Cites the source
Every answer links the document and section, so anyone can check it.
RAG for business: an AI assistant that answers from your documents
Digitup Studio is a technical studio in Madrid that builds RAG assistants: they search your manuals, policies, contracts or catalogue, answer with that information and link the source of every answer, with per-user permissions and without retraining any model.
In short
- RAG (retrieval-augmented generation) makes an AI model search your manuals, policies, contracts or catalogues first and answer with what it finds, citing the source.
- It stays up to date without retraining, and each user only sees what they are allowed to see.
- We have applied it to the Mundomar Cruceros catalogue and to Rescrito's PDFs.
- Projects from €1,000, with delivery from 2 weeks.
What it is
What is RAG in one sentence?
RAG is a technique that makes an AI model search your documents first and then write the answer with what it found, showing where it came from.
The acronym stands for Retrieval-Augmented Generation, a term from a 2020 research paper by Patrick Lewis and other authors. A model from OpenAI, Gemini or Anthropic does not know your returns procedure; RAG gives it that information at the moment of answering, without changing the model. It works in three steps:
- Index: documents are split into fragments and stored in an index that searches by meaning (embeddings).
- Retrieve: for each question, the system picks the most relevant fragments that this person is allowed to see.
- Generate: the model writes the answer from those fragments and links the source document.
Uses
What does a company use RAG for?
To have a chatbot over internal documents or a smart document search that gives the answer already written somewhere, with its source, without asking whoever knows.
The same internal questions, again and again?
Manuals and policies
The exact step of the procedure, with a link to the current version.
AI for businessCustomers lost in the catalogue?
Product catalogue
Natural-language search and recommendations, as in Mundomar Cruceros.
Mundomar Cruceros case studySupport swamped by repeated questions?
Support knowledge base
Answers on your website, WhatsApp or Telegram, with handover to a person.
AI chatbotsClauses spread over hundreds of contracts?
Law firms and contracts
Clauses, deadlines and terms located, with document and page.
AI for law firmsCollective agreements and circulars everywhere?
Accountancies
Collective agreements, circulars and each client's documentation, never mixed up.
AI for accountanciesCover hidden in the policy wording?
Insurance brokers
Cover and exclusions of each policy, with the clause in plain view.
AI for insurance brokersLaw firms, accountancies and insurance brokers are possible uses that we describe on their own pages; our published projects with RAG are in travel and education.
Comparison
RAG, uploading files to a chat, or retraining a model?
It depends on how many documents there are, whether they change and who can see what.
| Aspect | Custom RAG | Uploading files to a chat | Retraining a model (fine-tuning) |
|---|---|---|---|
| What it is | Searches your documents and answers with them. | Attaching files to a ChatGPT, Claude or Gemini chat or project. | Retraining a model with examples to change how it answers. |
| If a document changes | It is re-indexed and used from that moment. | You have to upload it again, unless the tool has connectors. | You have to train again. |
| Who sees what | Permissions per user, role or customer. | Whatever each person uploads or connects to their account. | What is learned applies to anyone who uses the model. |
| Cites the source | Yes: document and section. | Depends on the tool. | No. |
| Cost | Development plus model usage. | Monthly fee per user. | Preparing data and training, plus usage. |
| Fits | Extensive, changing or confidential documentation. | Personal use with few files. | Tone, format or highly repetitive tasks. |
Digitup Studio does not do fine-tuning; we include it as a market alternative. OpenAI's Optimizing LLM Accuracy guide (consulted on 23 October 2026) recommends adding context when the model lacks information and tuning it when format or tone are not consistent; the two can be combined.
Sources
Which documents and systems can be connected?
Almost any source with text: documents, shared folders, your website, databases and your CRM or ERP.
- PDF, Word and spreadsheets
- Google Drive and Google Workspace folders
- Microsoft 365 documents and email
- Your website, FAQs and internal pages
- Databases such as PostgreSQL or MongoDB
- CRM, ERP and Google Sheets
Text that rarely changes, such as manuals, contracts or FAQs, is indexed. Data that changes every minute, such as stock, prices or bookings, is queried live through an API, as our AI chatbots do. If the assistant also has to act, for example create a task, we are talking about AI agents.
The assistant can run on your website, in an internal panel, on Telegram or as a WhatsApp AI chatbot.
Permissions
How do you control who sees what?
Each indexed fragment inherits the permissions of its document, and the search filters by user or role before the model sees anything.
If the model never receives a fragment the user cannot see, it cannot reveal it. On top of that, we combine:
- Sign-in and permissions per user, role, team or customer.
- Separation of information between customers or departments.
- A log of queries and of the sources behind each answer.
- Credential protection and backups.
- Human approval for sensitive answers to customers.
What each provider retains from the data it receives is explained in AI integration in your company. If personal data is involved, GDPR fit is something your adviser validates: we do not give legal advice.
How do you stop the AI from making up answers?
By answering only from your documents, citing the source, saying "I don't know" when there is no answer, and measuring before launch. No model is infallible, but this way the errors are visible.
Says "I don't know"
If it finds nothing relevant, it admits it and offers to pass the question to a person.
Measured with real questions
A set of questions from your team, with the correct answer, is used to compare versions and models.
A person reviews
A panel shows conversations, sources used and unanswered questions.
How is the assistant kept up to date?
When you change a document, the system re-indexes it and the assistant uses it from that moment, without retraining the model.
Syncing can be automatic with the source folder or manual from a panel. Old versions leave the index, and unanswered questions point to what is missing from the documentation.
Cost and timelines
How much does RAG for business cost and how long does it take?
Digitup Studio projects start from €1,000 and delivery from 2 weeks. An assistant with a few documents, simple permissions and a single channel is close to the minimum.
Above that there is no fixed rate. What moves price and timeline the most is:
- Volume and type of documents: a few clean PDFs are not thousands of scans with tables.
- Number of sources: one folder versus Drive, Microsoft 365, your website and the CRM.
- Permissions: everyone sees the same thing, or each user sees only their own.
- Channels: website, internal panel, WhatsApp, Telegram or your own software.
- Required quality: evaluation, human review and control panels.
We start with a first version on your real documents to validate answers and costs, and then continue with maintenance. Model usage, billed by the provider, and hosting of the index are separate. Depending on the project, you can pay for a fixed-scope project, over 6 or 12 months, or with Build & Run (see ways of working).
Where have we applied RAG?
For Mundomar Cruceros, a travel agency specialised in cruises, we built a RAG over the catalogue of ships and routes and a recommendation chat. For Rescrito (formerly Aithor), the AI SaaS for academic writing that we developed, the RAG over PDFs and documents connects answers to the user's own material.
Fit
When does RAG fit and when does it not?
It fits when there is a lot of documentation that people consult often. If we think it is not worth it, we will tell you.
When it fits
- Your team or customers repeat questions whose answer is already written down.
- The documentation is extensive, scattered or changes often.
- Not everyone can see the same information.
- Every answer has to cite its source.
When we are not the right option
- There are few documents and only one person consults them: uploading them to ChatGPT, Claude or Gemini is enough.
- You want to retrain a model with fine-tuning: we do not do that.
- The documentation does not exist or contradicts itself: start by organising it or with an AI audit.
- The budget is under €1,000 or delivery has to be in under 2 weeks.
Related services
Frequently asked questions
What is RAG for business?
It is an AI assistant that, before answering, searches your company's internal documentation (manuals, policies, contracts or catalogues) and writes the answer with what it finds. Unlike a document search tool, which returns a list of files, it gives the answer and links its source. Digitup Studio designs it, connects it to your systems and puts it into production.
Are my documents used to train the AI model?
Not by using RAG: this technique does not train any model. Your documents are stored in a project index and, on each query, only the question and the fragments needed reach the model. Whether the model provider retains that data, and for how long, depends on their product and plan, so we review it with you before using real data.
How is RAG different from fine-tuning?
RAG searches your documents when answering; fine-tuning retrains a model with examples to change how it answers. For knowledge that changes, such as procedures, prices or contracts, RAG fits better: it updates on re-indexing, respects permissions and cites the source. Fine-tuning is more useful for fixing a tone or a format. They can be combined, but Digitup Studio does not do fine-tuning.
Does it work with scanned PDFs?
Yes, but it needs a prior text recognition (OCR) step that turns the scanned image into indexable text. Answer quality depends on scan quality: crooked pages, stamps, handwriting or complex tables cause more errors. That is why, before quoting, we test with a sample of your documents and tell you what works as it is and what needs cleaning.
Can it be used on WhatsApp, on the web or on the intranet?
Yes. The same assistant can run in a widget on your website, in an internal panel or intranet, on WhatsApp Business API, on Telegram or inside your own software. Permissions and sources are the same on every channel: a customer asking on WhatsApp only receives public information or their own, and an employee receives what corresponds to their role.
Which AI models do you use?
We integrate OpenAI, Gemini, Anthropic, DeepSeek, Perplexity or others, depending on quality, cost, speed, privacy and type of task. In a RAG the knowledge is in your documents and not in the model, so changing the model that writes the answers does not require retraining anything. In the first version we can compare several using the same real questions from your team.
How much does a RAG project cost and how long does it take?
Digitup Studio projects start from €1,000 and delivery from 2 weeks. Above that there is no fixed rate: price and timeline depend on the volume and type of documents, the sources, the permissions, the channels and the integrations. Model usage and hosting of the index are paid separately and we estimate them with the first version.
Where is Digitup Studio located?
Digitup Studio is in Madrid, at Calle Alcalá 211. We work mainly remotely with companies from Madrid, from other parts of Spain and from abroad, and we meet in person when the project calls for it. We can work in Spanish, English and Russian. You can start with a 30-minute call with Anton Zverev, our CTO, with no commitment, or write to us at hola@digitup.es.
Does your team consult the same documentation every day?
Tell us which questions you want to solve. In a 30-minute call with Anton, our CTO, we will tell you how we would approach it, or whether it is not worth it.