ParadeDB built a benchmark runner to make experimentation easy. This post looks at a real example over three iterations (with wildly different results).
We (ParadeDB) recently started building out our benchmarking infrastructure for cross-backend comparisons.
Rather than taking the usual path of bundling a workload and execution into one neat package, we decided to build a reusable database benchmark runner based on grafana/k6 first.
"If I had looked at the lexical search and BM25 space in 2016, I would have said it was solved, and that catching up would be nearly impossible."
This interview with Tantivy creator Paul Masurel looks at how wrong I would have been; discussing challenging solved domains, open-source competition done right, and why long-fermented frustration is an underrated driver.
This is a no-nonsense walkthrough of doing hybrid search inside Postgres without spinning up a separate search service.
A few takeaway:
- Postgres’s native `tsvector/ts_rank` stuff works ok for basic text matching, but it doesn’t account for global term frequency like BM25 does , so rankings can feel “flat” or noisy as soon as you go beyond simple queries (it's also slow).
- Using a BM25 index (via extensions like `pg_search`) actually gives you relevance scores similar to what you’d expect out of modern search engines, and you can use stemmers/tokenization directly in SQL. BM25 is the star of this story.
- Vector search fills in the semantic gaps (so “database optimization” isn’t limited to exact keywords), but you still don’t want to throw out lexical relevance. The trick is making it additive, not just adding scores together.
- RRF (Reciprocal Rank Fusion) is a neat practical tool here. It sidesteps trying to normalize totally different scoring systems by just focusing on rank positions.
If you’re building anything where relevance matters (docs, product search, help articles) having BM25 + vector makes a big difference over vanilla FTS + embeddings alone. It also keeps everything in Postgres, which simplifies consistency/ops compared to an external search cluster.
Hey HN! Author here. We added faceted search capabilities to our `pg_search` extension for Postgres, which is built on Tantivy (Rust's answer to Lucene). This brings Elasticsearch-style faceting directly into Postgres with a 14x performance improvement over a CTE based approach by performing facet aggregations in a single BM25 index pass and making use of our columnar store.
You get the same faceting features you'd expect from a dedicated search engine while maintaining full ACID compliance. Happy to answer technical questions about the implementation!
Chinese, Japanese, Korean etc.. don’t work like this either.
However, even though the approach is “old fashioned” it’s still widely used for English. I’m not sure there is a universal approach that semantic search could use that would be both fast and accurate?
At the end of the day people choose a tokenizer that matches their language.
I will update the article to make all this clearer though!
Hello HN, author here. It seems like everyone is talking about 'hybrid search' (lexical/BM25 + semantic/vector) these days, so I wanted to show how it's possible (and fully customizable) using reciprocal rank fusion in SQL.