When does long context beat retrieval, and when does it lose?

Start with what actually changed, because it is smaller than the slogan.

Do I need a vector database if I only have a few thousand documents?

Almost certainly not, and the arithmetic gets there by a different route than the usual one.

Is Postgres with vector support enough?

For most of the projects I see, yes. The useful half of that answer is the point at which it stops being true.

Is BM25 really competitive, and what does hybrid retrieval mean?

BM25 is competitive, it never stopped being competitive, and the evidence behind that is older and stronger than the evidence behind anything else in this article.

What does this cost per month at three realistic scales?

Here is the arithmetic in full. Every unit price below comes from a vendor page read while writing this, and every total is my own multiplication, not a vendor's estimate. Assume 1,536-dimension float32 vectors at 6 KiB each, chunks of 500 tokens, six thousand input tokens and four hundred output tokens per model…

Do You Still Need a Vector Database in 2026? | Kevin Escoda

Long context did not kill retrieval and BM25 never died. The arithmetic that decides it, at three realistic scales, with the evidence graded.

Two claims are competing for your budget this year, and each is delivered as though it settled the question.

The first: long context killed retrieval. Windows reached a million tokens, so put the corpus in the prompt and delete the index. The second: the vector database was always a fad. Agents can grep, and a specialised store is a 2023 artefact kept alive by venture funding.

Both describe something real. Neither is an answer. The answer depends on three numbers: how large the corpus is in tokens, how many queries hit it per day, and how hard you need to filter. Change any one and the recommendation flips, and it flips at values you can calculate in about ten minutes.