RAG development: AI answers grounded in your own data
We build retrieval-augmented generation (RAG) systems that answer questions from your documents, wikis and databases – with sources shown, permissions respected and quality measured before launch. Your teams and customers get answers they can check, from content that stays current without retraining a model.
What is RAG?
Retrieval-augmented generation (RAG) lets a language model answer from your own content. When someone asks a question, the system first finds the most relevant passages in your documents and data, then the model writes an answer based on those passages and cites them. Your knowledge stays current without retraining the model.
of organisations either do not have, or are unsure whether they have, the right data management practices for AI – the foundation every RAG system stands on.
Source: Gartner survey of 1,203 data management leaders, February 2025
- Answers depend on your documents, policies or product data
- Content changes often and must stay current
- People need to see the source behind each answer
- Different people may see different information
- The task needs a new skill, format or tone rather than new knowledge – fine-tuning may fit better
- The answer comes from a calculation or a database query – tools and APIs do that better
- There is too little content to search
RAG, fine-tuning or long prompts: which should you use?
They solve different problems, and many production systems combine them.
| Approach | How it works | Best when | Watch out for |
|---|---|---|---|
| RAG | Finds the relevant passages in your content for each question and answers from them, with sources | Knowledge changes often, answers must cite sources, and access differs by user | Answer quality depends on retrieval, so content and permissions need preparing |
| Fine-tuning | Retrains a model on examples so it learns a style, format or narrow task | A focused, high-volume task with stable patterns | It doesn’t add fresh knowledge, and needs retraining when things change |
| Long prompts | Sends whole documents with every question | A few short documents or a one-off analysis | Cost and response time grow with every question, and there is no per-user filtering |
| RAG with a fine-tuned model | A tuned model answers from retrieved content | Specialist language plus current knowledge at scale | The most effort – justify it with evaluation results |
What goes into a RAG system that holds up
The model is the easy part. Quality comes from what surrounds it.
Knowledge pipelines
Connect documents, wikis, tickets and databases; clean, split and refresh them so answers stay current.
Permission-aware retrieval
Every search respects who is asking, so people only get answers from content they are allowed to see.
Hybrid search and re-ranking
Keyword and semantic search combined, then re-ranked, so the right passage reaches the model.
Answers with sources
Every answer links to the passages it used, so people can check it in one click.
Evaluation and guardrails
Real questions scored on every change, plus guardrails against prompt injection and data leaks.
Monitoring and cost control
Quality, freshness and cost per answer tracked after launch.
Part of AI product engineering
RAG is one capability of AI product engineering. It is also part of the services below, and works best with the capabilities listed under them.
Works best with
Other capabilities a RAG system usually needs, each explained in full on its own service page.
How we deliver it
- 01
Pick the questions
The questions people ask, the sources that answer them and how we will judge a good answer.
- 02
Prepare the knowledge
Sources connected, access rules mapped and a test set built from real questions.
- 03
Prove it
Usually a 2–4 week proof of concept on your own data, scored against the test set.
- 04
Launch and run
Released with monitoring of quality, freshness and cost, and improved as your content changes.
Selected client work
“The platform was successfully implemented, thanks to Assystant’s excellent development work. The team worked in a timely and effective manner and delivered items on time.”
Related guides
RAG vs fine-tuning: how to ground LLMs in your company data
Compare cost, accuracy, freshness and security – and learn when to combine both.
Read the articleWhat is agentic AI? Definition, use cases and how to start
AI systems that plan steps, use tools and act with limited supervision – and how to adopt them safely.
Read the articleWhat does it cost to build your own GPT solution?
Approach, data, integration, usage fees, security and maintenance – and how to keep cost under control.
Read the articleHow to scope an AI proof of concept that reaches production
Business outcome, use case, data checks, success criteria and a go/no-go decision.
Read the articleFrequently asked questions
What is RAG?
Retrieval-augmented generation (RAG) lets a language model answer from your own content. The system first finds the most relevant passages in your documents and data, then the model writes an answer based on them and cites its sources. Your knowledge stays current without retraining the model.
RAG or fine-tuning – which do we need?
Use RAG when the model needs your knowledge, especially if it changes often or must be cited. Use fine-tuning when it needs a new skill, format or tone. Many products combine both. Our guide “RAG vs fine-tuning” compares cost, accuracy and freshness.
How do you stop the AI making things up?
We instruct the model to answer only from the retrieved passages, show the sources, say so when nothing relevant is found, and test this on a set of real questions before every release.
Can it respect who is allowed to see what?
Yes. Retrieval filters content by each person’s permissions before the model sees it, so answers only draw on what that person may access, and access is logged.
How long does it take?
Usually 2–4 weeks for a proof of concept on your data, then a first production release in 8–16 weeks, depending on the number of sources and integrations.
Have you built systems like this?
Yes. For BIFI – Berlin Institute for Innovation we built a cognitive translation memory with advanced search over past translations, and for MealsPad a small language model that suggests recipes from what is in the kitchen.