Capacity's AI Answer Engine® leveraged Phi to deliver better results for their customers, faster

Capacity an all-in-one Support Automation Platform, provides organizations with the ultimate Answer Engine®. They needed a way to help unify diverse datasets across tens of millions of search results and billions of interactions and make information more easily accessible and understandable for their customers. By leveraging Phi—Microsoft’s family of powerful small language models offering groundbreaking performance at low cost and low latency—Capacity provides the enterprise with an effective AI knowledge management solution that democratizes knowledge on large teams securely and in a way that maximizes value to the customer. With Phi, Capacity’s Answer Engine® improved results quality and scale, so customers save both time and money by more quickly finding the rich information they invested in to do their best work.

What was the challenge?

Enterprise employees struggle to find the data they need searching through isolated, untagged content, leading to frustration and wasted time. To address this, Capacity’s Answer Engine® retrieves information across diverse enterprise systems, repositories and sources, instantly delivering the exact answers needed to inform work and make faster decisions. At the same time, AI can only go so far to unify and enrich this data. Capacity addressed the challenge by leveraging Phi using Azure Serverless API to experiment on the effectiveness of Language Model-based tagging infrastructure. They applied prompt engineering, adherence workflows, and at-scale testing to better prepare Answers for search and create a more universal Answer Engine®.

Why did Capacity choose Phi?

Capacity chose Phi-3.5-mini for its speed, cost-effectiveness, and deployment flexibility. With Azure Models as a Service (MaaS), Capacity was able to use the Phi family models without having to provision GPUs or manage back-end operations, saving their team time, effort, and cost. They used prompt engineering and metadata tagging to optimize search results, ultimately improving development speed and query processing efficiency. Additionally, the favorable MIT Open Source licensing of the Phi family provided a strong long-term strategy for their private cloud deployment, vectorization, and query routing activities.

"From our initial experiments, what truly impressed us about the Phi was its remarkable accuracy and the ease of deployment, even before customization. Since then, we've been able to enhance both accuracy and reliability, all while maintaining the cost-effectiveness and scalability we valued from the start." 
~ Steve Frederickson, Head of Product, Answer Engine

How did they solve for it?

To achieve their goal, Capacity implemented Phi-3-mini and Phi-3.5-mini Model-as-a-Service, using both 4k and 128K variants with some prompt engineering. This allowed them to accelerate development on their AI-powered Answer Engine and help their enterprise customers deliver the right information to their end users quickly and accurately.

When presenting an Answer to their customer’s end user, Capacity wanted their AI Answer engine to instantly present the full Answer along with all the content metadata around it, so the end user could feel confident in their search results. To accomplish this, Capacity engineers split the tasks for Phi into preprocessing and real-time flows. In preprocessing, they generated metadata such as title summaries for answers, keyword tags for search, and other information to the index. This pre-work was done offline and ahead of time. Depending on the tagging task required for each Answer, they calculated the needed token size then rerouted the query to the appropriate Phi model. At query time, Phi models pre-process the query to retrieve the most relevant content.

The split tasks for Phi enabled repeatable performance, keeping the responsive query times users expect while enhancing results with new functionality and increased retrieval relevance. At the same time, the cost-efficiency of Phi was able to produce the same or better qualitative results for preprocessing with a 4.2x cost savings as compared to the competing workflow.

The considerable cost savings on the preprocessing allows Capacity to scale to ever-growing datasets. While the increased retrieval relevance fosters sustained growth and enhances user satisfaction.

After integrating Phi, Capacity observed significant improvements in both performance and customer satisfaction. The AI-powered solutions provided faster and more accurate information retrieval, which reduced time users spent searching for information.

Additionally, the seamless integration of datasets with the Phi-3.5-mini model as a service significantly empowered Capacity to address a wide range of use cases with enhanced efficacy, ultimately elevating the user experience. Steve Frederickson, Capacity's Head of Product, Answer Engine, noted,

“Integrating our datasets with the Phi-3.5-mini model was effortless. We have found new opportunity in its speed, and the enriched customer experience of GenAI enables us to resolve customer issues more effectively, delivering a superior user experience."

Capacity also shared some valuable tips for other organizations looking to implement similar AI solutions. They recommended designing the system to optimize for query performance and retrieval accuracy, including adding metadata and keyword tags to optimize search efficiency. They also emphasize the importance of choosing the right AI model based on the capability and scalability, to balance speed and cost-effectiveness.

The next step

Implementing Phi has revolutionized Capacity’s approach to knowledge management, providing their enterprise customers with efficient and accurate information retrieval solutions. Their success highlights the potential of the Phi model family to transform enterprise operations and improve user experiences.

Looking ahead, Capacity plans to explore additional state-of-the-art models such as Phi-4-multimodal and Phi-4-mini for more complex reasoning tasks like multilingual support and image understanding scenarios. They also aim to fine-tune their solutions to enhance their knowledge graph and improve interoperability among different institutional knowledge bases.

By continuously innovating and leveraging advanced AI technology, the Capacity Answer Engine® is well-positioned to remain at the forefront of knowledge management solutions, helping organizations do their best work with the complexities of information retrieval and discovery.

Learn more about the Phi family of models here:

About Phi

Learn about the latest updates

Download the models