User Guide¶
This guide is for the operator — the technical person who runs Adapta and sets it up for their organization. You do not need to read the codebase to use it.
If you are deploying the stack for the first time (installing Docker, the GPU toolkit, downloading a base model), start with the README quick start — that covers getting the containers running. This guide picks up once the stack is up and walks through actually using it.
What you can do¶
Everything in the product happens inside a Project. A project is one of two types, and the type decides how you specialize the model:
| You want to… | Project type | What you provide | Hardware |
|---|---|---|---|
| Make a model answer from your documents | Knowledge (RAG) | Documents (PDF, DOCX, TXT, MD, HTML) | CPU |
| Change how a model behaves (tone, format, a skill) | Fine-tuning (LoRA) | An instruction dataset (or synthesize one from documents) | GPU |
| Make a model read your images (invoices, forms, QC photos) | Fine-tuning (LoRA) on a vision base model | A .zip bundle of image + prompt → response examples |
GPU |
Image understanding only — the platform never generates images. See Image understanding for the bundle format and a worked example.
These are different mechanisms, not two kinds of 'training'
The console never calls RAG "training." When you create a project it asks "How do you want to specialize your model?" → Give it knowledge (RAG) vs Change how it behaves (fine-tuning). If you're unsure which you need, the rule of thumb: facts the model should look up → Knowledge; a way the model should act → Fine-tuning.
The two paths, end to end¶
Knowledge (RAG):
upload documents → they're indexed → create an endpoint → get an API key
→ your app calls the endpoint → answers come back grounded in your docs, with citations
Fine-tuning (LoRA):
upload (or synthesize) a dataset → start a training job → watch the eval gate
→ if it passes, create an endpoint → get an API key → your app calls the endpoint
The fine-tune path has one gate the RAG path doesn't: a trained adapter cannot serve until it passes evaluation. This is deliberate — it's the product's safety promise that an unverified fine-tune never reaches your users. See the eval gate below.
Both together (the production pattern): a fine-tune project can also index documents. Its endpoint then answers from your documents, in your trained voice and format — facts from retrieval (with citations), form from the adapter, in one call. See Knowledge + behavior together for a worked real-life example.
Where to go next¶
- Operator console — the browser walkthrough for both flows, from register to a live endpoint.
- Knowledge + behavior together — when to use RAG vs fine-tuning (real-life cases) and the combined pattern end to end.
- Consuming the API — how your applications actually call the endpoint with the OpenAI SDK, plus citations and usage.