Choose a vector host by scale

financial reports, technical docs, papers with figures or formulas, contracts, medical records, or anything where tables and charts must survive.

long mixed documents where only a paragraph answers the query, and token cost at generation matters more than an extra model call.

the answer connects facts that never appear in one chunk (who works in the same department as the CEO's assistant).

production queries that include SKUs, acronyms (WCAG example), error codes (the intended example is a connection-refused code; captions say "econ refuse"), or…

the same text is embedded twice, or the same question is asked again.

a demo works on about 10 documents and breaks around 10,000, or answers go wrong in production.

local vector store for development.

before embedding.

after a workload exists, not before the first prototype.

before chunking, whenever context comes from files or URLs.

first end-to-end chain after chunks exist.

starting the local environment the course codes against.