OpenSearch Data Model Diagram

Items carry the story

• One document per crop: geometry, class, provenance, cluster, VLM, probe and score fields
• region_boxes is a nested list: each box has its own state, text, cluster and embedding
• pe_embedding (1024-d) drives search, near-dup and clustering

Projects own their data

• Every project has six indexes named op_prj_<slug>__<role>; settings and projection state share its configs index
• Paths and index names are fixed when the project is created
• A guard refuses any request that names another project's index

Search is separate

• Core /ingest indexes are independent of every project
• All embeddings are k-NN vector fields