Perigee is an AI answer engine that makes large documentation sets searchable in plain language: readers ask in the sidebar and get a cited answer in about a second.
Having the vector index sit in the same database as the rest of our data — that's what made this a weekend project instead of a quarter.
Tobias Renner · Co-founder, Perigee
The challenge
Perigee handles customer content on behalf of its own customers, so GDPR compliance was a gating requirement rather than a later milestone. Data residency, deletion guarantees and auditable access controls all had to be settled before the first enterprise pilot.
The index is not small either: 612,400 documentation sections today, growing by 14,000 to 38,000 new sections on an average weekday. Any storage decision had to hold up under that curve.
We picked Basalt over a dedicated vector store because it let us be compliant on day one, with almost no work of our own.
Why they chose Basalt
Row-level policies, scoped keys and a readable audit trail cover most of what a data-protection review asks for. Perigee wired consent state straight into the same tables the retriever reads from, so an erasure request removes the embeddings with the record.
What they built
A set of API endpoints for training content and asking questions against it, a dashboard for syncing repositories and rotating keys, and a small component library so a prompt box can drop into any docs page in a handful of lines.
Auth, database and vectors under one roof. The integration work we didn't have to do is the real number here.
Marta Schöneggar · Co-founder, Perigee
The results
Perigee cleared its first two enterprise security reviews without a custom architecture, and median answer latency held at 840 ms while the index roughly tripled.
Tech stack
- Northsailframework
- Aperturehosting
- Basaltdatabase & auth
- Runcacherate limiting
- Tallowpayments
- Beacontrailanalytics