AI EngineeringRAG & AI search
RAG & AI search
Retrieval-augmented generation connects a model to your own content, so every answer cites the document it came from, and says so when there is nothing to cite. We become part of your team to build it. The AI Audit finds and costs the first one; building it is a separate decision.
A model on its own answers from memory; a RAG system answers from your documents and shows its working.
For contracts, policies, technical manuals and case files, the citation is the product, because the person acting on the answer has to be able to check it.
What we build
OpenKit builds retrieval systems that answer from your own documents and cite the passage every answer came from.
- Search across your documents
- Contracts, manuals, tickets, case files, drives.
- Cited answers
- Every response linked to the exact source passage.
- Hybrid retrieval
- Semantic and keyword, tuned to how your documents actually read.
- Access-aware answers
- The system respects who is allowed to see what.
- Runs inside your controlled environment
- Where policy or contract requires it: see Private AI.
Every system ships with the evaluation set it was measured on, so accuracy stays a number your team can re-run rather than a claim from us.
The same discipline runs in Marella, the private knowledge platform we build and operate. Answers come from approved company sources with the cited passage open beside them, so a reviewer can check the source before anyone acts on it.
Retrieval is one of four builds under AI engineering, alongside agents, language-model development and voice.
Where a raw model lets you down
Four reasons a general model cannot answer a question about your business, and the reason retrieval exists at all.
- It makes things up
- A language model will produce a confident, plausible answer with no basis in fact, a risk when a wrong answer drives a business decision.
- Its knowledge is frozen
- Trained to a cut-off date, it cannot reliably speak to current regulations or this quarter’s numbers.
- It has no proprietary context
- It knows nothing of your policies or your internal history, so its answers stay generic.
- Retraining is expensive
- Keeping a model current by retraining is slow and costly. A retrieval system updates by changing the documents, not the model.
The same question, grounded
“What is our notice period for a break clause?”
Raw model
“A break clause typically requires six months’ written notice.” It reads fluently, it may well be wrong about your lease, and the model gives you no source to check it against, so the error surfaces only when someone acts on it.
Retrieval system
“The notice period is nine months, per clause 14.2 of the lease dated 3 March 2023.” It comes out of your own document and cites the exact clause. Where the wording is ambiguous or missing, it says so rather than filling the gap.
How the build runs once the audit has costed it
- In the audit
Ranked
Document audit and question workshop, ranked against the rest of the work
- Accuracy baseline measured
- On the roadmap
Costed
A cost, a place in the order of work and the hours it gives back, before anything is built
- Build or buy settled
- Sequenced by payback
- In implementation
Built
Your documents indexed and searchable, inside an agreed delivery allocation or a separately scoped build
- Retrieval tuned to how they read
- In use
Measured
Benchmarked against your baseline
- Cited answers checked by your team
- Next document set scoped
Beyond fetch and answer
Where a single retrieval pass is not enough, we reach for one of three patterns, and each one answers a different kind of question.
- Agentic RAG
- An agent breaks a complex question into parts, works out which source answers each part (a document base, then a live API, then a database) and writes one grounded answer from what comes back.
- GraphRAG
- Retrieval over a knowledge graph as well as text, so the system understands the relationships between records rather than only their wording. Ask it which project used the same supplier as the one a named person led and it can follow the links, where plain text search cannot.
- Multi-modal RAG
- Retrieval that reaches beyond text into images, diagrams and audio, so a model can consult a technical drawing or a recorded call to answer a question.
Where a retrieval build sits
- Scoped in an AI Audit
-
We map your documents and the questions, rank the work, and cost it on the roadmap before anything is built. Building it is a separate decision.
How the audit runs - Built with an Embedded AI Lead
-
More document sets, more teams, tighter tuning, carried as an agreed backlog inside the delivery allocation your engagement reserves. Where you want one piece of work rather than an ongoing engagement, we scope that on its own.
How the embedded engagement runs
The honest range for UK retrieval work
Published UK market ranges for taking a retrieval system from scope to a working deployment. What moves the figure most is the state of your documents rather than the model you pick, because most of the work is preparing them. OpenKit scopes each build to an outcome rather than a day rate, and our own fees are on the pricing page.
- Retrieval or knowledge system
- £15k-£50k
- One clean source and a single channel at the low end; multiple sources and complex retrieval at the high end.
- Multi-source / agentic retrieval
- £40k-£150k
- Several systems and reasoning steps coordinated behind one answer.
- Regulated environment
- +10-20%
- Security documentation, access controls, audit trails, evidence.
Source: OpenKit AI development cost guide, published UK market ranges.
When grounding is not the answer
Four cases where retrieval is the wrong tool, and we say so before anyone spends.
- There is nothing to ground in. With no body of documents, retrieval adds nothing: a plain model or a custom LLM fits better.
- The task is a lookup or a calculation. A database query is cheaper, faster, and more reliable than wrapping a language model around it.
- The source data is contradictory or stale. Retrieval will faithfully return the wrong thing; fixing the data has to come first, and we will tell you before building.
- Nobody will keep the knowledge base current. A retrieval system is only as good as its documents: without an owner, its answers decay as the content ages.
If the answer is a tuned model rather than retrieval, that is large language model development, and we will point you there instead.
“… The transparency is fantastic; you can instantly verify every AI output against source documents, which builds trust with our users.”
Lease review with verifiable citations against the original PDF is the BAiSICS build; the other two run the same pattern over commercial leases and clinical triage. All case studies
What teams ask before the first document audit
What is retrieval-augmented generation (RAG)?
Retrieval-augmented generation connects a language model to your own documents. Instead of answering from what it memorised in training, the model first retrieves the relevant passages from your knowledge base and then answers from them, with a citation. OpenKit builds custom RAG systems so the answer is grounded in your data and you can check where it came from.
How is this different from a chatbot on our website?
A website chatbot follows a script or answers from general training. A retrieval system reads your own documents and cites the passage each answer came from, so your team can check it.
Why use RAG instead of a raw LLM like ChatGPT?
A raw model invents plausible-sounding facts, has a fixed knowledge cut-off, and knows nothing about your internal information. Retrieval changes each of those: answers come from sources you can open, the knowledge is as current as your documents, and it is your data the model reads. For anything where a wrong answer has a cost, grounding is what makes the system usable.
Our documents are a mess.
Most are, which the build assumes. BAiSICS runs on large, complex, poor-quality documents. The document audit is the first thing we do, and it tells us what we are working with.
Can it hallucinate?
A RAG system answers only from your sources and cites them. Where no source covers the question, we build it to say it has not found one.
Does that remove the risk entirely?
No. A well-built system retrieves the right passages, answers only from them, cites each claim and refuses when it cannot find support, so wrong answers become rare and, more importantly, checkable. We design the refusal and citation behaviour rather than hoping the model behaves.
Does our data train someone else’s model?
No. We contract with model providers on terms that keep your content out of their training data. Where that is not enough for your obligations, the whole system runs privately inside your own environment.
How do you keep our data secure in a RAG system?
The knowledge base and the model run where you choose (a UK region or your own servers) with role-based access so a user only retrieves what they are permitted to see. OpenKit is ISO 27001, ISO 9001 and Cyber Essentials certified, and the systems we build are designed to operate under UK GDPR. Client data is used only to deliver the engagement and is not retained afterwards. Every answer carries an audit trail of the sources it used.
How accurate is it?
We measure it. The system is benchmarked against your own baseline, on your own questions, before anything rolls out to the people who will use it.
How do you measure retrieval accuracy?
Against your own material. We build an evaluation set from real questions and expert-approved answers, then measure retrieval quality and answer accuracy against it before launch. For EMQN’s assessment platform we tested six models across 150 scenarios in six languages, and reported the honest gap between per-criterion accuracy and strict whole-report accuracy so the human-in-the-loop decision was evidence-based.
Can a RAG system work with our existing documents and systems?
Yes. OpenKit integrates retrieval with the document stores, wikis, ticketing systems, and databases you already run, through their APIs. The retrieval layer indexes your content where it lives; we handle the messy part (inconsistent formats, scanned PDFs, mixed languages) so the system reads what your team actually has, not an idealised version of it.
Who offers RAG solutions for enterprises in the UK?
OpenKit builds bespoke RAG systems around your own data: ISO 27001 and ISO 9001 certified, deployed in a UK region or on your own servers, with no per-seat licence that scales against you. A UK team does the work. The engagement covers access, handover and who runs it.
What does a RAG system cost in the UK?
Published UK market ranges put a retrieval or knowledge system at roughly £15,000 to £50,000, and multi-source or agentic retrieval spanning several systems at £40,000 to £150,000 (OpenKit AI development cost guide). Regulated environments add ten to twenty percent for the security and audit work. What moves the figure most is the state of your documents rather than the model you pick, which is why the document audit comes first.
Do we own the RAG system, or is it locked to a vendor?
The code we write for you is yours, with the documentation to run it. Third-party components keep their own terms. OpenKit builds on open components and hands over the index, the pipeline and the deployment, running on infrastructure you control. It keeps working whether or not you engage us again, and you are free to change it or hand it to another supplier.
Not ready to talk? The free AI readiness check scores where you stand in about five minutes.
What could your business do with AI?
Find out with experts who become part of your team. We uncover opportunities, get ideas working, and help your people build on the results.
We reply within one working day.