Saphan

MODELS / ADAPTING WITHOUT STARTING OVER

Your line of work has its own words.
Its own gestures.
Its own exceptions.

A custom intelligence can take them into account. You still have to choose what to give it, connect to it or teach it.

Find the right path See the adaptation carried out in the lab

THREE DIFFERENT LEVERS

Custom does not always
mean training.

These approaches can be combined. We start with what the system must do, then we choose the lever.

GIVE IT ACCESS

“It must find
our information.”

Connect a search to your authorized documents and show the relevant sources. The information can be updated without training the model again.

Retrieval-augmented search — often called RAG. The quality of the search and the accuracy of the citations remain to be tested.

Try it on a small corpus

CONNECT TOOLS TO IT

“It must prepare
an action.”

Connect the useful functions: read a stock level, propose a record, check a format. Define what it can do on its own and what waits for approval.

Instructions and tools. A good connection can matter more than a bigger model.

Try preparing a job sheet

TEACH IT A BEHAVIOR

“It must work
in a precise way.”

Adapt an existing model using examples: classifications specific to the business, a recurring format, a style or a targeted behavior.

Fine-tuning, including LoRA: specialized training, not the creation of a foundation model from scratch.

Open the learning project

LEARNING / A FULL PROJECT IN ITS OWN RIGHT

Learn your examples.
Succeed on the ones
it has not seen, too.

The lab carried out a LoRA adaptation of a small open model. For your business, the question is not only “can we train?”, but “what does it actually improve?”

UNDERSTANDDefine a task, usable examples and what an error would cost.
LEARNFirst compare an existing model, then adapt if this lever is justified.
TESTKeep examples out of the training, test the errors too and compare the cost of use.
Talk about a specialized model What our internal trial establishes
Illustration of a compact computing workstation in a workshop
Saphan vision · generated illustration.

THE ENGINE IS NOT THE WHOLE SYSTEM

A base of your own.
Engines
that can evolve.

Local when it meets the need. A cloud service when its capacity justifies using it. Both can have their place.

Changing models means redoing the tests: results, speed, cost, license and the data sent. Interchangeability is built; it is not a matter of a button.

Explore a comparison test bench

Useful questions

Are my documents automatically used for training?

No. Searching them to prepare an answer and using them in training are two different processes. The scope, the data and the providers are decided before they are used.

Do I need to buy a big machine?

Not necessarily. It depends on the model, the volume, the response time and where it is used. The trial also helps to choose between existing hardware, an additional local resource and a remote service.

Can a completely new intelligence be created?

That is a research direction distinct from an integration service. It requires its own hypotheses, resources and experimental results. It belongs in Explore, not in a standard delivery promise.

To go deeper: RAG vs fine-tuning comparison, AWS · adapting models with PEFT, Hugging Face.