Cloud or local AI

Bristlenose needs an AI model for the analysis pass. For most people that is a cloud model reached with an API key. A local model is also supported, for cases where data must not leave the machine.

What a cloud model sees

With a cloud provider — Claude, ChatGPT, Gemini, or Azure OpenAI — Bristlenose sends the transcript text to the provider's API for analysis. It does not send the audio or video. The major providers state that data sent through their API is not used to train their models.

If the interviews were recorded on Zoom, Teams, or Google Meet, the recordings have already passed through a cloud service once; sending the transcript for analysis is a similar step.

For setup, see the per-provider guides under How-to guides. For the exact list of when data leaves the machine, see Privacy and data.

Running a local model instead

If transcript text must not leave the machine — an internal policy, or an unusually sensitive study — you can run a local model with Ollama. Nothing is uploaded for the analysis pass.

The trade-offs:

• Local models are slower than cloud models.
• They need a capable, recent machine with enough memory.
• They return well-formed structured output less reliably (about 85% of the time, against about 99% for cloud models), so runs occasionally retry.
• On the steps that need most reasoning — clustering quotes and naming themes — a local model is likelier to miss nuance than a frontier cloud model. A larger local model narrows the gap.
• AutoCode is not available locally.

Which to choose

For most studies, a cloud key is the simpler and stronger option, and the cost is small. A local model is the right choice when policy or sensitivity requires that nothing leaves the machine.