Three engineers in the room before you commit.
Compare architectural takes, catch the bug none of them would catch alone, and ship with the trade-offs visible.
Keep it in pgvector. You are at 4M rows; the operational cost of a second store is higher than the query latency you would save right now.
Stay on pgvector but add an HNSW index and a partial index on workspace_id. You will get the latency without splitting the system.
Migrate when you cross 50M rows or when p95 query time crosses 200ms, whichever comes first. Build the abstraction now, defer the move.
Stay on pgvector. Add HNSW + workspace partial index this sprint, abstract the query layer behind a `VectorStore` interface so the eventual move is mechanical. Re-evaluate at 50M rows or 200ms p95.
What changes for you
See the boring-and-correct take, the clever take, and the contrarian take side by side.
Subtle race conditions, wrong assumptions about your stack, or an outdated API surface get caught when three reviewers compare.
The verdict comes with reasoning and trade-offs, not just a code block.
Which mind to ask
Each model leans differently. Maya keeps the differences visible.
Best for long-context code reading, careful refactors, and walking through unfamiliar codebases.
Best for general problem-solving, broad framework knowledge, and writing the code from spec.
Best for fast iteration, polyglot snippets, and reading API docs or screenshots of stack traces.
Where teams reach for it
Hear from three minds, get one verdict.
New accounts get $1 of trial credit. No card required to begin.