Skip to content

RAG development

RAG development services

We build AI that answers questions from your own documents, and shows where each answer came from.

We reply within 24 hours.

The AI looks it up before it answers.

RAG stands for retrieval-augmented generation. The system searches your own documents first, then writes the answer from what it found. Answers stay tied to your current documents, so a policy change does not mean retraining a model. Every answer can list its sources for a person to check.

The idea was set out in a 2020 research paper by Patrick Lewis and colleagues. NIST's glossary describes a RAG system as a model paired with a separate search over a knowledge base. The search finds the relevant passages, and the model uses them to write its answer.

For a business, this solves two problems at once. A language model on its own does not know your policies, prices or contracts. And when it does not know, it can still sound sure. RAG gives the model your facts to work from, and lets a person check each answer against its source.

How we build it.

To build a RAG system, gather the documents, split them into passages and index them for search. Then connect a language model that answers only from the passages it finds, and test it on real questions before launch. We also set who can see which documents, so the system never shows staff what they should not read.

Most of the quality in a RAG system comes from the search, not the language model. If the search returns the wrong passage, the answer will be wrong however good the model is. So we spend real time on how documents are split, how tables and scanned pages are read, and how the search ranks what it finds.

Before launch, we collect a set of real questions from the people who will use the system, with the answers they expect. We measure how often the system finds the right passage and how often the answer matches. The system does not go live until those numbers meet the bar we agree with you.

AI over your company knowledge base

Wikis, policies, contracts and manuals become one place to ask. Each person sees answers only from the documents they are allowed to read.

Access rules come from the systems you already use. If a contract folder is open to the legal team only, answers drawn from it reach the legal team only. When someone's access changes, their answers change with it.

Your knowledge base also changes every week. New policies arrive, old manuals are retired and prices move. We set the system to pick up changes on a schedule you choose, so answers always come from the current version. Old versions can stay searchable for audit, clearly marked.

For a longer guide to this kind of system, read enterprise search with RAG.

  • HR and policy questions from staff
  • Contract and clause look-ups for legal and sales teams
  • Product and support answers from manuals and help pages
  • Technical answers from engineering documents and past tickets

When RAG is the right choice

RAG and fine-tuning solve different problems. RAG brings facts to the model at the moment of the question. Fine-tuning changes the model itself, so it learns a style, a format or a narrow task.

AWS's guidance on comparing RAG and fine-tuning makes the trade-off clear. Fine-tuning can take hours to days, so it suits documents that rarely change. A fine-tuned model also does not cite its sources. RAG can take in new documents within minutes and point to where each answer came from.

So we reach for RAG when answers must come from documents that change, and when people need to check the source. Many of the systems we build use both. For the full comparison, read when RAG beats fine tuning.

We ship our own AI

We build AI products designed to run on your own hardware.

Radha, the copilot in Visylix, runs on an on-premise language model with 111 tools. The same team builds RAG systems for you.

See Visylix
Example

From idea to production.

How we work
  1. Discover

    We learn your goal, your data and your systems. You get a written plan before work starts.

  2. Build

    A senior team builds in short cycles. You see working progress every two weeks.

  3. Launch

    We test it hard, then launch in stages. Nothing goes live until it meets our standard.

  4. Run

    We watch it, fix it and improve it with you, for as long as you need.

Questions about RAG.

How do you create a RAG system?

Gather the documents, split them into passages and index them for search. Then connect a language model that answers only from the passages it finds, and test it on real questions.

Is ChatGPT a RAG model?

Not on its own. ChatGPT answers from what it learned in training. It becomes part of a RAG system when it is given your documents to search first.

What is RAG in data management?

It is a way to put your existing data to work. Your documents stay where they are, and the AI reads from them when someone asks a question.

Should we choose RAG or fine-tuning?

Choose RAG when answers must come from documents that change. Choose fine-tuning when the model must learn a style or a narrow task. Many systems use both. Read RAG or fine-tuning for the detail.

RAG or agentic RAG?

Agentic RAG lets an AI agent decide what to look up, search several sources in turn and check its own answer before replying. It helps with questions that need more than one document. Plain RAG is simpler and cheaper, and it is the right start for most document questions.

How much does RAG development cost?

It depends on scope. We give you a clear estimate after a short discovery call. The main drivers are the number of documents and how often they change. How many systems hold them, and whether it must run on your own servers, matter too.

Sources

  1. Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020)
  2. NIST CSRC glossary: retrieval augmented generation
  3. AWS Prescriptive Guidance: comparing RAG and fine-tuning