Agent Configuration

Configure system prompts, LLM settings, versions, and publishing for your Thotis RAG agents.

Thotis RAG supports two engine types:

Chat (chat): a conversational assistant that retrieves from datasets and generates answers. Backed by a RagFlow chat assistant.
Agent (agent): a workflow-based agent with multi-step reasoning, tool use, and orchestration. Backed by a RagFlow agent with DSL.

The system prompt defines the agent's personality, constraints, and instructions. It is sent to the LLM at the beginning of every conversation.

Best practices:

Be specific about the domain and tone.
Include explicit constraints (e.g. "Only answer questions related to French higher education").
Keep it under 2000 tokens for optimal performance.

The opener is the first message displayed to the user when a new session starts. It sets expectations and can suggest initial questions.

In the Settings tab you can configure:

Parameter	Default	Description
Model	deepseek-chat	The LLM model used for completions
Temperature	0.1	Controls randomness (0 = deterministic, 1 = creative)
Max tokens	2048	Maximum tokens in the response
Top P	0.9	Nucleus sampling threshold
Frequency penalty	0	Penalizes repeated tokens
Presence penalty	0	Penalizes tokens already in the conversation

Every change to an agent creates a draft version. Drafts can be validated and published:

Draft: editable, not yet live.
Published: the active version served to users. Only one version can be published at a time.
Rolled back: a previously published version that was replaced.

Edit the agent configuration.
Click Publish to validate the draft (checks: system prompt present, datasets bound, RagFlow healthy).
On success, the version is pushed to RagFlow and a runtime manifest is created.
The previous published version is automatically rolled back.

You can rollback to the previous version at any time from the Deployments page. Rollback re-publishes the last known good version.