RAG systems
A model that has not read your documents will improvise, and improvising is expensive. Retrieval is how you make it quote instead of guess.
What it is
We build enterprise retrieval-augmented generation systems: your private corpus indexed, chunked, embedded and searched, with the retrieved passage handed to the model as the only context it is allowed to answer from.
The hard part is not the embedding. It is evaluation: knowing which question your retrieval fails on, and finding out before a customer does.
How we build it
Fix the corpus before the index
Duplicate policies, superseded versions and scanned PDFs sabotage retrieval more than any chunking strategy saves it. We clean first, because garbage retrieved confidently is worse than nothing retrieved.
Build the question set before the pipeline
We write the questions the system must answer, with the passage that should be retrieved for each, before tuning anything. Without it you are optimising a number you invented.
Make refusal the default
When retrieval comes back weak, the system says it does not know. A confident wrong answer from your own knowledge base is the fastest way to lose trust in the whole thing.
Who it is for
A good fit if
- Organisations with a large internal knowledge base nobody can search.
- Support and compliance teams that must answer from an authoritative source.
- Anyone whose data cannot leave their own infrastructure.
Not a fit if
- Corpora too small to justify retrieval. A long context window is cheaper.
- Questions with no written answer anywhere in your documents.
See it
In buildA reference build
Hybrid retrieval over a private corpus
Keyword search and vector search fail differently, so the vault runs both and merges the results before the model sees anything. The point is not elegance: it is that a question phrased nothing like the document still finds the document.
- Keyword and vector retrieval merged, not either-or
- Every answer cites the passage it came from
- Runs against a private corpus with no external call for the data itself
This is a system we run, not a client case study. We do not publish client numbers we cannot show you the workings for.
Talk to someone who builds this
No qualification call, no discovery deck. Describe the work and we will tell you whether this is the right line for it.