LoRA Runtime

Properties

Set on the deployment / ecosystem.properties:

predictor.model.type=lora
predictor.lora.run.id=<run_id>
predictor.lora.embedding.db=logging
predictor.lora.user.collection=lora_user_embeddings
predictor.lora.item.collection=lora_item_embeddings
predictor.lora.customer.key=<customer_field>
predictor.lora.offer.key=<offer_field>
predictor.lora.metric=cosine

Replace <run_id>, <customer_field>, and <offer_field> with values from your project and flatten schema. Bind PrePredictLoRA and PostScoreLoRA on the deployment step (or set plugin.prescore / plugin.postscore to the class names your project export emits).

model.type=lora skips H2O MOJO and Dynamic Engagement, the same way similarity does. Ranking uses cosine or dot product per predictor.lora.metric.

Plugins

Push fails if the bound run_id has no embedding documents in those collections.

Feedback

Closed loop: POST /invocations then POST /response with uuid and offer_name. Endpoint is singular /response.