How AI processing works

PR-TOP supports four configurable AI providers. The active provider is selected by the platform operator via environment variables — therapists do not choose the provider directly.

• OpenAI — available models: gpt-4o-mini, gpt-4.1-nano, gpt-4.1-mini, gpt-4o, gpt-4-turbo, o4-miniaiProviders/openai.js:10
• Anthropic — available models: claude-3.5-haiku, claude-4-sonnetaiProviders/anthropic.js:11
• Google Gemini — available models: gemini-2.0-flash, gemini-2.5-flash, gemini-2.5-proaiProviders/google.js:10
• OpenRouter — a gateway to any of the above plus open-source models (deepseek, qwen, llama and hundreds more) via a single unified APIaiProviders/openrouter.js:11

The operator configures the deployment with one or more providers. Ask the platform operator which provider is currently active.

Only the minimum data required to complete each AI task is sent. No client names, Telegram IDs, email addresses, or other personally identifying fields are ever included in AI API calls.

• Session transcription: only the audio or video file is sent for the duration of the API call. No client identifiers or records are attached.
• Session summarization: the decrypted transcript text is sent to the AI provider. It is decrypted in memory immediately before the call and not written to any secondary location.
• Exercise generation: only the therapist's plain-text instructions are sent. No client diary entries or personal data are included.
• Semantic search: only the search query text is sent to generate a vector embedding. No diary or personal data is included.
• Chatbot assistant: only the user's question and relevant excerpts from the PR-TOP knowledge base are sent. No client-specific records are included.

AI is invoked for the following discrete, therapist-initiated tasks only:

1. Speech-to-text transcription of session audio and video recordings
2. AI-draft session notes and summaries — always require explicit therapist review and editing before useaiUsageLogger.js
3. Exercise content generation — therapist-triggered, result is editable before delivery to clients
4. Vector embedding generation for semantic search queries
5. Natural language query interface (Pro and Premium plans)
6. Between-session assistant chatbot (answers from the curated knowledge base, never from client diary data)

Data submitted to AI providers via REST API is processed transiently. Under standard API terms, providers do not store it in a persistent database:

• OpenAI API: processed in OpenAI's infrastructure (United States). API input/output is not used to train OpenAI models by default.
• Anthropic API: processed in Anthropic's infrastructure (United States). API data is not used for model training.
• Google Generative Language API: processed under Google Cloud terms. Data is not used to train Google models.
• OpenRouter: acts as a gateway and routes requests to the underlying provider. Each provider's own data-residency terms apply.

None of the above providers are based in the EU. If EU data residency for AI processing is a hard requirement, discuss provider configuration with the platform operator.

Each AI provider publishes a DPA or equivalent policy that covers how API data is handled:

• OpenAI — Data Processing Addendum
• Anthropic — Privacy Policy (API data)
• Google — Cloud Data Processing Addendum
• OpenRouter — Privacy Policy

If your practice requires a signed DPA directly between your clinic and each AI provider, verify with the platform operator which providers have been contracted and which DPAs have been countersigned.

Under standard API terms, AI providers do not retain submitted data beyond what is needed to complete the request:

• OpenAI API: inputs and outputs may be retained for up to 30 days for safety monitoring, then deleted. Not used for training.
• Anthropic API: data is not retained beyond the API call under current usage policies. Not used for training.
• Google Generative Language API: data is not used to train Google's models; retention follows Google Cloud data-processing terms.
• OpenRouter: forwards requests to providers; each provider's own retention terms apply to their copy of the data.

Provider policies may change. Review the latest version of each provider's DPA before making contractual commitments to clients.

PR-TOP does not currently offer a global 'AI off' account toggle. AI features are opt-in at the task level — nothing is sent to an AI provider unless the therapist explicitly triggers a task:

• Transcription is started manually — therapists choose to run it on individual sessions.
• AI-draft summaries are generated only when the therapist clicks 'Generate summary'.
• Exercise generation is therapist-initiated.
• Semantic search and the natural language query interface are invoked only when the therapist types a query.
• The chatbot is only active when the therapist opens and sends a message to it.

If your practice requires that no client data ever reaches an external AI provider, ask the platform operator to configure the deployment without AI provider API keys. In that configuration, transcription and AI summaries are unavailable; all other features (diary, sessions, exercises, SOS alerts) continue to function normally.