fbpx
About

About Apollo

Meet Apollo 

Consumer companies worldwide can self-deploy and fine-tune Apollo to sell their products and services with AI for the first time. Its Mixture of Models (MoM) architecture relies on specialized small language models rather than a single LLM, achieving up to a 95% reduction in errors and hallucinations compared to ChatGPT, Gemini and Co-pilot. 

To be able to represent consumer companies responsibly and safely in front of their customers, Apollo innovatively solves three critical barriers. First, the capability to effectively integrate with any number of internal systems. Second, it maintains strict control by adhering to the deploying company’s policies and regulations. Lastly, it addresses the issue of hallucinations with an innovative approach to continuous fine-tuning.

The Mixture of Models (MoM) Architecture 

Apollo runs on our groundbreaking Mixture of Models (MoM) architecture. This framework activates specialized, small language models for individual interactions, rather than relying on a single large language model for the entire conversation. In the MoM architecture, each user message prompts two models consecutively: first, a routing model that decides which workflow model (e.g. customer reviews or web search) would give the most accurate answer; then, a specialized workflow model that formulates the answer.

The Routing Model 

The routing model classifies and routes each user message to the relevant workflow model. For example, if the user asked about differences between products, the routing model will route the user message to the “Product Comparison” workflow model. There’s only one routing model in the MoM architecture, and while it can be fine-tuned, it cannot be removed or changed. 

The Workflow Models  

In the MoM architecture, each Workflow Model in the mixture functions like a specialized tool, designed to handle specific tasks. These tasks range from fetching answers about company policies from embeddings to summarizing reviews, requesting missing information, or recommending products based on user queries. The MoM architecture allows for any number of Workflow Models to be part of the mixture, and new Workflow Models can be added, changed or removed through the Control Panel. Companies can rapidly define new Workflow Models and set precise triggers and guidelines for their activation. 

Fine-tuning Apollo

Both the Routing Model and the Workflow Models can be fine-tuned within the playground. The MoM architecture enables breakthrough fine-tuning capabilities. First, each Workflow Model can be fine-tuned on specific messages as opposed to a data set of entire conversations. This enables continuous improvement in the form of fine-tuning an already fine-tuned model, and so forth. Secondly, the fine-tuning can be done right from the playground, with a human evaluator able to simply correct the mistakes of either the Routing or Workflow Models, and submitting a fine-tuning job that will apply to the relevant models in the mixture. What sets MoM further apart is the ability to fine-tune each model in the mixture independently, allowing companies to tailor each model’s performance to their exact specifications without impacting any other aspect of Apollo.

Key Innovations of Apollo & The MoM Architecture

  1. Specialized LLMs Vs. General LLM: The MoM architecture doesn’t just rely on a single large language model. Instead, it employs an ensemble of specialized, smaller models, each fine-tuned for specific interactions. This approach is akin to having a team of expert models, each skilled in a different area.
  2. Task-specific Design: As a domain-specific foundation model for e-commerce, Apollo enables companies to set guardrails and ensure the model operates within them. 
  3. Safety & Control: The MoM architecture ensures absolute command over the model’s output in every interaction and situation, safeguarding against unexpected outcomes.
  4. Unmatched Reliability: With responses sourced from designated, transparent endpoints, continuously fine-tuned through human feedback, Apollo reduces errors and hallucinations by up to 95% compared to ChatGPT, Bard or Bing. 
  5. Tool Activation: The MoM architecture offers unmatched connectivity to any number of APIs, company docs, and other relevant tools and destinations. 
  6. Efficiency: Smaller models are faster, cheaper and alleviate the need to compete over computational resources. Our ensemble of models in the Apollo Mixture of Models (MoM) architecture are trained in-house. 
  7. Continuous Fine-tuning: Each Workflow Model in the mixture is task-specific and can improve from one fine-tuning job to the next, leading to ongoing, daily gains in performance, accuracy and alignment to the needs of the deploying company. 
  8. FTHF (Fine-tuning through Human Feedback): MoM introduces an innovative approach to fine-tuning LLMs, enabling companies to directly fine-tune Apollo within the playground environment. This is achieved by simply correcting the model’s mistakes with structured feedback. 
  9. Situational Awareness: Apollo possesses advanced situational awareness, relying on an internal network of sensors to identify the state of each user-agent interaction in order to classify each customer interaction correctly.  

Developed over the past six years in collaboration with 60,000 human agents solving millions of consumer tasks, Apollo and the MoM architecture address the challenges that kept consumer companies from selling their inventory with AI. If you’re interested in being a part of our mission to transform the world from E-commerce to AI-commerce, visit our careers page.

Explore further