Skip to content

Adapta

A self-hosted platform for customizing and serving private language models. Upload documents or train a fine-tuned adapter on your own GPU — then expose the result as an OpenAI-compatible API. Your data never leaves your servers.

Prototype maturity

This is a one-person project at prototype stage. The pieces are wired up — RAG, LoRA training, eval gate, multi-tenant serving, vision fine-tunes — but the system has not been independently audited or validated at real team scale. If you are deploying with sensitive data, review the auth, error-handling, and eval-gate code paths yourself before going to production. See Brand §0 for the full disclosure.

This documentation serves two purposes: it is the operational and developer reference for the platform, and it is a concrete learning resource for anyone building or studying AI infrastructure on a real, production-grade codebase.

Where to start

  • AI Infrastructure Guide

    Not sure what GGUF, LoRA, embeddings, or a vector store are? Start here. This section explains the full technology stack — every component, why it was chosen, and how it fits — written for engineers who know web development but are new to AI infrastructure.

    AI Infrastructure Guide

  • User Guide

    For the operator running the platform. How to create projects, give a model knowledge from your documents, change how it behaves through fine-tuning, and hand an application a scoped API key.

    User Guide

  • Developer Guide

    For engineers working on the codebase. Architecture, the contract-driven workflow, the three CI gates, and the auto-generated code reference. Read Learning the system before your first change.

    Developer Guide

  • Reference

    The locked facts: product scope, the live API reference generated from the OpenAPI spec, the operations runbook, and the architecture-decision record.

    API Reference

What the platform does

A company deploys Adapta on its own server with Docker Compose. Inside that deployment, teams create Projects. Each project ends in a private model endpoint consumed with a scoped API key:

Knowledge (RAG) — upload documents → they are parsed, chunked, and embedded into a per-project vector store → the model answers grounded in your documents, with citations pointing at the source passages. The model's weights never change. CPU-only.

Fine-tuning (LoRA) — provide an instruction dataset → a GPU worker trains a LoRA adapter → the adapter must pass an evaluation gate before it can serve → the endpoint serves base model + adapter. On a vision base model, the dataset is a zip bundle of image + instruction examples and the endpoint accepts images — for invoice extraction, visual QC, document AI.

Both together — a fine-tune project can also index documents. Its endpoint then injects retrieved context and applies the adapter in one call: facts from retrieval with citations, tone and structure from the fine-tune.

The POST /v1/chat/completions endpoint is the only protocol customer applications call. If your code already calls OpenAI, it already calls Adapta.

New to RAG, LoRA, GGUF, or embeddings?

The AI Infrastructure Guide explains every concept in terms of engineering ideas you already know — databases, caches, job queues, and CI pipelines. No prior ML background required.

Learning paths

Your goal Recommended path
Understand the AI infrastructure AI Infrastructure Guide → Stack & decisions → Learning the system
Set up and use the platform User Guide → Operator console → Consuming the API
Contribute to the codebase Developer Guide → Architecture → Workflow
Understand a specific decision Stack & decisions — each component has an explicit "why not X?" section
Operate the platform in production Operations runbook

How this documentation stays correct

Docs rot. This system is built so the parts that drift hardest can't:

Part How it stays in sync
API reference Rendered from specs/openapi.yaml — the same file the server is tested against. Change the API, the docs change.
Code reference Auto-extracted from source docstrings (mkdocstrings). The reference is the code.
Internal links mkdocs build --strict runs in CI. A broken internal link fails the build.
Roadmap Canonical source at TODO.md in the repository — the roadmap page links to it directly rather than embedding it (the file links to source paths that only resolve in-repo).
Scope vs. tasks Reference docs describe what is. The Roadmap is the only place with open work. The two are never mixed.