Local Mode data flow

Outcome

In Local Mode, Synthezia uses local programs and a loopback Ollama service for the core audio, transcription, summary, chat, and embedding path. Install the required local models before expecting this path to work without model-download network traffic.

Local processing diagram

Audio or recording
  -> local FFmpeg conversion
  -> local whisper.cpp transcription
  -> local Ollama summarization and chat
  -> local embedding and indexing
  -> local SQLite, audio, and model storage

The Ollama runtime communicates through a loopback address on the same device. It is a local HTTP service, not an external AI-provider request.

What stays on the device in this path

A selected media file or live microphone recording is converted locally to a normalized WAV file.
whisper.cpp receives the local audio path and local Whisper model, then returns a transcript and segments for local storage.
Summaries and per-session chat send transcript and session context to the local Ollama runtime.
Provider-aware transcript and summary embeddings use the local Ollama embedding endpoint while Local Mode is selected.
Global search first stores transcript chunks in SQLite and then adds background embeddings through local Ollama.

Network needed separately

Local Mode is not the same as never connecting to a network. A Whisper model download contacts the configured Hugging Face source, and an Ollama model pull asks the local runtime to retrieve models from its configured registry. Those downloads are separate from processing a recording after the models are present.

Common failures

If a required Whisper or Ollama model is missing, Synthezia asks for model preparation or reports that the local runtime is unavailable.
If the local Ollama runtime cannot start, summaries, chat, and local embeddings cannot complete; local transcription remains a separate Whisper path.
If the device is offline before the required models are installed, model acquisition cannot complete.