Your RAG bases
Selected corpora
Context retrieved
Protection to match
Secured knowledge basesYour knowledge.
Your knowledge.
The AI models you choose.
Your rules.
Bring your organisation’s knowledge together and put it to work through a multi-AI platform, keeping access, processing and budget under control.
Your business context The choice of models A data path you control
Narratheque on video
A multi-AI platform at the heart of your work.
What you need, before the technology“We want to work
“We want to work
with our own data.
Not depend on a single AI.”
Your answers are not only out on the internet. They are in your procedures, your files, your lessons learned and what your teams know.
You want to plug AI into your data while choosing what may be sent, to whom, within which framework and for what result.
The Narratheque answerYou bring the context,
You bring the context,
the multi-AI platform brings the models.
RAG is what grounds the AI in your data: it searches for and prepares the knowledge a request needs. Multi-LLM lets you pick the model that suits the task, the users and your processing requirements.
Documents & filesWebsitesConnected sources
+
Multi-LLM
MistralClaudeOpenAIGeminiLocal models
Models chosen by
the operation, the users
and how sensitive the data is
InternallyAssistants, search, writing, summaries
Open to an audienceChatbots for clients, partners or visitors
Schematic view. The sources, models and processing available depend on the configuration. An external model may receive the context the answer requires.
Three complementary technologies
Retrieve. Structure. Protect.
The first two organise how AI searches and uses your knowledge. The third adds protection for identifying information. They answer different needs and can be combined, depending on what is available.
Vector RAG
“I want to find the right information in my documents.”
The system searches the corpus for the relevant passages and uses them as context to draft an answer.
ExampleFinding a returns procedure across the manuals and the support documentation.
Understand document searchLLM Wiki
“I want to organise scattered knowledge and connect the ideas.”
An AI-assisted wiki approach structures the corpus into connected, summarised knowledge, with references to maintain and to check.
ExampleOrganising how a new joiner is onboarded into a coherent set of subjects.
Explore the wiki approachEncrypted anonymisation
“I want to limit how much sensitive information is exposed.”
The protection mechanism identifies and replaces or encrypts the elements it covers before the intended processing. Its scope must be checked against your documents.
ExampleSubstituting names and contact details in the context used to draft an answer.
Understand the stronger protection“Encrypted anonymisation” refers here to a protection mechanism whose scope must be specified. If the identifiers can be restored, this is reversible pseudonymisation. Encrypting storage, on its own, does not hide the content from the model.
Multi-LLM, in the service of the workMulti-AI platform: choose the model
Multi-AI platform: choose the model
according to the work in hand.
The knowledge base remains your footing. You can choose different models to write, to summarise or to handle a sensitive context, among those available in your environment.
- By operation
Expected quality, length of context, cost and speed.
- By user
Set which models and which bases each team may reach.
- By data
Choose the protections and where the processing happens.
Example of rules to set
CommunicationsPublic content · writing
Models allowed for preparing publications.
Customer serviceProduct documentation · support replies
Trade corpus and human approval before sending.
Authorised teamSensitive documents · summaries
Protected path or local processing, as your requirements demand.
An illustration of governance: which restrictions are possible per user and per operation must be confirmed in the configuration you settle on.
Sovereignty & keeping your knowledgeDo not confuse storage,
Do not confuse storage,
processing and reuse.
To know what you are really protecting when you plug AI into your data, you have to look at the whole path, from the document you import to the result you send on.
Where does the data go?
Check where it is stored, how documents are prepared, where generation happens and which tools receive the result.
What may it be used for?
Check the provider’s retention and reuse terms, particularly around training or improving models.
What are you making reachable?
Decide which knowledge may feed an answer and which audiences may receive it.
And what about your knowledge being distilled?
A chatbot can pass on part of your expertise through its answers. Even without retraining the model, what is made reachable can be collected and reused. A dedicated public corpus, framed answers and access limits reduce that exposure without removing it.
Examine the security frameworkYou choose the destinationA multi-AI platform, internally
A multi-AI platform, internally
or open to an audience.
For your staff
Question the documents, find a method, prepare a summary and share the references the work needs.
DécouvrirThrough a chatbot
Open a selected corpus to clients, partners or visitors. Each audience gets a scope of its own.
DécouvrirAn internal base is not meant to become public by default. Opening it through a chatbot should be an explicit decision about the content and the audience.
A first scope is enoughA multi-AI platform starts with the data
A multi-AI platform starts with the data
that answers a real need.
One base to begin with, then others to separate departments or audiences. The Team plan follows that progression.
Your first project
- Choose the task to make easier.
- Gather the useful sources.
- Set the access rights and the processing.
- Test the answers and open up the use.
An agency or a trade adviser can support you at every step.
Frequently asked questions
What you want to know about the multi-AI platform.
Does RAG train a new model on our documents?
The principle of RAG is to search a corpus for knowledge and then hand context to a model. It does not require retraining that model. The provider’s processing and reuse terms still have to be checked.
Do we have to choose just one of the three technologies?
They do not play the same role. Vector search and the wiki organise access to knowledge; the protection mechanism reduces how much of certain information is exposed. Which combinations actually apply depends on the configuration.
Can we change model without rebuilding our bases?
Separating the knowledge foundation from the models lets the corpus remain your footing. Compatibility and any processing adjustments have to be checked for each configuration.
Can you promise our knowledge will never be distilled?
An absolute guarantee would be misleading. What you can do is limit what is sent to providers and what is given away in answers, then set rights, reuse terms and access limits to match.
