + Zone-trigger | Real-time person detection, cross-frame tracking, and named-zone enter/exit events for production-assistant use cases | docs/features/detection-tracking.md |
OCR and image redaction/pixel mutation are explicit non-goals — see
docs/ARCHITECTURE.md.
Design principles (short version)
- No HTTP in the consuming product —
importand call a method. - Every pillar supports local-or-remote execution (
process/thread/socket/grpc), including the real-time detection/tracking path. - Where model strength matters more than latency (VQA), the backend is pluggable: a small local model by default, or delegation to a configured remote endpoint — the call shape never changes.
- Reuses embedeer's proven worker-pool/provider-loader/server patterns rather than reinventing them for vision inputs.
- No tight coupling to any specific downstream consumer.
Full rationale in docs/ARCHITECTURE.md.
Installation
npm install @jsilvanus/seedeerLicense
MIT