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, 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) | Finetune (adapter) | An eval set (or synthesize one from documents) and an adapter to import (import converts and gates it on CPU) | CPU |
The platform never generates images; vision base models can read them (see Consuming the API).
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 (adapters). If you're unsure which you need, the rule of thumb: facts the model should look up → Knowledge; a way the model should act → Adapter.
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
Finetune (adapter):
upload (or synthesize) an eval set → import an adapter → it must pass the eval gate
→ create an endpoint → get an API key → your app calls the endpoint
In-product training was removed on 2026-10-10. Adapters are imported instead, and the import and its CPU eval gate run on every upload; a finetune endpoint exists once an adapter has passed the gate. The gate is the product's safety promise that an unverified adapter never reaches your users. See the eval gate and the bring-your-own-adapter guide (what to import, how it is scored).
Both together (the production pattern): a finetune project can also index documents. Its endpoint then answers from your documents, in your adapter's 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 adapters (real-life cases) and the combined pattern end to end.
- Migrating to Adapta — coming from Flowise, Cognita or R2R, or from a setup that trained adapters for you: concept mapping, what has no equivalent, and the API path.
- Consuming the API — how your applications actually call the endpoint with the OpenAI SDK, plus citations and usage.