The idea is compelling: upload your notes, journal entries, emails, documents, and photos into an AI system that knows your life well enough to help you find things, connect ideas, and act as a knowledgeable assistant. This category of tool — variously called personal AI, second brain, life OS, or knowledge management AI — has grown substantially, and building a private version using your own data is now a practical option for non-developers.
But the pipeline between “your documents” and “a useful personal AI” runs through several services, each with its own privacy posture. Understanding where your data actually goes at each step is not a paranoid exercise — it is the basic due diligence that determines whether your personal AI archive remains personal.
What a RAG Personal AI Actually Does With Your Files
Most personal AI systems that can work with your own documents use a technique called retrieval-augmented generation, or RAG. Here is what that pipeline looks like:
Step 1: Ingestion. Your documents, images, voice notes, and other files are uploaded to the system. For a cloud-based personal AI service, this means your files leave your device.
Step 2: Embedding. Each piece of content is converted into a numerical representation called an embedding — a vector that captures the semantic meaning of the content in a form the AI can compare and search. This conversion is done either by a model running on the service’s servers, or via a call to a third-party embedding API (often OpenAI or a similar provider). Your content is sent to whichever service does this conversion.
Step 3: Vector storage. The embeddings are stored in a vector database. The raw content may also be stored in full, or may be chunked and stored in pieces. Either way, a representation of your content persists on remote servers.
Step 4: Retrieval. When you ask a question, the system searches the vector database for content semantically related to your query, pulls the relevant chunks, and combines them with your question as context for the AI.
Step 5: Generation. The query plus retrieved context is sent to a large language model — almost always via an external API — to generate the answer. The context derived from your documents is visible to the LLM API at this step.
At steps 1, 2, and 5, your actual document content — or a close derivative of it — is transmitted to external services. The privacy properties of those services determine the privacy of your personal AI system.
The Three Layers of Data Exposure
The personal AI service itself. The service you sign up for — whether it is a commercial product like Notion AI, Mem.ai, or a newer personal AI app — stores your documents and your interaction history. Its privacy policy governs how long this data is retained, whether it is used to train their models, who else has access, and what happens if the company is acquired or goes bankrupt.
Many personal AI services explicitly use user content to improve their models. This may be how they keep costs sustainable, but it means documents you uploaded as private notes become part of a training corpus used to improve a general-purpose product.
The embedding and LLM APIs. Cloud-based personal AI services typically call external LLM APIs (OpenAI, Google, Anthropic, Cohere, and others) to process your queries. The terms of service for these APIs — particularly their API tiers used by developers — generally do not use the data for training by default. But “not training on it by default” is different from “not logging it.” API providers typically retain request data for some period for safety, abuse monitoring, and operational reasons.
If the personal AI service is calling an API and your document content appears in the prompt at step 5, the API provider has access to that content for the duration of the request and potentially for the retention period they specify. Understanding which APIs your personal AI service calls is part of understanding its privacy profile.
Third-party integrations. Many personal AI tools offer integrations with your email, calendar, cloud storage, note-taking apps, and productivity tools. Connecting these does not just give the personal AI access to your data — it creates persistent OAuth tokens that the service holds, and may create ongoing sync that continuously imports new content. Granting access to an integration is a much broader commitment than uploading a single document.
Photos as Personal AI Input: A Higher-Stakes Category
Text documents have sensitivity. Photos often have more, and in different ways.
When you upload photos to a personal AI system that can understand and index image content, the system must analyze the images — either on its servers or by calling a vision API (again, typically from a third-party LLM provider). This analysis may extract:
- Faces and people: Even without explicit facial recognition, modern vision models can describe who appears in a photo (“a woman in her 30s with dark hair,” “two children on a playground”).
- Locations: Landmarks, storefronts, vehicle plates, and GPS metadata embedded in image files can reveal where and when photos were taken.
- Sensitive contexts: Medical imaging, documents photographed for storage, medication bottles, personal moments. Vision models will attempt to describe whatever is in the image.
The sensitivity of photo-based analysis is qualitatively different from text analysis. A journal entry you wrote contains what you chose to put into words. A photo contains everything you captured, including things you did not consciously intend to record.
The On-Device Difference
A personal AI system that runs entirely on your device — storing embeddings locally, running inference locally, never sending document content to external APIs — has a fundamentally different privacy profile.
On-device personal AI exists: tools like Obsidian’s local AI plugins, Ollama with a local language model, and several open-source RAG frameworks can run the full pipeline on your machine. The trade-offs are real: local models are less capable than cloud frontier models, the setup requires more technical effort, and queries run more slowly.
But the privacy guarantee is also real: your document content does not leave your device to be processed. If you load a local embedding model and a local language model, your files, your embeddings, and your queries stay local.
The question each person has to answer is whether the capability difference between local and cloud AI is worth the privacy difference. For a casual note search assistant, the gap may not matter. For a personal AI that you are feeding with medical records, financial documents, sensitive communications, or years of private journal entries, it matters more.
What “Private” Actually Means for Each Service
Reading a personal AI service’s privacy policy with attention to specific questions will tell you more than a general summary.
Does the service use your content for model training? Explicitly, not by inference. If the policy says “we may use your content to improve our services,” that typically means yes. If it says “we do not use your content to train our models,” note whether this applies to free tiers, paid tiers, or all tiers.
Which third-party APIs does it call? Some services disclose their model providers. Others do not. You can sometimes infer from the service’s pricing structure — a service that costs much less than the API calls it would need to make is likely not making those calls for every user, or is using a weaker model.
What happens when you delete your content? Deletion should remove the raw content, the embeddings derived from it, and any cached versions. The policy should specify this explicitly and give a timeline.
What are the data retention policies for your interaction history? Your queries are as revealing as your documents. The questions you ask reveal what you are worried about, what you are planning, and what you are curious about. How long are these stored?
What happens in an acquisition? Personal AI services are attractive acquisition targets — they hold deeply personal data and a loyal user base. The privacy policy should specify what happens to data in an acquisition scenario. “Subject to change with notice” is not the same as “data will not be used differently.”
Building a Private Personal AI Using Separate Services
One approach that avoids trusting a single service with the full pipeline is to use separate services for separate parts of the pipeline.
Store your source files — documents, photos, notes — in a storage service chosen for its privacy properties: encrypted at rest, no content analysis, GDPR/CCPA compliant, transparent deletion policies. This service holds your original files and is your authoritative archive.
Run embedding and retrieval locally, using an open-source vector database (Chroma, Qdrant, Weaviate, and others support local deployment). Embeddings are computed from local copies of your files.
Use a local language model for queries, or call a cloud API only when queries explicitly require it — and understand that the query plus retrieved context will be visible to that API.
In this architecture, your source files live in a storage service that never analyzes their content. The AI pipeline runs locally or with explicit, bounded API calls. Your archive and your AI assistant remain separable, which means you can change or delete either without losing the other.
This is more complex to set up than a commercial personal AI service. It is also more transparent about what is happening with your data at each step.
The File Storage Foundation
Whatever approach you use for personal AI, the foundation is where your source files live. That storage determines what gets ingested, what can be deleted, and what persists after you stop using any particular AI tool.
If your source files are stored in a service that runs its own content analysis, trains models on uploaded content, or shares data across its product ecosystem, then building a “private” personal AI on top of it is partially undermined before you start. The AI pipeline is only as private as the archive that feeds it.
Choosing storage with appropriate privacy properties — encrypted at rest, no content analysis of uploaded files, no AI training on user content, transparent deletion — is the step that determines the ceiling on what privacy the whole system can achieve.
daftei stores files without running content analysis, without training AI models on uploaded content, and with encryption in transit and at rest. It does not show ads and does not sell user data. Files uploaded to daftei can serve as the source archive for a personal AI system without adding a layer of analysis that the storage service performs on its own.
The AI tools you use on top of that archive are a separate decision — one you can update, replace, or remove without affecting the underlying files.
What to Actually Do Before You Upload Your Life
Before connecting any personal AI service to your files and photos, these are worth knowing:
-
Read the privacy policy for model training clauses. Look for language about using content to improve services or train models. If it is unclear, assume yes.
-
Audit which third-party APIs the service calls. If this information is not in their documentation, ask their support team. “We do not disclose our model providers” is an answer, and it tells you something.
-
Start with less-sensitive content. Test the system with content you would not mind being broadly visible before uploading your most sensitive documents.
-
Check the deletion workflow before ingesting. Know how to delete your content, and verify that deletion extends to embeddings and cached content, not just the raw files.
-
Keep your source files in storage you control independently. Do not let the personal AI service become the only place your files exist. The storage and the AI layer should be separable.
A personal AI trained on your own life can be genuinely useful. The privacy requirements to make it actually private are achievable. They require knowing what the pipeline does, not assuming that “personal” and “private” mean the same thing.