Topic
4 videos.
BM25 is a formula that scores how well a document matches a query. It counts how often the query words appear, allows for document length, and gives more weight to words that are rare across the collection. Many search tools have used it for years as their basic method of keyword matching.
This is a technical topic, and every talk here is advanced. A calm place to begin is A no nonsense intro to BM25, which explains the settings called k1 and b and shows code. BM25: The Most Important Text Metric in Data Science is from 2023 and starts with the weaknesses of an older method, TF-IDF. Mike King, of iPullRank, then argues in a Kalicube Summit talk that search has moved on to meaning while many SEO tools still lean on BM25. That shift is where vector search comes in.
Sells in this field
Mike King argues that search now works on meaning while many SEO tools still lean on BM25 and TF-IDF, and goes on to relevance engineering and fan-out queries.
Teaching or conference talk
BM25 is a ranking formula for scoring how well documents match a query, and this video breaks down term frequency, document length, the k1 and b settings and a code demo.
Teaching or conference talk
TF-IDF has weaknesses, the presenter argues, and BM25 is built to answer them. The video lays out the problems first, then the BM25 formula.
Older: check current guidanceSells in this field
James Briggs and Sebastian Bruch from Pinecone host a workshop on sparse-dense embeddings, with SPLADE and BM25 examples and a discussion of how the two compare.
Older: check current guidance
