Skip to main content

History AI: Enhancing Ongoing Conversations with Contextual Awareness to create vector query

Overview​

History AI is built to augment Language Model capabilities by applying context from previous interactions within a conversation to enhance the query to build a new vector query. Unlike Context AI, which is tailored for the initial query of a session, History AI takes charge from the second query onwards. It adeptly handles the complexity of evolving discussions, ensuring that each response is informed by past exchanges, thus providing the best vector query.

This AI is only used when the ReAct is not enabled

Note: Check the AI settings --> ReAct & LLM Agent Settings --> Enable ReAct agentic workflow = Unchecked When Enable ReAct agentic workflow = Checked, then the React Reasoning Engine creates a history which is passed to the Rag knowledge base Agent

The image shows a section of an AI configuration interface for managing 'History AI Prompts', with a detailed explanation for creating queries based on conversation history. The prompt requires crafting a query for a vector database, considering the ongoing conversation without providing extra explanation, as the system does not involve human interaction. Examples illustrate how to generate vector queries from conversation history. A system prompt guides users on processing chat history within a Retrieval-Augmented Generation (RAG) system to understand the context and create optimized database queries. There’s a list of dynamic tokens for query customization on the right, and a 'Save' button to apply changes

Implementing History AI​

The process involves configuring two main components:

  1. The Prompt: The prompt for the History LLM.
  2. The System Prompt: The System Prompt for the History LLM.

Dynamic Tokens​

Dynamic tokens act as placeholders that inject dynamic content into the prompt when it is executed. History AI supports a range of dynamic tokens that you can utilize:

  • {query}: The user's inputted question.
  • {image_description} The description from Vision AI
  • {history}: The history of the conversation, allowing the AI to build upon previous interactions.
  • {title}: The title of the webpage where the search is conducted, which may provide context for the search.
  • {origin}: The original URL of the page, which might influence the type of information sought.
  • {language}: The language of the page, ensuring that results are returned in the correct language.
  • {referrer}: The URL that directed the user to the current page, which could influence their search intent.
  • {attributes}: Additional context that can be inserted, typically added via javascript to the search control.
  • {org_name}: The name of the organization for which the AI is configured, helping tailor the AI's function to that entity.
  • {purpose}: The stated purpose of the AI bot, guiding the AI in how to prioritize and refine search results.
  • {org_url}: The domain URL of the organization, which the bot can use to prioritize content from that domain.
  • {tz=America/New_York} : Timezone creates the text of what is the time and date for the LLM to use
  • {image_upload_description} : Image descriptions of images uploaded by the site visitor into the search box processed by vision AI

Best Practices​

When writing prompts and configuring the system, here are some best practices to consider:

  1. Contextual Relevance: Always prioritize the user's intent and the context in which the query is made.
  2. Clear Token Usage: Use dynamic tokens precisely and only where they will enhance the quality of the query.
  3. Test and Iterate: Regularly test the prompts with real queries to ensure they perform as expected and iterate based on the results.
  4. Documentation: Maintain clear and comprehensive documentation so that the purpose and function of each token are easily understood by users configuring the AI.

Conclusion​

By accurately configuring the prompt and system prompt, and utilizing the dynamic tokens, History AI is capable of transforming a simple search into a complex, context-aware vector query. This results in a smarter search experience that saves time and delivers more relevant results.