integrate openvocab search

[cmosig]

https://github.com/cmosig/deadtrees-search-modular

What to build

• New queue task type

  • Add an “embedding” task type to the processor queue (settings tables, enums, API/processor models).
  • When a COG upload completes (after cog + metadata), enqueue an embedding task for that dataset/orthophoto.

• Processor pipeline step

  • In processor/src/processor.py, add a handler that runs the patch embedder on the COG in processing_path using the local OpenCLIP ViT-H/14 weights (similar to deadwood model placement).
  • Output: list of patch embeddings + bounding boxes in pixel coords, tied to the dataset/orthophoto ID.

• Supabase schema

  • Add a table/column for patch embeddings (e.g., v2_patch_embeddings with dataset_id, patch_bbox, embedding vector, model_id, created_at).
  • Use pgvector for the embedding column; enforce FK back to the dataset/orthophoto record.
  • Add minimal indexes (ivfflat/lsm) for cosine search on the embedding vector.

• Backend API (FastAPI)

  • Add an endpoint to run text encoding on CPU via the modular TextEmbedder (load weights locally once).
  • Endpoint takes text + dataset/orthophoto filter, returns top patches via Supabase/pgvector similarity.
  • Reuse existing Supabase client + auth/token verification patterns.

• Frontend wiring

  • Hook search UI to the new endpoint (send text).
  • Display patch hits (orthophoto ID + bounding box). No webserver in the embedding package itself.

• Ops/Config

  • Place OpenCLIP weights alongside existing models under /assets/models, and add env/config entries for the path.
  • Ensure processor image/conda env includes deadtrees-search-modular and pgvector client deps.
  • Logging: follow existing UnifiedLogger/SupabaseHandler; update status flags for the embedding task.

Acceptance criteria

• Upload completes → queue contains embedding task → processor runs embedder on the COG → embeddings + bboxes stored in Supabase with FK to dataset/orthophoto.
• Text query endpoint returns pgvector-ranked patches for a given dataset/orthophoto filter.
• CPU-only text encoding; GPU for patch embedding (if available).
• Tests: unit test for enqueue flow, processor handler call, and pgvector query stub/integration.

[linear]

DT-165