All work

Semantic docs search

  • RAG
  • Vector search
  • Evals
  • 12k
    Documents indexed
  • <800ms
    Median query time
  • 0
    Uncited answers

The problem

Staff couldn't find anything in a decade of accumulated internal documentation. Keyword search returned either nothing or four hundred results, so people asked colleagues instead — which is expensive.

What I built

Chunked and embedded the corpus, added hybrid keyword-plus-vector retrieval so exact terms like part numbers still work, and constrained the model to answer only from retrieved passages with inline citations. Built a small eval harness of real staff questions to catch retrieval regressions before deploy.

The outcome

Answers arrive in under a second with links to the source document. Because every answer is cited, people trust it enough to actually use it — and when it's wrong, they can see why.

Need something like this?

If this looks close to a problem you have, the fastest way to find out is a short call.