The chatting and memory flow of Sentient is designed to provide users with an intelligent, personalized experience that goes beyond standard chatbot interactions. Users can converse with Sentient just like any other chatbot, but the underlying system is capable of performing various actions based on the context of the conversation, making it much more dynamic and responsive.

Orchestrators

Sentient’s core functionality is driven by orchestrators, which are responsible for interpreting and processing user messages. Upon receiving a user input, the orchestrator analyzes the message and decides which action should be taken. The possible actions include:

1. Normal Response: The system simply provides a direct response to the user’s message.
2. Memory Action: The system identifies that the message contains personal or contextual information that should be stored or updated in the user's memory (knowledge graph).
3. Agentic Tool Call: The system recognizes that the user requires a specific task or tool to be executed (e.g., generating a report, setting a reminder, etc.).

Message Classification: Personal vs General

User messages are classified as either personal or general:

• General Queries: These are typical chatbot interactions where the model does not need to access any personal information. The system processes these queries and provides a response without any reference to the user's knowledge graph.
• Personal Queries: When the system detects that a query pertains to personal or sensitive information, such as asking about the user's preferences, interests or previously mentioned facts, the model will access the user’s personal knowledge graph. It will search for relevant data within the graph and use that context to provide a more accurate and personalized response.

Internet Context Retrieval

In cases where the user’s query cannot be answered with the data available within the personal knowledge base, Sentient may need to retrieve information from the internet. This process is handled by another orchestrator, which evaluates whether the query requires external context.

If the system determines that additional information is needed, it will:

1. Generate an Anonymized Search Query: To protect the user’s privacy, any personal data is removed from the search query before it is sent out. The query is anonymized to ensure no sensitive user information is shared during the search.
2. Search and Retrieve Context: The model performs an internet search to gather relevant information and context to answer the user’s query.
3. Use Retrieved Context for Response: After retrieving the information, the model uses it to formulate a complete and accurate answer for the user.

Memory Actions

Memory Actions are designed to extract facts and key details from user messages, which are then stored in the knowledge graph. These facts might include things like the user’s preferences, key dates, or frequently requested tasks. For example, if a user mentions a new project or updates their contact details, Sentient will identify these pieces of information, extract them, and add them to the user’s knowledge base for future reference.

Agentic Actions

When a user requests a specific task or action, such as setting a reminder, creating a report, or sending an email, Sentient triggers Agentic Actions. These actions interact with external tools or APIs to execute the required task. For instance, if a user asks Sentient to set a reminder for an appointment, the system will interact with the appropriate tool to create that reminder and then confirm with the user.