Services · Build
Retrieval-augmented generation (RAG)
AI that answers from your knowledge, not the internet's. RAG finds the right passages in your own approved content and hands them to the AI before it writes, so every answer is grounded, current and traceable. We build it into our own product, slideAcross, and help organisations get it right in theirs.
RAG in four steps
Plainly put, every answer goes through four stages:
- Your knowledge: documents, policies and data, written and approved by people
- Retrieve: find the passages that actually answer the question
- Generate: the AI writes using only those passages
- Check: show the source, and measure how often it's right
Why RAG
General AI doesn't know your organisation
Tools like ChatGPT and Gemini can write about almost anything. But ask them about your products, your policies or your results and they either don't know, or guess with total confidence.
The problem with general AI
- It hasn't read your documents, so it fills the gaps with plausible guesses
- It can't show its working: there's no source to check an answer against
- It goes out of date, and doesn't know when your policies change
- It speaks for you without knowing your positioning, evidence or voice
What RAG changes
- Answers come from your approved content, not the open internet
- Every answer can point to its source, so people can check it
- Update a document and the answers update, with no retraining
- It can say "I don't know" when your content doesn't cover the question
Built here first
RAG at the heart of slideAcross
slideAcross, our presentation platform, is built on RAG. Your content libraries are the knowledge; the slides are the answers.
- Slides from your library: choose a topic and slideAcross builds the slide from your own approved content
- Questions in the moment: when the audience asks something unexpected, ask by voice or typing and it finds the right passage and builds the slide on the spot
- Grounded by design: the AI is there to present your expertise, not to invent it
We're taking it further in our current development: measured evaluation of every answer, smarter chunking of long documents, search that combines keywords with meaning, and a "show source" link behind every generated slide. Everything we learn goes straight into our work for clients.
More about slideAcrossOur view
Most RAG projects fail on content, not code
The AI model gets the attention, but an answer is only as good as the passage it was given, and retrieval is only as good as the content underneath. Messy, duplicated or out-of-date documents produce confident, wrong answers.
That's why we start with the foundation: content written by people, based on clear evidence, and structured so the right passage can be found.
Content library creationWhat makes RAG work
Six things we get right
Each of these is a common point of failure, and each is something we design for from the start.
Foundations
Accurate, approved, human-written content, with an owner and a review date. Everything else depends on it.
Chunking
Splitting documents into passages the right size to retrieve, keeping their headings and context so nothing is taken out of its setting.
Retrieval
Search that combines exact keywords with meaning, then reranks the best matches, with a threshold below which it doesn't answer at all.
Grounded generation
Instructions that keep the AI to the passages provided, with answers that cite where they came from.
Evaluation
A test set of real questions with known answers, so accuracy is measured, not assumed, and every change can be shown to help.
Security and governance
Protection against instructions hidden in documents, access that respects who can see what, and clear rules on data retention.
How we can help
From first idea to measured results
Whether you're exploring RAG, building it, or wondering why your existing AI assistant gets things wrong.
RAG readiness review
Where RAG would add value, whether your content is ready for it, and the risks to plan for.
Content foundations
A human-written, evidence-based knowledge base, structured for retrieval from day one.
Build or pilot
A RAG tool built around your content and your users: an assistant, a search, or answers inside your own systems.
Evaluate and improve
Already have an AI assistant? We'll measure how often it's right, find where it fails and fix it.
Pricing
Priced to your project
RAG projects range from a short review to a full build, so we'll always price against your needs. Tell us what you're trying to achieve and we'll send a priced proposal.
A priced proposal for every project
review, foundations, build or evaluation
Charity or non-profit? We offer substantial discounts. Ask us about it.
Got an idea? Let's make it affirmable.
Tell us what you'd like AI to answer, and what it needs to know. We'll reply within two business days.