PII for model APIs #692
|
I just want to understand, what happens to our PII when we use enterprise models as API endpoints? Do these data goes to model request as context? Is it safe to use this way? |
Replies: 2 comments 1 reply
|
From the public source, treat anything in the chat context as sent to the selected provider API. The local/privacy bit mainly means Odysseus is self-hosted and stores API keys locally encrypted; it doesn't mean remote API calls are automatically PII-redacted. So for enterprise models, the safety answer depends on your provider contract/retention settings plus your own redaction policy. If you need strict handling, use a local model or put a scrub/proxy layer in front of the endpoint and avoid sending raw PII in the first place. |
|
Thank you, I think local models are way to go for the use case |
|
Yes. If PII is present in messages, tool results, or retrieved context, it can become part of the remote model request. Provider retention rules still matter, but the safest control is to remove or transform the data before the API call. For a TypeScript application or gateway, Flare Redact does this locally and can use reversible placeholders when the application needs the original values back. I maintain it here: https://github.com/umudhasanli/flare-redact |
From the public source, treat anything in the chat context as sent to the selected provider API.
src/llm_core.pybuilds OpenAI-style / Anthropic / Ollama payloads frommessagesand posts them to the configured endpoint, so if PII is in a user message, tool result, retrieved memory, or document context, it can be part of that request.The local/privacy bit mainly means Odysseus is self-hosted and stores API keys locally encrypted; it doesn't mean remote API calls are automatically PII-redacted. So for enterprise models, the safety answer depends on your provider contract/retention settings plus your own redaction policy. If you need strict handling, use a local model or put a scrub/proxy l…