Overview
SafeBase’s AI pipeline contains many different steps to generate answers to security questions. This pipeline handles parsing questions out of questionnaires, retrieving relevant content from your Trust Library, generating answers, and post-processing those answers to ensure that they are accurate and aligned with the answer format that the questionnaire expects. Today, AIQA is built on top of Google’s Gemini large language models (LLM), which were specifically chosen for their availability, scalability, ease of integration into our current tech stack, and the state-of-the-art performance provided over other AI tools.
At SafeBase, we recognize the critical importance of data security, privacy, and transparency. AIQA offers enterprise-grade security and privacy, coupled with unmatched accuracy.
We use a retrieval augmented generation (RAG)-based pipeline that leverages well-known foundational models.
The underlying LLM does not store data that is sent to it via our prompts, nor are they trained on any data sent to them.
AIQA makes use of vector databases that are segregated per customer and stored within our VPC
SafeBase does not use customer data for training purposes, and our foundational model providers do not train their models on your data either.
How SafeBase handles Customer data
Data Storage and Processing
Customer data is stored in a database in our VPC
AI models cannot query customer data directly. Relevant snippets needed to answer a given question are passed to the model in context only.
Models & Training
We do not use customer data for training models
We only use approved Enterprise-grade foundational models from
providers such as OpenAI, Anthropic, and GoogleCurrently, the main model powering our AI is Gemini (Google). Our chosen
model may be subject to change based on product quality & our internal SafeBase assessment.
We will never use a model that uses customer data for training.
Cloud Infrastructure
Our cloud infrastructure is powered by Google (Vertex Agent Builder)
We utilize Vellum, a 3rd-party SaaS tool, to build advanced LLM workflows. Vellum only uses snippets of information to run LLM chains, it does not use customer raw data directly.
We do not store any data in Vellum.
Opting in & out of AI
AI is only used when requesting generative answers to a question (e.g., answering a question using AIQA). In the future, SafeBase may implement other AI-powered features, and customers will always have the ability to opt out of generative AI use.
Access to AI features is gated with RBAC to control which users may utilize them.
Compliance and Security
SafeBase uses a layered security approach to protect the application and customer data. Details can be found at https://trust.safebase.io/.
AI Subprocessors
Google - data storage & foundational models
Vellum - LLM pipeline
The most up-to-date list will always be available on SafeBase's Trust Center
Questions or Concerns
For any inquiries, please reach out to security@safebase.io. We are dedicated to
handling your data with the highest standards of security and privacy.