Under Settings → AI/LLM → AI Connectors you create connections to language models. A connector bundles the provider, model, credentials, and sampling parameters — it’s the foundation for all AI features. The overview shows, per connector, its name, provider, and properties (active/inactive).

+ Neue VerbindungFilter zurücksetzen
NameProviderEigenschaften
defaultOpenAI-kompatibelAktiv
semantic-compareOpenAI-kompatibelAktiv
geminiGoogle VertexAktiv

Reserved names: “default” and “semantic-compare”

Two connector names have a special role:

Name Effect
default powers LegalAi — the AI assistant in the case context. Without a default connector, LegalAi isn’t available.
semantic-compare enables semantic verification within LLM extraction templates — comparing extracted values by meaning (rather than plain text comparison).

Additional connectors with freely chosen names (e.g. gemini, sonnet) can be connected as well and assigned to individual extraction templates.

Creating a connection

+ New connection opens the connector dialog. It covers common providers via an OpenAI-compatible access method, but also supports cloud models (e.g. Google Vertex via project/endpoint/location).

KI-Konnektor
Name der Verbindung
Organisation
API Key
URL
Defaultmodel
Temperature
0.2
Timeout
60
MaxTokenSize
 
topP
 
topK
 
Thinking Modus
Token Budget
ausSSL-Zertifikatsprüfung deaktivieren
Fallback
Für Cloud-Modelle: Project ID · Endpoint ID · Location
Speichern
Field Meaning
Connection name freely chosen — or reserved (default / semantic-compare, see above).
Organization optional organization identifier for the provider.
API Key the provider’s secret access key.
URL the API’s endpoint address.
Default model the model used by default.
Temperature controls randomness — higher values increase the diversity of outputs.
Timeout the REST client’s time limit for a request (default 60).
MaxTokenSize the upper limit on the number of tokens per request.
topP / topK sampling parameters (nucleus and top-k sampling, respectively); topK = -1 considers all tokens.
Thinking mode / Token budget activates the model’s internal “thinking” and limits the maximum tokens usable for it.
Disable SSL certificate verification only for special cases (e.g. internal endpoints) — leave enabled as a rule.
Fallback a backup connector used if the request fails.
Project ID / Endpoint ID / Location details for cloud-hosted models (e.g. Google Vertex).

Where to get an API key for each provider

You generate the API key for the connector dialog directly in the respective provider’s developer portal — not in DEPLAW. The most common providers at a glance:

Provider Generate API key
OpenAI platform.openai.com/api-keys
Anthropic (Claude) console.anthropic.com/settings/keys
Google (Gemini / AI Studio) aistudio.google.com/app/apikey
Google Vertex AI (project/endpoint/location) console.cloud.google.com/vertex-ai
Mistral console.mistral.ai/api-keys
Azure OpenAI via your own resource in the Azure portal (resource → “Keys and Endpoint”)

Local and self-hosted models

Because the connector uses a freely chosen URL and an OpenAI-compatible access method, the language model doesn’t have to sit with a cloud provider: you can just as well connect a local or self-hosted LLM (on-premise or in your own data center) by pointing the URL at its endpoint.