AI Transparency Notice
Last updated: May 12, 2026
1. Introduction
This notice explains how GoldenRetriever uses artificial intelligence (AI) to process your files and provide search and question-answering functionality. It is provided in compliance with the EU AI Act (Regulation (EU) 2024/1689) and the General Data Protection Regulation (GDPR).
2. How We Use AI
Embedding Vector Generation
When you index files, the App converts their content into mathematical vectors (embeddings) that represent semantic meaning. This is powered by Google's Gemini Embedding 2 model (3072-dimensional vectors) called through the public Gemini API on your own API key. Depending on the file type, the App sends either extracted text or the original file bytes (for PDFs, images, iWork, RTF, HTML, audio segments, and video segments) inline to the Gemini API. The resulting vectors are stored locally on your Mac in a bundled Qdrant instance. We do not store your vectors and the bytes are not retained by us at any point.
Audio & Video Transcription and Description
Audio tracks and video frames are transcribed and described by Gemini 2.5 Flash as part of indexing. The same direct-from-your-Mac, your-own-API-key model applies. A local WhisperKit-based transcription engine is in development but is not yet enabled in shipping builds.
Semantic Search
When you search, your query is embedded via the Gemini API and compared against your locally stored vectors using Qdrant's hybrid search — combining dense semantic similarity with BM25 keyword ranking via Reciprocal Rank Fusion. Results are ranked by similarity score only — no automated decisions are made about relevance beyond mathematical comparison.
AI-Powered Q&A
When you ask a question, relevant chunks are retrieved via hybrid search and sent as context to your chosen Q&A provider:
- Google Gemini (default) — model
gemini-2.5-flash - OpenAI — default model
gpt-5-mini - Anthropic Claude — default model
claude-sonnet-4-5 - Ollama — local-only, default model
llama3.3onhttp://localhost:11434 - Custom — any OpenAI-compatible HTTP endpoint you configure
Provider terms govern retention and training-use of your data once it reaches that provider. For Ollama or a self-hosted custom endpoint, Q&A traffic can stay entirely on your own infrastructure.
3. What AI Outputs Mean
Search results are rankings by combined semantic-and-keyword similarity — not endorsements of accuracy. You should review results and verify them against original files.
Q&A answers are generated by your selected AI model, not retrieved verbatim from your files. They may contain errors, including hallucinations (plausible but false information), misinterpretations, or logical errors. Do not rely on AI answers for legal, medical, or financial decisions without independent verification.
4. Provider Data Protection
- Your file content is sent directly from your Mac to your chosen AI provider on your own API key — it does not pass through the Company's servers.
- Retention, training-use, and regional processing are governed by your provider's terms with you. For Google's Gemini API, that is Google Cloud Terms applicable to your account.
- All API traffic is encrypted in transit using TLS.
5. EU AI Act Classification
GoldenRetriever's AI systems are classified as Limited Risk under the EU AI Act. The App does not make automated decisions with significant legal or material effects on individuals. You initiate all AI processing, review all outputs, and retain final authority over any decisions. This complies with EU AI Act Articles 50 and 52 (transparency obligations) and GDPR Article 22 (automated decision-making).
6. No Automated Decision-Making
GoldenRetriever does not use AI to make decisions that produce legal or similarly significant effects. Specifically, AI is not used to approve/deny subscriptions, flag accounts, restrict access, or make eligibility determinations. All AI outputs are informational and presented for your review.
7. Opting Out of AI
You can disable or reconfigure AI features at any time in the app:
- Open GoldenRetriever → Settings → AI Provider to switch your Q&A provider, point at a local Ollama instance, or set a custom endpoint.
- To stop new embeddings being generated, stop watching folders (Library → Watched Folders) or remove your API key.
- Previously stored embeddings remain on your device and can be deleted in Settings → Data Management.
Without an AI provider configured, search and Q&A are unavailable.
8. AI Models
| Model | Purpose | Provider | Where it runs |
|---|---|---|---|
| Gemini Embedding 2 (3072-d) | Generate embedding vectors for files and queries | Google (Gemini API) | Google's servers, on your API key |
| Gemini 2.5 Flash | Transcription, media description, default Q&A | Google (Gemini API) | Google's servers, on your API key |
| OpenAI gpt-5-mini (optional) | Q&A only, if selected | OpenAI | OpenAI's servers, on your API key |
| Anthropic claude-sonnet-4-5 (optional) | Q&A only, if selected | Anthropic | Anthropic's servers, on your API key |
| Ollama llama3.3 (optional) | Q&A only, if selected | Local / self-hosted | Your device or your own infrastructure |
| Qdrant hybrid search | Dense + BM25 search via RRF | Self-hosted (bundled binary) | Your Mac |
9. Your Rights
You have the right to access information about AI processing, request deletion of your data, receive data in a portable format, and not be discriminated against for exercising these rights. For full details, see our Privacy Policy.
10. Contact
For AI-related questions, contact us at [email protected] with the subject “GoldenRetriever — AI Inquiry.”
If you have concerns about EU AI Act or GDPR compliance, you may lodge a complaint with your national data protection authority.
In Plain Language
GoldenRetriever uses AI to search your files and answer questions about them. By default it uses Google's Gemini API on an API key you provide — your file content is sent from your Mac directly to Google, not through us. You can switch the Q&A model to OpenAI, Anthropic, or a local Ollama model. Search results and AI answers aren't guaranteed to be accurate — always verify them. AI does not make decisions that affect you; you do.