AI for Search and Recommendation: Semantic Search, Collaborative Filtering, and Personalization Engines

Status: public · Confidence: medium (0.8) · Basis: verified_sources

## TL;DR

AI search and recommendation systems usually combine retrieval, ranking, and personalization. Dense retrieval and late-interaction search help find candidates, while recommender models learn from user-item interactions and feedback.

## Core Explanation

Search starts with a query and a corpus. Dense retrieval maps queries and passages into vectors, while late-interaction methods such as ColBERT keep token-level matching signals. Recommendation starts with users, items, and behavior data; neural collaborative filtering is one way to model user-item interactions beyond simple matrix factorization.

LLMs add new interfaces and ranking signals, but they do not replace evaluation. Search and recommendation systems still need offline relevance tests, online experiments, safety controls, freshness checks, and transparency about personalization.

## Further Reading

- [Dense Passage Retrieval](https://arxiv.org/abs/2004.04906)
- [ColBERT](https://arxiv.org/abs/2004.12832)
- [Neural Collaborative Filtering](https://arxiv.org/abs/1708.05031)
- [LLMs for Recommendation Survey](https://arxiv.org/abs/2305.19860)

## Related Articles

- [Recommender Systems](./recommender-systems.md)
- [AI Search Engines](./ai-search-engines.md)
- [Federated Learning](./federated-learning.md)