expectedwrong hindsight

Rockset Was the Answer

OpenAI acquired them yesterday, so now the answer is somewhere else.

2 min read 223 words #enterprise #retrieval #RAG #rockset #openai
hindsight — evolved

rockset was acquired by openai shortly after this post. the answer got absorbed into the thing creating the question. the observation about enterprise retrieval being a mess was validated by openai deciding they needed to own the database layer.

Enterprise retrieval is a mess — not in the "we need better tooling" sense, but in the "we built five different pipelines and none of them agree on what the current data is" sense.

The vector database solved a specific problem and created six others. Staleness, hybrid search, reranking, chunking strategies that turn out to matter enormously and in ways nobody documented — it compounds. By the time you're doing RAG at enterprise scale, you've got freshness requirements colliding with query complexity requirements colliding with cost requirements, and the whole thing is held together by a Notion doc someone wrote in Q3 2023 that is now load-bearing.

Rockset looked like the answer. Real-time ingestion, converged indexing, SQL on fresh data without an ETL pipeline standing in the way — the thing enterprises actually need when they want to query their own data without waiting a day for it to become queryable. Smart architecture for a genuinely ugly problem.

OpenAI acquired them yesterday.

So: the retrieval problem is real, the solve path was interesting, and the solve path is now inside a company whose primary interest is probably not your enterprise data freshness requirements. What's left is a collection of companies that built on Rockset, scrambling, and everyone else back to the spreadsheet.

The problem doesn't go away. The water finds a new crack.