Architecture · in regular use
Document RAG agency in Nice
The method that makes a model answer from YOUR documents, citing its sources, instead of inventing.
What we do with it
Document RAG at our clients, concretely
Used on a sentiment-tracking tool for AI models, with LangChain and a document store. It is the architecture we choose for natural-language search across your files.
The limits
When we recommend something else
If your documents are neither current nor structured, RAG amplifies the mess instead of solving it.
Our work
Projects we delivered with Document RAG
Paradigm
Platform for a Nice business centre: registered address, digitised mail, messaging and assistant.
Document RAG: The assistant answers from the centre's own documents, through OpenAI file search.
Read the Paradigm case study
Method
How we bring Document RAG into a project
Scoping starts from real questions: the ones your staff currently put to an experienced colleague or search for in vain across shared folders. We gather a list of them, each paired with the document that holds the correct answer. That list becomes the project's evaluation set, and it also decides which sources are indexed first.
Preparation covers how documents are split so that each passage still makes sense on its own, which metadata are worth keeping (date, department, version) and access rights: a user should only receive extracts from documents they can already open. We assess retrieval, which must surface the right passage, separately from answer writing, which must stick to it. Keeping the two apart means the right link in the chain gets fixed when an answer is wrong.
Once deployed, the index updates whenever a document is added, changed or removed, and every answer keeps a record of the passages it used. Costs split between the initial indexing and each question asked, and we present them to you that way. Finally, we hand over the procedure for adding a source, excluding one or rerunning the evaluation.
For your business
What it changes in practice
In many small businesses, part of the know-how lives in a few people's heads: the current procedure, a clause in a standard contract, the answer given to a customer last time. Every question interrupts someone, and one absence is enough to hold up a file. A tool that answers from your documents, showing where the information comes from, cuts down those interruptions and helps a newcomer find their way without constantly asking colleagues.
It also has a less obvious effect: it exposes documents that contradict each other or were never updated, because its answers quote them. On confidentiality, only the passages relevant to a question are sent to the model, and existing access rights are respected. Your documents and their index stay in a store you own, and the model that writes the answers can be replaced without rebuilding everything. Staff come to trust it through the sources on screen, which they can check for themselves.
Our reading of it
What we take away from Document RAG
Assisted document search answers one precise question: what do our documents say about this? The principle is to split your content, index its meaning, retrieve the passages closest to the question, then let the model write only from those passages, sources shown. What changes everything is that the answer states where it was found.
Quality comes from the preparation, not from the model: splitting that respects document structure, version handling, excluding whatever the user has no right to read. We store the vectors in PostgreSQL rather than adding a specialised database for as long as the volume allows — one less dependency to administer. The sequence of steps is written with LangChain once it becomes conditional. On a ten-page corpus none of this machinery pays for itself, and a plain model call is enough.
Decision guide
Should you choose Document RAG?
RAG or fine-tuning a model on our documents?
To make a model answer from your files, RAG is almost always the right choice. Documents stay in a store you control, an update takes effect as soon as it is re-indexed, and each answer can cite the original passage. Fine-tuning changes how a model behaves, not what it reliably knows: you cannot remove a document from it or trace where a statement came from. We keep fine-tuning for teaching a style or an output format.
RAG or plain keyword search across our files?
If your staff know the exact terms and are looking for a specific document, a good keyword search engine is enough and costs nothing per query. RAG makes a real difference when questions are asked in everyday language, when the answer is spread across several texts, or when it is worded differently from the question. It then returns a written answer with its sources. A hybrid of keyword and semantic search often gives the most consistent results.
RAG or simply pasting all our documents into the model prompt?
Sending the whole collection with every question works for a handful of short documents. Beyond that, each request becomes expensive, responses slow down and the model makes poorer use of information buried in the middle of a long text. Above all, you transmit all your data every time, including material unrelated to the question. RAG selects only the relevant passages, which cuts cost, limits the data exposed and lowers the risk of error.
Client questions
What clients ask us about Document RAG
Can the tool still make up answers?
The risk drops considerably but does not vanish. The model is instructed to answer only from the retrieved extracts and to show its sources, so you can check them in one click. When no relevant passage exists, it should say so rather than fill the gap. Quality also depends on your documents: conflicting or outdated versions will lead to conflicting answers, however well the system is built.
Are our documents used to train an AI?
RAG does not train any model: your documents are split up and indexed in a dedicated store, and only the useful passages are passed on when a question is asked. For sensitive files, both that store and the model can be hosted in the European Union or on your own infrastructure, so nothing leaves your perimeter. We set out where each component sits during scoping.
Do we need to tidy up our files before starting?
Often, yes. A RAG system reports what it finds: if several versions of the same procedure coexist, or outdated documents are never archived, the tool will quote them with equal confidence. So we begin by identifying which sources are authoritative and which should be excluded. If that sorting cannot be done, we tell you before any development is committed.
Will every member of staff see every document?
No, not if the project is designed properly. Each indexed passage keeps the permissions of its source document, and retrieval filters results according to who is asking. A salesperson will therefore not receive an extract from an HR file they could not open in your shared folders. We test that separation with accounts for different roles before the system goes live.
What kinds of documents can be included?
Plain text, PDFs, office documents, intranet or wiki pages and archived emails can all be brought in without particular difficulty. Scanned documents first go through character recognition, and the quality depends on how clean the scan is. Complex tables and diagrams lend themselves less well to the process, so we check a sample of your files to see what is usable before setting the scope.
Going further
Document RAG sits within our Artificial intelligence
A technology does not make a project. The method around it decides the outcome: scoping, prototypes, short cycles, and a clean way out.
Neighbouring technologies
A Document RAG project in mind?
The first conversation is free. We will tell you plainly whether this technology is the right one, and if not, which is.
Updated on