Retrieval is where AI succeeds or fails
Almost every broken AI feature we are called in to fix is a retrieval problem, not a model problem. We build the corpus, the index and the evaluation that decides whether an answer was any good.
Three reasons retrieval work comes to us
The corpus first
Before any index exists, we look at what you actually have: formats, duplication, access rules, how often it changes, and what is simply wrong. Indexing bad material faster does not help.
→Corpus audit: formats, duplication, freshness, gaps→Chunking designed against document structure, not a fixed size→Metadata and permissions modelled up front→Source of truth agreed where documents disagree→A refresh pipeline, because the corpus keeps movingIndex and query design
Vector search alone rarely wins. We combine it with keyword and filters, then tune retrieval against questions your experts actually ask.
→Hybrid search: vectors, keywords and metadata filters→Re-ranking where precision matters more than recall→Query rewriting for how people really type→Row- and document-level access enforced at query time→Latency and cost measured per query shapeEvaluation and citations
We build an eval set with your domain experts and hold every change against it — and every answer points back to the passage it came from.
→Eval set written with your experts, owned by you→Retrieval metrics separated from generation metrics→Citations to the exact passage, not the document→Regression testing on every index or model change→Dashboards for unanswered and low-confidence queriesBook a call→Measured, not asserted
Retrieval quality is a number. We report it before and after every change, and we separate it from how well the model writes.
What we build it on
A short, deliberate list — each in production on a system we maintain. See the full partner stack.
Clients we work with
Retrieval case studies
What we will not do
Retrieval is not always the answer either. These are the jobs we hand back.
$./retrieval-review→