A writes:

Recently I’m trying to build a semantic search system with my own data and I came across your blog post. I found quite a few papers using “Recall@K” as an evaluation metric (e.g. Semantic Product Search by Amazon, Embedding-based Retrieval in Facebook Search by Facebook, Embedding-based Product Retrieval in Taobao Search), but it is unclear how they obtain the total number of relevant documents (or items) for their query-document pairs.

While it is totally possible to hire a lot of annotators to figure out which documents are relevant to a search query, I don’t think that is economically feasible at all. Do you have any idea how engineers in industry figure out the total number of relevant documents (or items) for their query-document pairs? Many thanks!

If I had to build a search engine from scratch, I would:

• Start with lexical matching such BM25 or what’s available in Elasticsearch or Solr
• Deploy this in production and collect data on what users click on (i.e., labels)
• Then, use these labels for semantic search

I think using human annotators can work, but probably only for defects or edge cases, given how costly it is.

Have a question for me? Happy to answer concise questions via email on topics I know about. More details in How I Can Help.