Where does my data go?

Documents, embeddings, conversation history, and vector indexes are stored in your self-hosted Dify stack. Only data required for inference leaves your network. It is encrypted end to end and processed only inside attested and verified confidential computing environments. Privatemode does not retain the data after processing.

Which Dify features work with Privatemode?

Chat apps, agent apps, knowledge bases with RAG, text embedding, and vision input all work with Privatemode. Any Dify feature that uses an LLM, embedding, or speech-to-text model through the OpenAI-API-Compatible plugin is supported.

Do I need /v1 in the base URL?

Yes. Use `http://host.docker.internal:8080/v1` for all model types, including embedding and speech-to-text. Omitting `/v1` causes a 404 during credential validation with plugin version 0.0.59.

How does encrypted inference affect performance?

Privatemode runs on hardware-based confidential computing, which adds only a small performance overhead on the server side, typically a few percent. The local proxy encrypts requests and responses with negligible added latency. Overall response times depend mainly on the selected model, request size, and network latency.

Confidential AI workflows with Privatemode and Dify

Dify is an open-source platform for building LLM applications. With Privatemode, every inference request is encrypted end to end and processed inside attested and verified confidential computing environments. Documents, embeddings, and conversations are stored on your infrastructure.

Introduction

Build AI apps without exposing your data

Introduction

Building AI apps in Dify shouldn't mean exposing your data

The challenge

Knowledge bases don't stay on your servers

Dify’s chatbots, agents, and workflows work best with powerful cloud models. With a conventional AI provider, however, prompts and document content used for RAG or embedding generation are available in plaintext during processing on the provider’s infrastructure. Self-hosting Dify does not change this.

The solution

Chat, agents, and RAG, encrypted end to end

The Privatemode proxy runs next to your Dify stack. Before any data leaves your system, the proxy uses remote attestation to verify that the Privatemode service is running in an expected confidential computing environment. Only after successful verification is the data encrypted and sent to that environment. Confidential computing protects the data during processing from both the infrastructure provider and us.

Benefits

Why use Privatemode AI with Dify?

End-to-end confidentiality

Learn how

Cloud-grade AI models

See models

Free tier available

How to get started

Set up confidential AI in Dify in just a few steps

Get your API key

If you do not have a Privatemode API key yet, generate one for free here.

Run the Privatemode proxy

The proxy verifies the integrity of the Privatemode service using confidential computing-based remote attestation. The proxy also encrypts all data before sending and decrypts data it receives.

Deploy Dify

Clone Dify Community Edition and run the given commands. Open http://localhost and create your admin account. All services, including Postgres, Weaviate, and Redis, run locally.

Install the OpenAI-API-Compatible plugin and configure models

In Dify, go to Settings, then Model Provider, then Marketplace, and search for OpenAI-API-Compatible to install it. This plugin handles the connection to Privatemode without any custom code.

Then add three model types in the Model Provider settings, all pointing to http://host.docker.internal:8080/v1:

LLM: a chat model
Text Embedding: an embedding model
Speech-to-Text: a speech-to-text model

Pick your model IDs from the Privatemode model overview. Set the API key to any placeholder value. Authentication is handled by the proxy.

Done!

You have successfully integrated Confidential AI into Dify. Create chatbots, agents, or knowledge bases using the configured models.

FAQ

Frequently asked questions about using Privatemode with Dify

No. Dify's built-in OpenAI-API-Compatible plugin handles the connection. You configure model names and the base URL, then select the models in Dify apps.