In an era where data is invaluable and vulnerable, ensuring robust privacy and security is not just an option—it's necessary. Many AI-powered tools process and store data in ways that can expose sensitive information, but Medullar takes a fundamentally different approach. Our design philosophy centers on user privacy at every stage of data handling. This post'll explain how our system leverages Retrieval-Augmented Generation (RAG), securely processes and stores data, and how our model stands apart from competitors like Perplexity AI and ChatGPT.

Retrieval-augmented generation (RAG) is an innovative AI technique that enhances language models by retrieving external knowledge from a vector database. Instead of relying solely on pre-trained data, RAG supplements user queries with real-time, domain-specific context without compromising raw data.

How It Works in Medullar

1. User Query: A user asks a question.
2. Secure Search: Medullar's system securely searches an encrypted vector database for relevant knowledge.
3. Contextual Enrichment: The retrieved content is appended to the original query as raw text.
4. LLM Processing: The enhanced query is sent to a large language model (LLM), which generates a more accurate and context-aware response.

RAG ensures that responses are current and highly relevant without storing the original documents by integrating real-time knowledge extraction into the query process.

The graph above shows a very simplified flow of what's happening in Medullar when the user asks a question to a Space

No Permanent File Storage – Ever

At Medullar, safeguarding your data is built into our DNA. Here's how we handle your files:

• Temporary Encrypted Storage: When you upload a document, it is temporarily stored in a private Azure Blob Storage instance. Each file is encrypted using a unique encryption key per user, ensuring that even this transient data is secured.
• Knowledge Extraction & Vectorization: The content is processed to extract the knowledge it holds. Using OpenAI's embedding models, the content is vectorized and stored in a fully encrypted vector database. Notably, the vectorized knowledge is encrypted with AES-256 and protected by a unique key for each workspace.
• Immediate Deletion: Once the necessary information is extracted and securely stored, the original file is permanently deleted from our system. There is no lingering raw data—only encrypted, abstracted knowledge remains.

Our Commitment to Security

• Military-Grade Encryption: All stored knowledge is encrypted using AES-256, ensuring that not even Medullar personnel can decipher your data.
• Granular Access Controls: Strict access management ensures only authorized users can retrieve and interact with the stored knowledge.
• Data Lifecycle Management: Our processes adhere to the best data handling and retention practices, ensuring that data exists only as long as needed for immediate processing.

The graph depicts how Medullar Spaces ingests user data, independent of whether it's a file, image, video, text, or a search result from Medullar Search

While AI-powered search tools such as Perplexity AI and ChatGPT offer powerful capabilities, they handle data in ways that can compromise privacy. Here's a quick breakdown of how Medullar compares to them:

• File Storage
  • Medullar: Files are never permanently stored. They're only temporarily encrypted storage for the necessary time while Medullar processes them.
  • Perplexity AI: Files may be stored with optional sharing settings.
  • ChatGPT: Files or conversations may be stored for model training.
• Knowledge Storage
  • Medullar: Knowledge is fully encrypted and vectorized. It is a zero-knowledge model, meaning Medullar itself can't decipher user data.
  • Perplexity AI: Extracted knowledge may be stored in plaintext or with less stringent encryption.
  • ChatGPT: Offers limited control over what knowledge is stored or how it's used.
• Collaboration
  • Medullar: Secure Spaces allow sharing processed knowledge, never raw files.
  • Perplexity AI: File sharing can expose raw content to other users.
  • ChatGPT: Shared files and conversations may be accessible within the platform.
• Security Model
  • Medullar: Employs a zero-trust architecture with AES-256 encryption per workspace.
  • Perplexity AI: Provides unclear or less robust encryption details.
  • ChatGPT: Uses OpenAI's standard security and data-handling policies.

What This Means for You

• Absolute Privacy: Your original files are never stored, ensuring that even in collaborative environments, only processed, encrypted knowledge is shared.
• Zero-Knowledge Assurance: With Medullar, even our systems administrators cannot decipher your data, protecting you against external and internal threats.
• Enhanced Security for Collaboration: When you share a Space for collaboration, only the abstracted, encrypted knowledge is available—never the actual file.

As AI and data privacy regulations evolve, Medullar remains committed to staying at the forefront of secure, privacy-first technology. Our ongoing research and development aim to enhance our encryption protocols further, refine our RAG techniques, and expand our compliance certifications. With Medullar, you're not just choosing an AI-powered search tool—you're choosing a partner that prioritizes your data privacy as much as you do.

At Medullar, the future of AI lies in empowering users without compromising privacy. By ensuring that:

• Files are only temporarily held and immediately deleted.
• The extracted knowledge is stored in a fully encrypted.
• Collaboration is handled with the utmost security.

We offer a secure, enterprise-grade solution for AI-powered search and knowledge discovery.

Ready to experience an AI-powered search that truly respects your privacy?

Try Medullar today!