RAG implementations
Retrieval pipelines, chunking, embeddings, reranking, citations, tuned for accuracy.
- Engagement
- Fixed-scope engineering, often on a system you already have
- Typical timeline
- 3–4 weeks
- Starts with
- Fifty real questions and the answers your own experts consider correct.
You already built this. It demonstrates beautifully and then answers a question about the wrong product line, using a document that was superseded last year. Nine times in ten the model is not at fault, the retrieval step handed it the wrong pages and it did its best with them.
Answers grounded in your own data, with sources attached.
What we build
Concrete artefacts, handed over and documented.
- 01
An evaluation set of real questions paired with the passages that should have been found, which is the only way to tell improvement from luck
- 02
Chunking tested rather than assumed: whether tables, headings and clause numbering survive is measurable, and it usually decides the outcome
- 03
Hybrid search combining meaning with exact keyword matching, because part numbers and policy codes fail on meaning alone
- 04
A reranking step that reorders candidates before the model sees them, normally the single largest gain available
- 05
Retrieval quality on a dashboard, so a regression is caught internally rather than by a customer
What changes
Answers stay tied to the correct and current source document
Quality changes are proven against a fixed set instead of debated
Cost per answer falls, because good retrieval needs far fewer tokens than a very large context window
How it runs
01Build the evaluation set
Fifty real questions, the passages that answer them, and agreement from your experts that those are the right passages.
02Retrieval experiments
Chunking, embedding and search strategies compared on the same set, with the losing options recorded so nobody retries them.
03Rerank and hand over
The reranking layer, the dashboard, and a written account of what moved the numbers and by how much.
Chosen per project. Named here so you can see the shape of it.
- pgvector
- Pinecone
- OpenAI
- Cohere Rerank
- Braintrust
Questions we get asked
Tell us the outcome, not the tooling.
Send us the situation you are in. We will tell you which discipline it belongs to, what we would do first and what it costs, including when the answer is to wait.