Skip to Content
DocsModulesLoRA Recommender

LoRA Recommender

The LoRA Recommender is a retrieval module, not a generative LLM feature. It trains a low-rank residual on top of frozen nomic-embed-text embeddings, then scores offers in the Runtime closed loop like other recommenders.

Operator entry: Solutions → LoRA Recommender (/lora).

Training data

Feature rows typically come from logging.ecosystemruntime_flatten, produced by Generate CPR Feature Store. Align date windows with your Two-Tower jobs when both modules share the same deployment.

Embeddings and run metadata live in:

  • logging.lora_user_embeddings
  • logging.lora_item_embeddings

Operator flow

  1. Configure a LoRA job (rank, corpora, train/validation split).
  2. Train — fits the residual on frozen encoder outputs (rank=0 disables the residual for baseline comparisons).
  3. Concept test — offline checks before export.
  4. Export — artifacts for deployment binding.
  5. Bind on deployment — set runtime properties (below) and Push to the target runtime.

Public HTTP routes also exist under Workbench public_api/routers/lora.py for scoped integrations.

Runtime scoring

Set on the deployment / ecosystem.properties:

predictor.model.type=lora
predictor.lora.rank=8
predictor.lora.user_embedding_collection=logging.lora_user_embeddings
predictor.lora.item_embedding_collection=logging.lora_item_embeddings
plugin.prescore=…PrePredictLoRA
plugin.postscore=…PostScoreLoRA

Exact property keys follow your project template; see Workbench deployment export for the bound class names.

Closed loop unchanged: POST /invocations then POST /response with params as a JSON string.

Project Generate

Project Generate can emit LoRA-related keywords and deployment snippets when the prompt describes embedding-based retrieval — use it after the flatten store and training corpora exist.

Last updated on
Access & OperationsOverview