Skip to content

Model channels

A model channel is a provider + API key + endpoint that supplies models to your account. You don't need one to get started — a default chat model is open on every account — but adding your own lets you use a specific provider, your own key, or a private/self-hosted endpoint.

Add a channel

  1. Open Settings (gear icon, top-right) → Model Service.
  2. Click + Add channel.
  3. Pick a provider from the dropdown.
  4. Enter your API key, and a custom Base URL if needed.
  5. The platform auto-detects available models for that key.
  6. Choose the active model for this channel and save.

You can toggle a channel on/off without deleting it, and keep several channels for different providers side by side.

Add a model channelAdd channel (shown in Chinese): pick a provider → enter API key + base URL → auto-detect and select models

Supported providers

Anthropic (Claude)OpenAI
Google GeminiAzure OpenAI
DeepSeekKimi (Moonshot)
Zhipu (BigModel)MiniMax
DashScope (Alibaba Qwen)Volcengine
Ollama (local)LM Studio (local)
Custom (any OpenAI-compatible endpoint)

For a custom OpenAI-compatible endpoint, choose Custom, paste the base URL and key, and the platform will list its models.

Choosing the chat model

The model used for a conversation is selected in the chat panel, from the models your channels expose. New accounts start with a default chat model open, so this is optional until you want something else.

The knowledge-base models are platform-wide (not here)

A knowledge base needs three models under the hood:

  • Embedding — turns text into vectors for search.
  • Rerank — reorders retrieved passages by relevance.
  • Contextual — writes a short context prefix for each chunk at ingestion time.

These run on the platform's global defaults and are not configured in your channels. This is deliberate: all projects share one vector store, so the embedding model (and its vector dimensions) must be consistent across every user. Letting each user pick a different embedding model would make existing vectors incompatible. So retrieval "just works" for everyone, with no setup.

Advanced

If a deployment does allow overriding these per-user, it's an administrator setting — a regular account never needs to touch it. Your channels are for the chat model, and for image / speech models if you use those features.

Next