[FEATURE] LLM Feature Store Integration — Embed features for ML pipelines

[gelluisaac]

Description

Integrate LLM-generated features (embeddings, scores, explanations) into the
AstroML feature store for use in ML pipelines.

Scope

Connect LLM outputs to the feature store for downstream consumption.

Files to Touch/Create

• astroml/features/llm_features.py — LLM feature definitions
• astroml/features/embedding_features.py — Embedding features
• astroml/features/scoring_features.py — LLM score features
• astroml/features/llm_generators.py — Feature generation
• astroml/llm/features/integration.py — Integration layer
• astroml/llm/features/compute.py — Feature computation
• astroml/features/pipeline.py — Update to include LLM features

Feature Types

1. Embedding Features:

   • Transaction description embeddings
   • Account behavior embeddings
   • Alert description embeddings

2. Score Features:

   • Fraud probability from LLM
   • Explanation confidence score
   • Uncertainty estimates

3. Meta Features:

   • Prompt version used
   • Model used
   • Latency attribution

Implementation Details

• Feature computation as part of pipeline
• Materialized views for fast access
• TTL and refresh policies
• Feature versioning
• Backfill support for historical data

Acceptance Criteria

• LLM features available in feature store
• Features updated in real-time pipeline
• Backfill completes in <24h
• Feature quality matches standalone LLM calls
• Downstream models can use LLM features
• Feature versioning supports rollback

Integration Points

• Fraud detection model can use LLM scores
• Recommendation system uses embeddings
• Alert prioritization uses LLM confidence

Labels

enhancement, llm, feature-store, integration