The core architecture comprises three layers:
1. Website Interface: Serving as the primary user-data interface, the website requires minimal resources for operation. The majority of processing is offloaded to the vectorDB and language model instances.
2. VectorDB Layer: Housing numeric representations of organizational data segments, the vectorDB facilitates efficient storage. It performs a similarity search based on query meaning, not just string matching, and returns relevant data segments to the central UI.
3. Language Model: Accessed via API calls to providers like Anthropic, OpenAI, or private GPU instances such as Sagemaker, the language model is responsible for generating answers. It operates on the user query and context provided by the central UI.
Upon receiving a user query, the VectorDB conducts a nuanced similarity search, understanding the query context. It then transmits relevant data segments to the central UI.
Subsequently, the central UI dispatches the question and context to the Language Model. If contextually appropriate, the model crafts an answer and returns it to the UI for presentation.
If you’re intrigued by the idea of implementing a comparable system for your organization, enabling natural language interactions with company documents, we encourage you to get in touch with us.