To tackle ePilot’s challenge, Firemind developed an AI-powered solution centered around a Retrieval-Augmented Generation (RAG) model. The process began with ePilot providing a set of email threads, which Firemind used to create a custom knowledge base. This knowledge base was built using Amazon Bedrock and Amazon OpenSearch Serverless, where the email threads were indexed and embedded into a vector database, allowing for efficient retrieval of context-specific information.

The solution involved several key steps:

Data Preparation & Indexing: The raw email threads provided by ePilot were first cleaned and converted into a format suitable for indexing. These were then stored in Amazon S3 and indexed using an embeddings model, specifically Titan Embeddings G1, to ensure the retrieval of relevant chunks of information from the knowledge base.

Email Summarisation: For each email thread, the latest message was extracted, and an AI model (Claude) was used to generate a topical summary. This summary served as a query to the knowledge base, retrieving up to 30 relevant chunks of text that were semantically aligned with the query.

Response Generation: The relevant information retrieved from the knowledge base, along with the summarised email content, was fed into the AI model to generate a comprehensive response. This included a summary of the entire email thread, suggested next steps, and a reply email draft that aligned with the tone and context of the conversation.

The entire process was designed to be quick, with the AI generating a response within 10 to 30 seconds. This marked a significant improvement over the manual process, which could take anywhere from 5 to 30 minutes or more. The solution was integrated seamlessly into ePilot’s CRM platform, enabling users to manage email communications more efficiently and accurately.