Configuring the AI backend
The chat is the only part of the platform that does not answer until it is configured. No language model ships with it — the choice is yours, and it decides where conversation content flows.
On this page
Models for chat, SQL and lavaflow
Chat, SQL functions and the Pyro step can use different models. SQL and running Pyro steps require a local provider or API access; personal Claude/ChatGPT subscription sign-ins cannot be used there. Subscription sign-in for chat is available only when explicitly enabled by the operator.
Only the chat waits
Login, worksheet, import, governance and search run without any backend — the embedding model for search ships with the platform.
Who may do this
Platform administrators configure the backend under KI → KI-Anbindung (AI → AI backend). An ordinary user does not see the page.
Providers and access routes
| Route | What it needs | Where conversation content goes |
|---|---|---|
| Claude (Anthropic) | An API key from the Anthropic console | To the provider |
| OpenAI | An API key from OpenAI | To the provider |
| Azure OpenAI | Resource name, API key and deployment name as the model | To the configured Azure resource |
| AWS Bedrock | Region, model ID and a Bedrock API key or IAM key pair | To the configured Bedrock service |
| Google Vertex AI | Project, region and service-account JSON; Gemini or Claude as the model | To the configured Vertex AI service |
| OpenAI-compatible | Address and key of a self-hosted model, for example behind vLLM | Wherever that model runs — on your own network: nowhere outside |
| Sign in with Claude | No installation-wide key; each user signs in with their own account | To the provider, under that user's own account |
Setting it up
Sign in as a platform administrator
In the local compose stack that is flaky-admin.
Open KI → KI-Anbindung
Pick a provider and paste the key. For a self-hosted model, add the base URL as well.
“Verbindung testen und speichern” (test and save)
The key is verified and then stored encrypted. It cannot be read back — not even by an administrator.
Select a model
Once the connection stands, you can choose the specific model offered by that provider.
What the choice means for your data
A self-hosted model behind the OpenAI-compatible option keeps conversations on your own network too. With a hosted provider, conversation content and tool results used in chat are sent there. Optional AI classification may also send names and types. The embedding provider is configured separately.
Signing in with your own Claude account
Instead of one shared key for the installation, each user can sign in with their own Claude account. The agent then runs in an isolated environment per request, and the gates remain fully in force — rights still come from Keycloak, not from the model account.
AI usage, prices and daily budget
Under Betrieb → Überwachung, “KI-Kosten” shows recorded token usage by user and model. Euro amounts are calculated only for models with configured prices; unpriced usage remains visibly unpriced. Claude subscription usage is not presented as an API bill. Embedding tokens are not yet included in this tracking.
Administrators can maintain prices in EUR per million tokens and import list prices on demand from the LiteLLM reference list, converted at the ECB rate. Manual contract prices are never overwritten. Source and exchange-rate date remain visible. Without internet access, prices can be entered manually.
A daily token budget can be set as a default and per user. It is checked before a chat run; a running model call may still exceed the limit. This is a token quota, not a guaranteed euro spending cap.
When the chat stays silent
- “No backend configured” — the expected state before setup, not an error.
- The connection test fails: with a self-hosted model, check the base URL and the egress guard first — an allowlist that does not name the target host blocks the call.
- Answers arrive but data is missing: that is not the model but your rights. The agent sees only what your identity may see in Trino.
Applies to: This page describes lavalake 0.201.0. Product and documentation live in separate repositories — when in doubt, what the console shows is authoritative.
Something missing?
If these pages do not answer your question, we will walk through the platform against your own case in half an hour.
Book a demo