Module 12 Lesson 2: Data Privacy and Anonymization
Protecting the prompt. How to ensure sensitive user data like PII doesn't end up in your AI logs.
Data Privacy: Cleaning the Input
Even though Ollama stays on your machine, your Logs and Vector Databases might still contain sensitive information. If a user pastes their Social Security Number or a Credit Card into your chatbot, and you save that chat history to a database, you have a security liability.
1. What is PII?
PII (Personally Identifiable Information) includes:
- Names, Emails, Phone Numbers.
- Physical Addresses.
- Financial IDs.
- Medical Diagnoses.
2. The "Pre-Processing" Layer
In a production app, you should use a "Cleaner" before the text ever reaches Ollama.
The Workflow:
- User input: "My email is alex@example.com. Help me write a bio."
- PII Filter: Use a Python library like
Presidioto find labels. - Result: "My email is [EMAIL]. Help me write a bio."
- Ollama: Only sees the cleaned version.
3. Scrubbing Vector Databases
In Module 10 (RAG), we learned to index documents. If you index an "Employee Salaries" PDF, that data is now in your Vector Store.
- Problem: Someone could ask the AI: "What is the salary of the CEO?" and the AI will find the answer in the vector store and reveal it.
- Solution: Metadata filtering. Tag your documents with "Public" vs "Private" and only allow the AI to search "Public" documents for normal users.
4. Disabling "Telemetry"
Ollama is very privacy-respecting, but some apps built on top of it (like some UI extensions) might try to send usage statistics back to their developers. Always check the settings of your UI layer (Open WebUI, Enchanted, etc.) to ensure "Usage Reporting" or "Anonymous Analytics" is turned OFF.
5. Compliance by Isolation
The ultimate privacy tool is the Firewall. By running Ollama on a machine that is physically disconnected from the internet, you have a 100% guarantee that no data can be sent to a third party. (See Lesson 4: Air-Gapped Environments).
Key Takeaways
- PII should be stripped before it reaches the AI log.
- Logs and Vector Stores are the primary places where "leakage" happens in local setups.
- Use automated filters to anonymize data in production apps.
- Verify the privacy settings of the UI tools you use with Ollama.