LoRA Recommender Module
Train and deploy a LoRA Recommender in Workbench, then score ranked offers through the same Runtime closed loop as other recommenders (POST /invocations, then POST /response).
Operator entry: Solutions → LoRA Recommender (/lora).
It is a sibling of Two-Tower. Two-Tower (predictor.model.type=similarity) is unchanged.
Where it fits
| Stage | Component |
|---|---|
| Train / export | Workbench Solutions → LoRA Recommender |
| Live score | ecosystem-runtime with predictor.model.type=lora |
| Plugins | PrePredictLoRA and PostScoreLoRA (bind on deployment) |
Training data
Feature rows typically come from logging.ecosystemruntime_flatten, produced by Generate CPR Feature Store. Align date windows with Two-Tower jobs when both modules share a deployment.
Embeddings and run metadata live in:
logging.lora_user_embeddingslogging.lora_item_embeddings
Operator flow
- Configure a LoRA job in the Workbench UI (training corpora and split).
- Train, then Concept test and Export embeddings when you are ready to deploy.
- Bind on deployment — open the deployment step, set Model Type → LoRA Recommender, and bind the training
run_id. Set runtime properties and Push. Push fails if required embedding data for that run is not present. - Optional: Generate Notebook or Use in Python for a standalone training script.
External integrations: public gateway LoRA routes with scopes lora:configs and lora:train.
Closed loop
When collecting feedback: POST /invocations, then POST /response (singular) with that row’s uuid and offer_name. params is a JSON string.
In this module
Related
- Two-Tower Module — dual-encoder retrieval
- CPR Feature Store — flatten source for training
- Runtime plugins — Pre-Predict and Post-Score binding
- Public APIs — LoRA