AI Architecture Handbook

The KAP AI backend as a Python package.

Introduction

The KAP AI backend lives in ai/ as a single pixi-managed Python package. The distribution is kavion-ai-server; the importable package is kavai (source under ai/src/kavai/). It exposes its capabilities to the web application over a thin HTTP/SSE Gateway speaking the AG-UI contract — the web-side wiring and the wire contract are in the Web Handbook (Web/AI Interface and the AG-UI contract).

All active engines are Google ADK / agent-runtime based: dataadk (the packaged reference), argus (the in-repo DataADK track, default), orion (next-gen ADK 2.x), and kawa (a tool-using AG-UI agent that runs in its own isolated environment). The delivery reference for the AI Assistant requirements is kavai_server in the KavApps submodule; KAP’s ai/ is the platform’s prototyping ground and the canonical source for the shared DataADK packages.

In this handbook

Page	What it covers
Package layout	The kavai source tree, subpackages, the CLI entry point, and packaged data.
Runtime & systems	The Gateway / ADK Runtime / Kawa adapter processes and ports, request routing, and the four systems.
Environments & operations	The pixi features and environments, the serve-*/test-* tasks, and how to run it locally.
DataADK packages	The kavai-agent-contract and kavai-dataadk packages, the contract, and their boundary rules.
Integrity pipeline	The API 571/584 integrity stages, schemas, reference data, and contract endpoints.