PinnedDaniel TunkelangRanking vs. Relevance: 2 Pitfalls and How to Avoid ThemRelevance takes priority over desirability, but desirability dominates small differences in probability of relevance.Apr 29Apr 29
PinnedDaniel TunkelangLLMs and RAG are Great, But Don’t Throw Away Your Inverted IndexIt is tempting to believe that we can dispense with the inverted index in favor of embedding-based retrieval. But there are a few…Mar 29Mar 29
PinnedDaniel TunkelangSparse and Dense RepresentationsAI-powered search moves from sparse bags of words to dense embedding-based representations. But sparse vs. dense is a false dichotomy.Apr 15Apr 15
PinnedDaniel TunkelangAI-Powered Search: Embedding-Based Retrieval and Retrieval-Augmented Generation (RAG)Replacing traditional search with AI-powered search means embedding-based retrieval and possibly retrieval-augmented generation (RAG).Apr 83Apr 83
PinnedDaniel TunkelangSemantic Equivalence of e-Commerce QueriesSemantic Equivalence of e-Commerce Queries by Aritra Mandal, Daniel Tunkelang, and Zhe Wu. KDD 2023 Workshop on E-Commerce and NLP (ECNLP).Aug 7, 20231Aug 7, 20231
Daniel TunkelangIs Similarity Objective?Many search problems involve content and query similarity. Is similarity objective? In theory, no. In practice, it’s often close enough.Jun 24Jun 24
Daniel TunkelangBags of Documents and the Cluster HypothesisThe bag-of-documents model is a corollary to the cluster hypothesis. The model is likely to fail if a query violates the cluster…Jun 101Jun 101
Daniel TunkelangBags of Queries as Sparse Document RepresentationsThe bag-of-queries model provides a sparse document representation that can be useful as either a positive or negative relevance signal.May 28May 28
Daniel TunkelangIs Targeted Advertising EthicalUsers have — or should have — two choices: pay cash or accept targeted advertising. There is no free lunch.May 71May 71
Daniel TunkelangLLMs and RAG are great. What’s Next?In the next few years, I believe that we will see LLMs focus less on size and more on function calling and tool use.Apr 18Apr 18