Versa Chat and API

Launched in August 2023, the Versa platform is approved for use with all UCSF data while offering a safe and protected environment for faculty, staff, and learners to leverage generative AI technologies. Versa keeps the data inside UCSF and complies with HIPAA requirements. The platform consists of two products – Versa Chat and Versa API.

Versa Chat is a web interface that allows users to engage in interactive conversations with large language models such as Azure OpenAI’s GPT-4o and GPT-4. Upon completion of this required online training course, the tool is free by request to anyone at UCSF. VPN is required to access Versa for remote users. See our wiki for model details.

Users can input prompts and engage with the GPT model within their browser, including incorporating protected health information (P4) and de-identified clinical data (P3).

Complete the custom training on Versa (required for access), request access, or get support. 

Versa API extends Versa Chat's capabilities by providing programmers and researchers with tools to automate tasks and process large datasets. UCSF IT partially subsidizes the use of this API, and detailed pricing information can be found at Versa pricing.

With software endpoints designed to facilitate model training and inference in a secure environment, Versa API programmatically calls the Azure OpenAI GPT models (e.g., using Python or R in a Jupyter notebook), enabling incorporation of generative AI capabilities into your data pipeline.

Versa API is approved for use with UCSF’s protected institutional and clinical data (P2, de-identified P3, and identified P4). See our wiki for the API endpoint details features supported, such as Chat Completion and Embeddings.

The Versa API gateway can be accessed from a RAE (Research Analysis Environment) virtual workstation with high level of support from the UCSF ARS (Academic Research Services) team or from other UCSF secure P4-approved environments with limited support (recommended for highly technical users).

Complete the custom training on Versa (required for access), request access, or get support. 

Versa products are cleared for use with UCSF protected data, including P2, P3, and P4 data. This is inclusive of being HIPAA compliant for UCSF data as well. Versa is not approved for PHI from ZSFG, SFDPH, or the VA, and sensitive data from these entities, including electronic health records, cannot be entered into Versa without IRB approval. 

The AI Tiger Team built Versa to be as minimally retentive of data as possible. Users prompts and responses to Versa Chat and API are transmitted securely to the model vendors (e.g., Microsoft Azure) but are not retained beyond that user-session (aka a single conversation in Chat or after the response is returned in the API). This intentional lack of data retention is intentional and intended to help Versa users be excellent stewards of UCSF data. 

Prior to making any model available in Versa, the AI Tiger Team has taken steps with the model vendors to set up every model offered (e.g., GPT-4) in Versa so that no data is retained by the model vendor beyond that session. This means the model vendors are explicitly not allowed to use any Versa data in future model training, refinement, fine-tuning, or any similar activity. UCSF data is thus kept secure in Versa and not saved by any third party. 

The UCSF IT roadmap for generative-AI services continues to evolve rapidly, along with this once in a generation speed of technological development. The AI Tiger Team, in coordination with the AI Health Governance group, Academic Research Services, Office of the CRIO, and FAS Executive Leadership Group is managing the Generative-AI roadmap.

Highlights of the Gen-AI roadmap include:

• UCSF-wide access to Versa announced April 4, 2024.
• Continued expansion of the variety of Large Language Models (LLMs) from existing vendors and from a broader array of vendors. Specifically, access through Versa to LLMs from AWS, Meta/Facebook, Anthropic, and Google are all in progress. The priority is to give users a range of choice between models based on model size, speed, and cost effectiveness.
• Appropriate usage guidelines are being developed within each UCSF mission area as more is learned about the risks, costs, and ethical and appropriate use guidelines for this powerful technology.
• Versa chat, the Health Insurance Portability and Accountability Act (HIPAA)-compliant web application, is positioned to fill faculty and staff requests for the public version of ChatGPT across the enterprise in a cost-effective and secure for business use manner.
• Versa API gateway, a suite of AI/LLM models to expand functionality and data sources such as other publicly available LLMs, UCSF’s Enterprise Data Warehouse (EDW), UCSF Intranet data, and research data assets available on Information Commons. Access to these data will follow the existing approval processes to ensure compliance with UCSF policies and procedures related to data access for quality improvement and research.

Completed roadmap highlights

• [June, 2024] Free access to GPT-4o model launched in Versa Chat and API
• [June, 2024] PTUs (prioritized bandwidth) added to GPT-4 model in Versa Chat, more than doubling the speed of Versa Chat
• [February, 2024] Retrieval Augmented Generation (RAG) access to UCSF data sources into prototype for interested data source owners
• [January, 2024] Free access to the OpenAI premium models, GPT-4-Turbo, in Versa Chat
• [October, 2023] Free access to the OpenAI premium models, GPT-4, in Versa Chat
• [August, 2023] Initial user group access to Versa Chat with access to Azure cloud service for GPT-3,and GPT-3.5-Turbo
• [August, 2023] Initial user group access to Versa API with access to Azure cloud service for GPT-3, GPT-3.5-Turbo, GPT-4, and associated models