Semantic Search

TLDR — The key innovation which has allowed for huge advances in semantic search has primarily been improvements in training methodology. Naive approaches to contrastive learning don’t present a challenging enough problem to force models to learn interesting representations. Erroneously sampling false negative examples greatly hurts model performance. New techniques…

This post surveys recent techniques for semantic product search, especially neural network models. We introduce and classify a number of models, and describe the commonalities and differences in their characteristics such as model architecture, loss function, etc. We believe familiarity with the various choices of the building blocks in semantic…

Databases like arXiv are essential in researching any topic, and come in especially handy for going down avenues of curious study. It’s important in pursuing any learning that one can find applicable and relevant resources, which is half the struggle of study. Living in the Silicon Valley where the pace…

Let’s rank passages in a dataset using the nearest neighbor search — Dense embeddings are a game changer in machine learning, especially in search engines and recommender systems. Dense embeddings are currently being applied in ad-hoc information retrieval, product search, recommendation engines, etc. Many companies are presently adopting some form of embedding-based search in their workflow. Some articles on how dense embeddings…

Search your large collection of documents in the most effective way using open source tools Semantic search is the task of retrieving documents from a collection of documents (also known as a ‘corpus’) in response to a query asked in natural language. Powered by the latest Transformer language models, semantic…

TLDR — Single vector embedding search model’s are efficient for search but they create an information bottleneck which limits performance. In this review we introduce a class of models which represent text using multiple embeddings such as ColBERT. The literature shows that these models are much better at domain transfer…

Databases have been the foundational technology that has shaped our world over the last fifty years. Databases have progressed from flat files to relational to NoSQL, and all of these types of databases are still in use. …

TLDR — Ensembling keyword and embedding models for search is one of the quickest and easiest ways to improve search performance over the standard embedding-based search paradigms. There is a large amount of evidence in the machine learning literature that supports that this helps with in-domain performance, out-of-domain generalization, as…

We’ll implement a naive semantic video search using OpenAI’s CLIP model in Python (ignoring audio). (GitHub repo for this post here.) By the end of the post, we’ll be able to search for content within a video by describing it with words: query = "A man hanging from a boom…