On this page
AI Platform for Internal Tooling
aiinfrastructurevector-search
A self-hosted retrieval layer that makes private documentation searchable.
A platform that lets teams query their own internal documentation with natural language, without sending data to a third party.
What it does
It ingests Markdown and PDF sources, embeds them as vectors, and answers questions grounded in those sources. Retrieval is combined with plain SQL so results are auditable and predictable.
Design decisions
- Self-hosted everything — embeddings and inference run on the company's own infrastructure.
- Grounded answers only, with citations back to the source pages; the model is not allowed to invent facts.
- Boring storage — a Postgres vector index rather than a purpose-built database, because the operations team already runs Postgres.
What I learned
Grounded retrieval is mostly an information-architecture problem. The quality of the answer depends far more on how the documents are chunked and labelled than on which model you pick.