_peregrine_

as a tpuf employee i can say we feel very well represented in this article. author did a really nice job. good read too

confirmed - we still write by hand

instant bookmark!

this is actually not how cursor uses turbopuffer, as they index per codebase and thus need many mid-sizes indexes as opposed to one massive index as this post describes

unfortunately i'm not able to share the customer or use case :( but the metrics that you see in the first charts in the post are from a production cluster

the solution described in the blog post is currently in production at 100B vectors

seems like a good rule of thumb to me! though i would perhaps lump "cost" into the "until it breaks" equation. even with decent perf, pg_vector's economics can be much worse, especially in multi-tenant scenarios where you need many small indexes (this is true of any vector db that builds indexes primarily on RAM/SSD)

yeah I mean that's basically what Javi talks about in the post... if you can throw hardware at it you can scale it (ingestion scales linearly with shards)

but the post has some interesting thoughts on how you do the high-scale ingestion while also handling background merge processes, reads, etc.

definitely interesting and related

nice one

Pretty solid at SQL generation, too. Just tested in our generation benchmark: https://llm-benchmark.tinybird.live/

Not quite as good as Claude but by the best Qwen model so far and 2x as fast as qwen3-235b-a22b-07-25

Specific results for qwen3-coder here: https://llm-benchmark.tinybird.live/models/qwen3-coder

Yeah we're always looking for new models to add

Yeah I mean SQL is pretty nuanced - one of the things we want to improve in the benchmark is how we measure "success", in the sense that multiple correct SQL results can look structurally dissimilar while semantically answering the prompt.

There's some interesting takeaways we learned here after the first round: https://www.tinybird.co/blog-posts/we-graded-19-llms-on-sql-...

We need to add it

Actually no, we have it up to 3 attempts. In fact, Opus 4 failed on 36/50 tests on the first attempt, but it was REALLY good at nailing the second attempt after receiving error feedback.