Vertical AI models

Aforza’s Vertical AI Models are the engines behind the scenes that drive the decision making and responses received by the end users. Aforza’s strategy is to use best-in-class models (e.g. GPT-4o, Gemini Flash 2.0) and then overlay and fine tune those with vertical specific knowledge and capabilities. This approach provides an experience tailored specifically for the use cases of the Consumer Goods industry with the world class performance of the latest horizontal LLM Models.

Models & Fine Tuning

Aforza is initially using a number of different OpenAI based models to underpin Ava’s capabilities. The choice for OpenAI was based on several benchmarks including speed, throughput, and performance. Additional models will be included at a later point.

Models are then subsequently fine tuned based on common Consumer Goods use cases such as Retail Execution and Trade Promotion Management. Fine tuning provides more specificity to the models and makes them perform better for a narrower band of use cases. The nature of this fine tuning is proprietary knowledge of Aforza.

Agents & Orchestration

Aforza uses the concept of Agents to manage the requests from users. An Agent is a dedicated assistant that is focused on a specific area of functionality. For example, a Sales Agent is focused on providing capabilities around order capture. Aforza trains and instructs all of its Agents on very specific use cases that are relevant to the Consumer Goods industry. Using multiple Agents means the level of accuracy is extremely high for each individual use case.

Upon receiving a message Ava directs the request to the appropriate Agent. Throughout a conversation multiple Agents might be brought in to provide their expertise and tools to aid in the resolution of the request.

Tools & Functions

Each Agent in Ava has access to a vast array of tools that allow it to perform specific actions. For example the Sales Agent has tools to search the product catalog and to create orders. These tools allow the agents to retrieve information or perform actions they need to complete a request. The tools are what elevate Agents from being purely information providers to meaningful personal assistants that can help do your job. Agents are able to call their ‘tool’ which is then fed to the LLM Service to actually execute the authenticated function and decide what information to return to the Agent. During execution of a chat the UI displays the current tool the Agent is using.

Prompt Injection & Content Filtering

Although Ava is an internal employee tool, it is still important to add appropriate guards and measures to protect abuse of the AI system. To that end Aforza has implemented stringent guards for detecting and stopping both prompt injection attempts and inappropriate content requests. Any request that is made will be evaluated against these guard rails and either rejected or approved for answering by one of Ava’s Agents.

Vector Stores & Grounding

Ava is able to analyse images, documents, and other file types. To do this the document is first processed as a Vector Store against the agent. This file is persisted for the length of the thread. This Vector Store is then searched and analysed by the agent, allowing it to provide insights and intelligence based on the file contents.

In addition to the ad-hoc analysis of files, custom agents can also have permanent Vector Stores added against them. Having permanent files associated with the agent means they can ground their answers on your company documentation. Using this source of truth lets you have agents that can give very granular, company specific answers to your users.