How One AI Assistant Works

Introduction

One AI Assistant enables users to explore data and generate charts with simple, natural language prompts. It interprets the request, uses a vector database to match it with the correct metrics, dimensions, and time selections, and delivers charts and tables tailored to the prompt. This setup gives users a streamlined way to access and visualize data without needing expertise in data models or query building. In this article, we will explore how One AI Assistant works. 

Vector Configuration

During the One AI Assistant initial setup process, site administrators use the One AI Configuration screen to select which metrics and dimensions the assistant has permission to use to generate charts and tables. These selected metric names (i.e., Headcount - EOP), dimensions names (i.e., Cost Center), and dimension nodes (i.e., Engineering as a node of Cost Center) are used to create the embeddings that will be leveraged by the assistant to match concepts in user prompts with the correct data elements in One Model. 

Embeddings are generated by sending the selected metric names, dimension names, and dimension nodes to OpenAI, who returns the embeddings to One Model. One Model then stores the embeddings in a vector database within our infrastructure. Vector databases are useful in that they enable searching by similarity. You'll learn why that's important further along in this article.

Embeddings are re-generated every time the vector configuration is saved as well as every time the data is processed on your One Model site. OpenAI retains these embeddings for logging purposes for 30 days after their creation and does not use this data to train its models. One Model has applied for Zero Data Retention (ZDR), and we are actively working with OpenAI toward approval. This setup ensures relevant data remains accessible for analysis while keeping other data secure.

Submitting a Prompt

When a user enters a prompt, such as “Show me headcount by cost center for the last 12 months,” the prompt is sent to OpenAI. Only the text of this prompt is sent to OpenAI at that time—not the embeddings. This distinction allows One AI Assistant to work efficiently by leveraging pre-stored embeddings and sending only prompt text when interacting with OpenAI. As with embeddings, prompts are retained by OpenAI for logging purposes for 30 days and are not used for model training.

 

Recognizing Key Elements & Matching to Embedded Data

When a prompt is submitted, One AI Assistant uses OpenAI’s large language model (LLM) to identify essential elements, such as metrics, dimensions, dimension nodes, and time selections within the prompt. For instance: 

  • Headcount is a metric
  • Cost Center is a dimension
  • Brisbane is a dimension node
  • Last 12 months is a time selection

These recognized elements are then sent back to One Model, where One AI Assistant queries the vector database described earlier in this article to find the closest matches in your data.Dimension nodes (e.g., “Engineering” within Cost Center) are also available in this process, allowing for specific filtering requests, such as “Show me headcount for the Engineering department only.”

Only data configured and permissioned for the assistant’s use during setup can be accessed, matched, and used in analysis. If a metric or dimension name was not embedded, it won’t be included in the assistant’s analysis, nor will it be shared with OpenAI.

Generating the Query & Displaying Results

With the matched metric(s), dimension(s), and time selection(s), One Model generates the query based on your request through its query engine. One AI Assistant then presents the data in the optimal or requested format, whether that’s a table, chart, or key value. Using the query panel, users can see exactly which metrics, dimensions, time selections, and filters were used to construct the data visualization to verify accuracy. 

Key Points to Remember

  • Initial Embedding Setup Only: Embeddings for metrics, dimensions, and nodes are sent to OpenAI only during the initial setup and site processing—not with every prompt.
  • Admin-Defined Embeddings: Administrators control which data elements are embedded during configuration, ensuring your queries align with the data you want to analyze.
  • Prompt Text Submission: Only the prompt text is sent to OpenAI each time you submit a request, not the embeddings.
  • Temporary Data Storage: Data sent to OpenAI (prompts and embeddings) is stored for 30 days until Zero Data Retention (ZDR) is implemented.

This setup allows you to interact naturally with your data while ensuring that One Model only shares prompts and selected data elements with OpenAI for accurate matching and insights.

For a more technical explanation, see the One AI Assistant Technical Overview, which covers embedding and query generation processes, data storage details, and a detailed architecture diagram.





Was this article helpful?

0 out of 0 found this helpful

Comments

0 comments

Please sign in to leave a comment.