RAG development
company in the UK.
Magora is a London software company that builds retrieval-augmented generation (RAG) systems: AI assistants and search that answer from your own documents, with a cited source for every answer.
What RAG is and
when you need it.
Retrieval-augmented generation (RAG) is a way of making a large language model answer from your own documents. The system searches your content for the passages that match a question, hands them to the model, and the model writes an answer from those passages and cites them. You need RAG when people must get accurate answers from information that is private, specific to your business or changing often.
Search first, then answer
A normal chatbot answers from what the model learned in training. A RAG system looks up your documents first, so the answer comes from your content and can link back to it.
Private, specific or changing content
Staff lose time searching policies, contracts or tickets; customers ask questions your documentation already answers; or you want AI that works from your data without training a model on it.
Simpler tools are sometimes enough
If the job is drafting or summarising for one person, an off-the-shelf assistant is usually enough. If you have few documents or a fixed set of answers, ordinary search or a FAQ page may do. We will tell you which.
RAG development
services.
These are the six kinds of RAG system we build. Each one is measured against your real questions and connected to the systems you already use.
Knowledge-base assistants
A chat assistant over your policies, handbooks, manuals and wikis. Staff ask in plain English and get a short answer with links to the pages it came from.
Document Q&A
Ask questions of contracts, reports, claims files or technical documents and get answers with the quoted passage beside them, instead of reading hundreds of pages.
Internal and enterprise search
Search that understands meaning as well as keywords, across file stores, tickets and records, with filters and permissions so people only find what they may see.
Support copilots
Suggested replies and answers for your support or sales team, drawn from past tickets and product documentation, with a person approving what is sent. Customer-facing chat agents are built the same way.
Agentic RAG
An AI agent that decides what to look up, searches several sources, checks what came back and then acts through your systems, with human approval for the steps that matter.
RAG inside your product
Answers, summaries and search added to the software you already run, without a rebuild. See AI & ML development for the wider range of AI work.
How a RAG system
answers a question.
Nine parts sit between your documents and an answer. Answer quality depends heavily on the first five, which is why we put as much care into ingestion and retrieval as into the choice of model.
Connecting your content
Connectors read PDFs, Word files, web pages, tickets, emails and database records, extract the text and keep it in sync when documents change.
Splitting for meaning
Documents are cut into passages that make sense on their own, keeping headings, tables and metadata such as source, date and owner.
Turning text into vectors
Each passage is converted into a numerical vector that captures its meaning, using an embedding model chosen for your language and subject matter.
Searchable index
Vectors are stored in a vector database with the original text and metadata, so the right passages can be found in milliseconds.
Finding the right passages
A question is turned into a search that combines meaning and keywords, then re-ranks the results so the best evidence reaches the model.
Answering with citations
The language model writes an answer using only the retrieved passages, and each statement links back to its source.
Respecting permissions
Retrieval is filtered by who is asking, so the assistant never shows a document the person could not open themselves.
Measuring answer quality
A test set of real questions with known answers is run on every change, scoring retrieval and answers so quality is measured rather than assumed.
Limits and monitoring
Refusals when nothing relevant is found, personal data redaction where needed, an audit log of questions and answers, and monitoring of accuracy and cost after launch.
The stack
behind our RAG systems.
Chosen per project for accuracy, latency, cost and where your data is allowed to live. The system is model-agnostic, so swapping the model later does not mean rebuilding the pipeline.
- OpenAI GPT models
- Anthropic Claude
- Google Gemini
- Llama, Mistral, Qwen
- LangChain
- LlamaIndex
- Bespoke orchestration
- Tool / function calling
- pgvector
- Pinecone
- Weaviate
- Qdrant
- Hybrid search & re-ranking
- Email & document stores
- CRM & ERP APIs
- Microsoft Teams
- Custom API integration
- AWS Bedrock
- GCP Vertex AI
- vLLM / Ollama self-hosting
- UK / EU regions
- Evaluations in CI
- Guardrails & approvals
- Audit logging
- GDPR by default
- EU AI Act readiness review
RAG that keeps your
data where it belongs.
The first question most clients ask is where their documents go. These are the options we design with; the right mix for your project is agreed in Discovery.
Your documents are not training data
We build on business API tiers, such as the OpenAI and Anthropic APIs, Azure OpenAI and AWS Bedrock, whose terms exclude training on customer inputs by default. We check the current terms for your use case in Discovery, and switch on zero or limited retention where the provider offers it.
UK or EU hosting options
Application data, vector stores and logs can live in UK or EU cloud regions. Where you need model processing in-region too, we choose a provider and model that offer it.
Private deployments
Models can run through Azure OpenAI or AWS Bedrock inside your own cloud account, or as open-source models self-hosted in your VPC or on your own servers, as we did for the University of Essex.
Access, redaction and audit
Role-based access to sources, personal data redacted before prompts where needed, and configurable logging so you can see what the system was asked and what it answered.
GDPR and regulated sectors
Data-minimisation review in Discovery, DPIA support, and experience in healthcare, financial services and education, where the rules on data are strictest.
EU AI Act readiness
A risk-tier review before deployment, with the transparency and documentation your system needs if you serve EU users.
RAG vs fine-tuning
vs a plain LLM.
Most business projects start with RAG, and some add fine-tuning later for style or format. A free prototype and Discovery show which approach does the job at the lowest running cost.
| Plain LLM (prompting a model API) | RAG over your data | Fine-tuning | |
|---|---|---|---|
| Best for | Drafting, summarising, classifying and general chat | Answers grounded in your own documents and records | A consistent style, format or narrow domain task |
| Knows your documents | No, unless you paste them in each time | Yes, it retrieves them at question time | Only what was in the training examples |
| Cites sources | Not reliably | Yes, each answer can link to its passages | No |
| When your content changes | Nothing to update, but nothing is known either | Re-index the changed documents | Retrain the model |
| Data you need | Little or none | The documents the answers should come from | Hundreds to thousands of good examples |
| Main trade-off | Can answer confidently from general knowledge instead of yours | Only as good as the content and the retrieval | Needs retraining as things change |
From documents
to working answers.
Same senior engineers from the first workshop to handover, weekly demos on your data, and no offshore rotation. The prototype is free; the price of the build is fixed after Discovery.
Free prototype
5 – 10 working daysA coded RAG prototype running on a sample of your real documents, so the first conversation about accuracy is about evidence rather than promises.
Discovery
2 weeks · from £3,000We map the document sources, access rules and the questions people actually ask, review data protection, and agree a fixed estimate that is yours to keep, whoever builds it.
Proof of concept
4 – 6 weeks · from £15,000The system runs on real data in a controlled pilot, with evaluations that show whether retrieval and answers are good enough to scale.
Production build
Typically 6 – 8 weeks after DiscoveryConnectors, vector database, access control, API and interface are built and integrated with your systems, with weekly demos on your data.
Run & improve
OngoingMonitoring of answer quality, cost and drift; re-indexing as documents change; and a plan for when the content, the process or the model changes.
What RAG development
costs in the UK.
A RAG pipeline with a vector database and API starts from £30,000 at Magora. These are indicative 2026 prices in GBP; every engagement starts with a free working prototype, and the final figure is fixed after Discovery, when the documents, access rules and integrations are known, then quoted as a fixed price. Support after launch is £2,000–£4,000 per month. See our UK software development pricing, or estimate your own with the cost calculator.
A coded prototype on your own example documents in 5–10 working days, before you commit to anything.
Two weeks: source audit, question set, data-protection review and a fixed estimate.
4–6 weeks. The system on real data in a controlled pilot, with evaluations.
Retrieval pipeline, vector database and API over your documents, with citations and access control.
A production-grade LLM application with evaluations, monitoring and UK / EU data residency.
Agents that automate internal workflows with tool use and human sign-off.
AI we've
put into production.
Not all of our published AI projects are RAG systems, and we do not present them as such. These three show the parts of a RAG project that matter most: answers limited to approved sources, data that stays private, and AI that works inside the tools staff already use.
Empath AI
- Problem
- Call-centre agents had no way to pick up undeclared signs of caller vulnerability, such as anxiety or cognitive decline, during a live call.
- Built
- A SaaS platform that analyses vocal biomarkers in live calls, turns them into guidance prompts shown to the agent inside Microsoft Teams, and adds call management reporting and anonymous diagnosis reporting.
- Outcome
- Agents get live, empathetic guidance in the tool they already use, and managers see caller and agent-wellbeing trends without exposing individuals.
- Link to RAG
- Shows AI guidance delivered inside the tools staff already use, which is where a support copilot sits.
BodySync
- Problem
- Clinicians spent long hours drafting personalised treatment plans by hand and looking up exercises.
- Built
- A web app that takes structured input (diagnosis, assessment findings, treatment preferences, goals) and drafts a phase-appropriate plan with OpenAI, restricted to pre-defined peer-reviewed clinical sources. Built for HIPAA and GDPR, with one-click send or PDF export.
- Outcome
- The MVP is in use in UK clinics, and further development is planned.
- Link to RAG
- Shows the core RAG principle: the model is limited to approved sources instead of answering from general knowledge.
University of Essex
- Problem
- Sensitive files had to be classified by confidentiality level, and cloud AI services were not allowed.
- Built
- A self-contained AI Orchestrator module running two local models: an LLM (Mistral 7B) that proposes new semantic tags and confidentiality rules for human review, and a fine-tuned entity-recognition model that applies the approved rules and writes a standard YAML file.
- Outcome
- Classification is automated without sending data to third-party cloud APIs, and the model is retrained as experts review new tags.
- Link to RAG
- Shows the private route: local models and no cloud AI, for data that cannot leave your infrastructure.
Also: Rollio, KMA 9x9, Interprefy. All projects →
What clients
say.
Verified reviews from our Clutch profile — 4.9 out of 5 across 71 client reviews.
“We were very happy with the overall experience.”
“Partnering with Magora for the development of the messenger was a decision we will never regret.”
“They're accessible, and it doesn't feel like I'm working with a faceless team.”
RAG development:
the honest questions.
Thinking about a wider AI build? See AI & ML development and AI agent development, the fixed-price software development process, or the checklist: is your business ready for an internal AI agent?
RAG stands for retrieval-augmented generation. Instead of letting a large language model answer from memory, the system first searches your own documents for the passages that match the question, then asks the model to write an answer using only those passages and to cite them. The answers stay tied to your content, and they stay current as your documents change, without retraining the model.
At Magora, a RAG pipeline with a vector database and an API starts from £30,000. An AI proof of concept starts from £15,000 (4–6 weeks), a production LLM application from £40,000, and Discovery from £3,000. The prototype is free, and support after launch is £2,000–£4,000 per month. These are indicative 2026 prices in GBP; the final figure is fixed after Discovery, when the documents, access rules and integrations are known, and the build is then quoted as a fixed price. See the full UK software development pricing or try the cost calculator.
A coded prototype on a sample of your documents takes 5–10 working days and is free. Discovery takes two weeks. An AI proof of concept takes 4–6 weeks, and a first production release typically takes 6–8 weeks after Discovery. The time depends mostly on how many document sources there are, how clean they are and how strict the access rules need to be.
Most business use cases start with RAG. The model answers from your own documents and records, which stay up to date without retraining, and every answer can point to its source. Fine-tuning fits when you need a consistent style, format or narrow task and have good examples to train on. The two can be combined. Discovery shows which one, or which mix, your case needs. The comparison table above sets them side by side.
Yes. That is a common RAG request: an assistant that answers staff or customer questions from your policies, manuals, tickets and files, and only shows each person what they are allowed to see. It can run on a business API tier that does not train on your inputs, inside your own Azure or AWS account, or on open-source models hosted on your own infrastructure when the data must not leave it. If a ready-made tool such as Microsoft Copilot already covers your documents well enough, we will say so.
Not by us, and not by the model providers we build on by default. We use business API tiers, such as the OpenAI and Anthropic APIs, Azure OpenAI and AWS Bedrock, whose terms exclude training on customer inputs by default, and we check the current terms for your project in Discovery. Application data, vector stores and logs can be hosted in UK or EU cloud regions. If model processing must also stay in-region, we choose a provider and model that support it, or deploy privately.
No RAG system is guaranteed to be right every time, so we design to catch errors rather than promise there are none. The model is instructed to answer only from the retrieved passages, every answer carries citations so a person can check it, and the system says it cannot find an answer when retrieval returns nothing relevant. A test set of real questions with known answers runs on every change, so a drop in quality is caught before release. Sensitive answers can wait for human approval.
Magora is one of them: a London software company, in business since 2010, that builds RAG systems, AI agents and LLM applications from a free prototype to production, with the same senior team throughout. Whoever you ask, check four things: can they show a working prototype on your documents, how they measure answer quality, where your data is processed, and whether the code, prompts and evaluation sets become yours. At Magora they do: source code, prompts and evaluation sets transfer to you on payment, and we sign an NDA on day one.
Have documents
people can’t search?
30-minute call with a senior AI engineer. Tell us what your team looks up every day — we’ll say whether RAG fits, what a free prototype on your documents would show, and what the build would cost.