Skip to content
AI and Generative AI

RAG and knowledge systems

Most organisations already hold the answers their staff spend the day hunting for. They are in policy documents, closed tickets, contracts, and wikis nobody has opened in two years. Retrieval-augmented generation makes that material answerable in plain language, with a citation attached to every claim.

The problem

Search finds documents. People need answers.

Keyword search returns forty results and leaves the reading to you. A model without retrieval invents a confident answer from nothing. The combination — retrieve the right passages, answer only from them, cite what was used — is what makes the result trustworthy, and getting the retrieval half right is most of the work.

You’ll recognise this if

  • Answering a routine customer question means opening four systems
  • New staff take months to become useful because knowledge is undocumented
  • The same question is asked in your support queue every week
  • Your intranet search is a running joke internally
What you get

What we actually deliver

A retrieval layer over your real corpus

Documents, tickets, records, and pages ingested, chunked sensibly, and indexed — including the unglamorous work of handling scanned PDFs, tables, and the seven near-identical versions of the same policy.

Answers with citations

Every response points at the passages it came from, so a reader can verify in one click. This is the feature that turns a curiosity into something staff will rely on in front of a customer.

Permission-aware retrieval

The index respects who is allowed to see what. Retrieval is a notorious way to leak documents sideways across an organisation, and it is designed against from the start rather than patched later.

Freshness and a way to correct it

A pipeline that keeps the index current as documents change, and a route for a subject expert to mark an answer wrong and have that correction stick.

How we work

The way we approach it

    Fix retrieval before touching the prompt

    When answers are poor, the cause is almost always that the wrong passages were retrieved. We measure retrieval quality on its own, against real questions, before spending any time on wording.

    Build the question set with your experts

    We collect real questions and the answers your best people would give, and that becomes the yardstick. Without it, quality is a matter of opinion and every change is a coin toss.

    Be explicit about what it doesn't know

    When the corpus doesn't contain an answer, the system says so and points to who might. Silence and invention are both worse than an honest gap.

Outcomes

What changes

  • Plain-language answers drawn from your own material, with sources attached
  • Less time lost to hunting across systems for something already written down
  • New staff productive sooner, without a colleague being interrupted for it
  • A measurable retrieval quality score that can be tracked over time
Questions

Asked often enough to answer here

How much content do we need for this to be worthwhile?

Less than people expect. A few hundred substantial documents is enough to be genuinely useful. What matters more than volume is that the content is reasonably current — a large corpus of contradictory, outdated material is harder to work with than a small accurate one.

Our documents are messy scanned PDFs. Is that a problem?

It is work, but it is normal work and we plan for it. Extraction quality from scans sets a ceiling on answer quality, so we assess a sample early and tell you what it will take rather than discovering it late.

Can it handle confidential material?

Yes, with permission-aware retrieval so answers are constructed only from what the asker is entitled to see. For material that cannot leave your infrastructure at all, the pieces can run inside your own environment.

Wherever you’re starting from, let’s figure out the next step.

Tell us what you’re building — we’ll tell you honestly whether we’re the right team for it.