Model settings

Two models do the work: a classification model picks the category, a reply model writes the draft. Use a small, fast model for classification and a larger one for replies, or the same model for both. Any OpenAI-compatible server works: vLLM (http://localhost:8000/v1), Ollama (http://localhost:11434/v1), llama.cpp server, LM Studio. Save before testing.

1 Classification model

Reads each email and picks a category. Always runs at temperature 0. Category corrections from reviewers are added to its prompt.

Server URL

Leave empty to use the reply model's server and API key — e.g. two models on one Ollama, or vLLM serving both.

Model name

Exactly as the server lists it, e.g. Qwen/Qwen3-4B or qwen3:4b.

API key (optional)

Only used when the server URL above is filled in.

Timeout (seconds)

Max tokens

The answer is a short JSON object; 512 is plenty unless thinking is on.

Structured output Sends a JSON schema so the model can only answer a valid category id (vLLM guided decoding / Ollama structured outputs). Turn off if the server rejects response_format. Disable thinking mode Sends chat_template_kwargs: {enable_thinking: false} for Qwen3-style hybrid models — much faster triage.

2 Reply model

Writes the draft reply from the category's response templates and reviewer feedback.

Server URL

Model name

e.g. Qwen/Qwen3-30B-A3B or qwen3:30b-a3b.

API key (optional)

Timeout (seconds)

Max tokens

Temperature

“New version” without feedback adds 0.3 to get a genuinely different draft.

Disable thinking mode Same switch as above, for the reply model. <think> output is stripped either way.

Our addresses

Email addresses and/or @domains that belong to us, separated by commas or new lines. When an email from one of these is selected in Outlook — a reply we sent, or internal mail — the add-in does not generate a reply (users can still choose “Generate anyway”). The signed-in Outlook user's own address always counts as ours. Listing the whole domain also skips mail between colleagues.

Reply signature

Available in the reply prompt as {{signature}}.

Both steps currently use the same model on the same server.