Service
RAG and enterprise search
RAG (retrieval-augmented generation) means the model answers from your documents rather than from memory. Zero-Shot builds RAG systems over corporate archives: every answer shows its source, documents the user is not allowed to see never enter the answer, and the system refuses to answer when no source supports it.
What we build
- 01An ingestion pipeline across file servers, SharePoint, Drive, email, wikis and databases
- 02A retrieval layer filtered by user permissions, so unauthorised documents never reach the answer
- 03Citations with every answer: document name, page and the quoted passage
- 04An answerability check — without sufficient sources the system returns 'not found' instead of prose
- 05An evaluation set of real questions with the correct source labelled by hand
How it runs
01
Collect the questions
We gather 100-300 real questions the system is expected to answer. What we measure is defined before anything is indexed.
02
Survey the archive
How many documents, which formats, which dates, who can access what. This step usually reveals the archive is messier than anyone assumed.
03
Retrieval first, answers later
We build and measure retrieval alone: does the right document appear in the top five. Discussing answer quality before that number is good is pointless.
04
Answer layer and limits
Once retrieval is strong the answer layer is added, with separate tests for uncited answers, permission leaks and mixed-up document dates.
What we measure
- Share of questions where the correct document appears in the top 5
- Number of answers not grounded in a source (target: zero)
- Permission leak test results (target: zero)
- Time to find an answer, before and after
- Share of answers drawn from outdated documents
When this is the wrong fit
If the answer lives in one structured database, RAG is the wrong tool — write a query instead. If the archive is full of contradicting versions, fix versioning first: the system does not resolve contradictions, it treats both as true.
By sector
The same work runs differently in every sector. The detail is here.
Common questions
How is RAG different from the model's own knowledge?
The model's own knowledge comes from training data; it cannot tell you when it was updated and cannot cite anything. With RAG the answer is generated from documents pulled from your archive at query time. Change a document and the answer changes, and every claim has a source you can open.
How do you control who can see what?
Permission checks happen in the retrieval layer, before any answer is generated. A document the user cannot open never enters the result set, so it cannot enter the answer either. A dedicated leak test suite runs against this on every release.
Does our data leave our infrastructure?
That is a deployment choice. The same system can run entirely on your own servers with open-weight models, or against a cloud provider. Where data goes is written down before setup; if regulation or contract keeps data in-country, we deploy locally.
Other services
What we've written on this
- Where to start with enterprise AIHow to choose a first AI project: which tasks qualify, in what order to proceed, and what to expect in the first eighty days.
- Why AI makes things up, and how to reduce itWhy language models state false things confidently. The real cause of hallucination, which countermeasures work, and which do not.
- Is it safe to send company data to AIWhat happens when company data reaches a language model: provider terms, training use, data protection law, cross-border transfer and self-hosting.
Last updated: 2026-09-05
Let's talk about this one
Tell us what you want to build and within two days we'll tell you who takes it and how we'd measure it.