Skip to content

Configuration

Jump to bottom
Lucas Skog edited this page Jul 10, 2026 · 3 revisions

Configuration

All settings live under Settings in the app. There are five groups: Audio, Transcription, Summarization, Speakers and Appearance. This page explains every field.

referat uses two independent providers: transcription (speech → text) and summarization (text → minutes). You configure them separately, so you can mix modes — for example transcribe locally and summarize in the cloud.

API keys never leave your machine in plaintext. Keys are encrypted with Windows DPAPI via Electron safeStorage and stored as ciphertext in %APPDATA%\referat\settings.json. The plaintext key is never sent back to the app's interface, and referat refuses to store a key if OS encryption is unavailable rather than writing it in plaintext.

Transcription

The service that writes out what was said. Any OpenAI-compatible /v1/audio/transcriptions endpoint works.

Fields

  • Preset — prefills the fields below for a known provider (see the table).
  • Base URL — the provider's /v1 root. referat appends /audio/transcriptions.
  • Model — the transcription model name.
  • Language — an ISO code such as sv. Leave empty to let the server auto-detect.
  • API key — optional; leave empty for local servers.

Presets

Preset Base URL Model (default) API key
Local server http://localhost:8000/v1 KBLab/kb-whisper-large no
OpenAI https://api.openai.com/v1 whisper-1 yes
Azure OpenAI https://<resource>.openai.azure.com/openai/v1 whisper yes
Custom (empty — enter your own) (empty) depends

Summarization

The service that turns the transcript into the actual minutes. Two API flavors are supported: OpenAI-compatible chat completions and Anthropic messages.

Fields

  • Preset — prefills the fields below.
  • API type — OpenAI-compatible or Anthropic.
  • Base URL — the chat endpoint root (/chat/completions is appended for the OpenAI-compatible flavor).
  • Model — the model name.
  • API key — optional for local servers, required for cloud providers.
  • Minutes template (advanced) — see below.

Presets

Preset API type Base URL Model (default) API key
Local server OpenAI-compatible http://localhost:11434/v1 llama3.1 no
OpenAI OpenAI-compatible https://api.openai.com/v1 gpt-4o-mini yes
Azure OpenAI OpenAI-compatible https://<resource>.openai.azure.com/openai/v1 gpt-4o yes
Anthropic Anthropic https://api.anthropic.com claude-3-5-sonnet-latest yes
Custom OpenAI-compatible (empty — enter your own) (empty) depends

Azure OpenAI specifics

referat targets Azure's OpenAI-compatible v1 endpoint, whose base URL is shaped:

https://<resource>.openai.azure.com/openai/v1

Replace <resource> with your Azure OpenAI resource name. Enter your Azure key in the API key field. referat sends the key both as an Authorization: Bearer header and as an api-key header, so it works whether the endpoint expects the OpenAI or the Azure header shape. Use your deployment name as the model.

Anthropic flavor

With API type: Anthropic, referat posts to <base>/v1/messages using the x-api-key header and anthropic-version: 2023-06-01. If your base URL already ends in /v1, referat strips it so the path doesn't become /v1/v1/messages. Leave the base URL as https://api.anthropic.com unless you use a proxy.

The minutes template

The minutes template (under Advanced) is the prompt that shapes the output. It contains a {{transcript}} placeholder, which referat replaces with the meeting transcript before sending. (If you remove the placeholder, the transcript is appended to the end of your prompt instead.)

The default template is in Swedish and asks the model for four sections — Sammanfattning (summary), Beslut (decisions), Actionpunkter (action items with owner and deadline) and Öppna frågor (open questions) — and instructs it to answer in the transcript's language and to use only information present in the transcript. Edit it freely to change the structure, language or tone; the default works without any changes.

Tip: prefer a non-reasoning model. Reasoning-heavy models can return an empty answer, which referat surfaces as an error instead of saving empty minutes. See Local AI Setup.

Speakers

Optional speaker identification ("who said what") — transcript segments are labelled Talare 1, Talare 2, … and the labels can be renamed; the names flow into the minutes when the protocol is regenerated. Off by default. Requires the local companion server — see Speaker Diarization for the full setup. In the app's Swedish UI the group is called Talare.

Fields

  • Identify speakers (Identifiera talare) — the on/off toggle. When off, meetings are processed exactly as before.
  • Server address (Serveradress) — the diarization server's address. Default http://localhost:8300.
  • Test connection (Testa anslutning) — checks the server's /health endpoint; a network error means the address is wrong or the server isn't running.

A diarization failure never blocks the minutes: the meeting gets a warning note and the protocol is still produced, just without speaker labels.

Audio

  • Microphone — which input device is recorded.
  • Record system audio — capture what plays on the computer (for example other people in a video call). This is what lets referat work with Teams, Zoom and Meet without a bot. On when possible; if system audio can't be captured, referat records the microphone only and tells you.

Appearance

  • Theme — Follow the system, Light or Dark.

Re-running onboarding

Settings includes an option to run the setup guide again, which re-walks the provider choice, connection test and microphone test.

Related pages

Clone this wiki locally