Abstract: A conversational artificial intelligence system includes a plurality of large language models configured to perform different functions. A routing system routes requests to a large language model best suited for a task associated with the request. A conversation manager coordinates a flow of a conversation between the system and a user using the plurality of large language models. A supervisory module operates in parallel with the conversation to audit responses generated during the conversation and provides feedback that contributes to construction of downstream responses to the user. One or more safety modules prevent the system from performing actions that are outside of defined constraints. The routing system may route requests based on latency requirements, domain, or classification tasks, and may fail over to a different large language model when a first large language model degrades.

Abstract: A conversational artificial intelligence system includes at least one large language model configured to execute a conversation with a user and a memory accessible across multiple conversations with the user. The system extracts attributes of the user, including at least preferences, from the conversation and stores the attributes in the memory. In a subsequent conversation, the system surfaces at least a portion of the attributes from the memory based on a current context of the subsequent conversation, and the surfaced portion personalizes a response to the user. The memory may store the attributes as a knowledge graph and may be stored in a tokenized format in a key-value cache to reduce latency. The memory may be accessible to a plurality of artificial intelligence agents serving different roles. Computer-implemented methods and non-transitory computer-readable media for operating the system are also disclosed.

Abstract: A healthcare conversational system for cardiovascular care includes a conversation interface configured to execute multiple turns of human-like conversation with a patient, emulating a human healthcare professional. A control logic generates control signals based on evaluating upstream conversation to construct downstream conversation. A second opinion module, operating in parallel with the conversation, analyzes information provided by the patient and provides feedback that causes the conversation interface to clarify or revise a response. The system may include a retro-improvement logic that detects when a previously expressed machine response requires improvement, selects from a plurality of specialist engines one that is specific to the response pair, and causes expression of an improved response.

Abstract: A conversational artificial intelligence (AI) system is configured to engage in a multi-turn conversation with a user. The multi-turn conversation is substantially focused on a target topic. A conversation analyzer analyzes the multi-turn conversation to detect and store at least some turns in the multi-turn conversation that deviate from the target topic and instead characterize life attributes of the user. A knowledge graph constructor builds a knowledge graph for the user based on at least some turns in the multi-turn conversation that characterize the life attributes of the user. The knowledge graph translates the life attributes of the user into the user's life biography, including life chronology, life preferences, life milestones, life events, or any combination thereof. A knowledge graph applicator uses parts of the knowledge graph in a subsequent multi-turn conversation with the user by contextually interspersing portions of the user's life biography in the subsequent multi-turn conversation.

Abstract: In some aspects, an artificial intelligence that is configured to complete a checklist during a conversation initiates the conversation with a human. The artificial intelligence receives a first human response from the human, extracts data from the first human response, and stores the data in a conversation summary. The artificial intelligence provides a first artificial intelligence response to the human. The artificial intelligence receives a second human response from the human, accesses the conversation summary, and provides a second artificial intelligence response to the human based at least in part on a portion of the data in the conversation summary. The artificial intelligence, based at least in part on determining that the checklist has been completed, ends the conversation with the human.

Abstract: In some aspects, an artificial intelligence that is configured to complete a checklist during a conversation initiates the conversation with a human. The artificial intelligence receives a first human response from the human, extracts data from the first human response, and stores the data in a conversation summary. The artificial intelligence provides a first artificial intelligence response to the human. The artificial intelligence receives a second human response from the human, accesses the conversation summary, and provides a second artificial intelligence response to the human based at least in part on a portion of the data in the conversation summary. The artificial intelligence, based at least in part on determining that the checklist has been completed, ends the conversation with the human.

Abstract: In some aspects, an artificial intelligence that is configured to complete a checklist during a conversation initiates the conversation with a human. The artificial intelligence receives a portion of a first human response from the human and generates multiple possible responses to the portion of the received first human response. The artificial intelligence selects, based at least in part on receiving additional portions of the first human response from the human, a first artificial intelligence response from the predicted multiple possible responses and provides the first artificial intelligence response to the human. The artificial intelligence, based at least in part on determining that the checklist has been completed, ends the conversation with the human.

Abstract: In some aspects, a multi-turn conversational system includes: an artificial intelligence to provide a conversation interface configured to execute multiple turns of human-like conversation with a user and a control logic, in communication with the conversation interface, and configured to generate one or more control signals based on evaluating multiple turns of upstream human-like conversation between the conversation interface and the user. The control signals contribute in part to construction of multiple turns of downstream human-like conversation between the conversation interface and the user.

Abstract: In some aspects, a multi-turn conversational system includes: an artificial intelligence to provide a conversation interface configured to execute multiple turns of human-like conversation with a user and a control logic, in communication with the conversation interface, and configured to generate one or more control signals based on evaluating multiple turns of upstream human-like conversation between the conversation interface and the user. The control signals contribute in part to construction of multiple turns of downstream human-like conversation between the conversation interface and the user.