RAG development

RAG development on your private company data

Retrieval-augmented generation (RAG) lets a language model answer from your documents instead of from what it learned in training. The system finds the passages that matter, passes them to the model with the question, and the model answers with citations a reader can check.

Book a scoping callSend us a project brief

Operated by Azinove SAS (France); we work with US teams remotely.

A demo takes an afternoon. A RAG system people trust takes more: the right content, kept current, retrieved with each reader’s permissions applied, and answers tested against the questions your team actually asks.

What a production RAG system includes

We build the whole pipeline, from the source systems to the answer on screen, and the monitoring around it.

  • Ingestion from where your knowledge lives: document stores, wikis, help centres, ticket history and databases
  • Permission-aware retrieval, so people only get answers from content they are allowed to open
  • Answers with source citations, and a plain “I could not find this” when the content is missing
  • Evaluation sets built from real questions, run on every change
  • Deployment in your cloud account or ours, with logging and cost limits

Permissions come first

The fastest way to lose trust in a knowledge assistant is to let it quote a document the reader should not see. We carry access rules from the source systems into the retrieval index and check them when each question is asked, not only when the content is loaded.

Retrieval is a search problem

Most weak RAG answers are retrieval failures, not model failures. We tune how documents are split, the metadata kept with them, hybrid keyword and vector search and re-ranking against your test questions, and we measure retrieval separately from the final answer so we know which part to fix.

Where our own products fit

GitDoc keeps documentation in Markdown from GitHub and serves it to AI agents over MCP, one way to give an agent dependable access to a body of knowledge. Botsmith, our managed website chatbot, is in development and answers from a client’s site and documents. Our guide to enterprise RAG on the blog sets out the architecture we follow.

Questions

What clients ask about RAG development.

Another question? Email hello@cleverai.ai

What is RAG, in plain terms?

A search step before the model answers. The system retrieves relevant passages from your content and gives them to the model with the question, so the answer is grounded in your documents and can cite them.

Should we use RAG or fine-tuning?

For answering from company knowledge that changes, RAG is usually the better fit: updates take effect without retraining and answers can cite sources. Fine-tuning changes style or behaviour; it is a poor way to store facts.

Can a RAG system respect our document permissions?

Yes, and it should. We bring access rules from the source systems into the index and filter results for each user at question time.

Which document formats can you work with?

Common formats such as PDF, Word, HTML, Markdown and plain text, plus records from databases and SaaS tools through their APIs. Scanned documents need text recognition first, which we plan for when it applies.

How do you check that the answers are right?

We build a set of real questions with their expected answers and sources, then measure retrieval and answer quality against it. The set runs again whenever the content pipeline, the instructions or the model changes.

Can it run inside our own cloud account?

Yes. The index, the application and model access can all be deployed in your cloud account, so documents stay on infrastructure you control.

Does a RAG assistant ever make things up?

It can, which is why we test for it. The model is instructed to answer only from retrieved sources, citations are shown, and the assistant says so when nothing relevant is found. Evaluation tracks how often an answer is not supported by its sources.

Related

Services, products and reading related to RAG development.

Talk to the team

Tell us about the RAG development project you have in mind.

Operated by Azinove SAS (France); we work with US teams remotely.

Book a scoping callSend a project brief