Important Issues¶
Table of Contents
Cold Start & Long-Tail Handling¶
Use pretrained content encoders
Use popularity prior & meta-learned embeddings
Rely on seller/group identity for long-tail bootstrapping
Data Labeling and Feedback Loops¶
Weakly supervised methods
Use LLMs to generate labels from noisy data
Online learning or continual training pipelines
System Design Aspects¶
Real-time embedding update pipelines
Feature freshness (e.g., contact count in last hour)
Ranking latency vs. relevance trade-offs
Efficient ANN index sharding