Dialog API

The Dialog API uses symbolic techniques to form a control plane, accelerator, and LLM router for text, voice, and multimodal inputs. It enables the rapid "twitch" reactions needed for maintaining a realistic dialog with a situated agent, while reducing your spend on LLM resources. Think of it as a neuro-symbolic harness that adds precision, repeatability, and response curation, for important user interactions.

The engine lets you combine syntactic and semantic rules. Rules are stored as static JSON files, and can be edited using a visual authoring tool. Key features of the engine include:

• Multistep transformation using both syntactic and semantic rules
• An efficient, chart-based approach that considers end-to-end context before committing to an interpretation
• A predicate-based user store that allows for the accumulation of working knowledge and long-term memories
• Output that is well-adapted to a situated and/or embodied agent, including support for multimodal inputs and non-verbal behaviors
• A compute-efficient design that allows for ultra low costs and latencies on domain-specific responses
• A framework for harnessing slower and more expensive external APIs, most notably LLMs

You normally call the API from a server application that you manage, in order to provide storage for conversation state. This allows an agent to form long-term memories about each user. Because logs and conversation state are stored on your terms, you remain in complete control of the sensitive information they contain.