What people actually install to build RAG in 2026: a two-registry measurement

No survey exists on "the share of teams using approach X", but something else does exist and can be counted: how many times each package was downloaded. On 27 August 2026 we queried the counters directly. Python, per month:

The first thing that stands out: the two most-downloaded packages in this list have nothing to do with RAG. A provider client and a token counter are needed by anyone calling a model at all, regardless of whether they are building document search.

Add up the dedicated vector databases and compare against an extension to ordinary PostgreSQL:

Under two percent apart. That is, the approach of "put the vectors in the database you already have" is comparable in prevalence to every dedicated vector store combined.

It fits our driver measurement too: PostgreSQL leads both ecosystems by a wide margin. If the database is already running and already maintained, adding an extension costs less than standing up and operating a second storage system.

A caveat on the weaviate row. Its Python client shows 111 million downloads a month while its JavaScript client shows 32 thousand a week. A ratio in the thousands is not explained by popularity alone; more likely the package arrives as somebody else’s dependency or is installed repeatedly in builds. We excluded it from the comparison above and do not treat that figure as a measure of use.

The same measurements in npm, per week:

Two differences between the ecosystems matter.

The orchestrator is different. Python is dominated outright by LangChain. In JavaScript its version draws three million a week against the Vercel toolkit’s twenty-three — eight times more. Anyone writing in both languages should know the "standard" tool is not shared.

The provider balance is different. In JavaScript Anthropic’s client has drawn essentially level with OpenAI’s — 35.9 million against 36.5. In Python it trails twofold: 208 million against 452. The same pair of providers, a different ratio.

Downloads are not projects and not users. The bulk comes from build pipelines. A package pulled on every build outranks one installed once.

A week and a month are not directly comparable. The npm and PyPI figures come from different windows and differently sized ecosystems; comparing across them is invalid, comparing within one list is fine.

Downloading a client does not mean RAG. The openai package is installed both by someone building document search and by someone simply asking a model to rewrite text. The counter cannot separate them.

No survey of the patterns in use exists. No vendor and no researcher publishes what share of teams uses reranking, hybrid search or semantic chunking. Anything of the form "85% of teams apply…" has no source.

What these numbers do show is the relative weight of approaches within one ecosystem. And the most practical point: before standing up a separate vector store, check whether an extension to the database you already run would do. By prevalence that route gives up nothing.

To check what database answers for you and how fast, use the port checker.