Considering a New Logic of Knowledge Storage and Retrieval: Seeking Advice and Research Resources

Abstract: I’m exploring an innovative approach to knowledge storage and retrieval, tailored for specific use cases like novel writing. This method aims to offer logical benefits beyond traditional semantic search techniques, potentially revolutionizing information access and utilization.

  1. Applicability in Novel Writing: This concept is particularly useful in creative domains such as novel writing. Imagine reading a paragraph you’ve written and sensing it lacks engagement. This realization could serve as a ‘key’ for querying specific writing techniques from an expert system, aimed at resolving such issues. For instance:
    Key: “When a character’s motivation feels unclear or weak.”
    Value: Detailed paragraphs of writing skills tailored to address this challenge.

  2. Why This Approach?:

  • Problem and Conditions Are Clear and Easily Identified: Using an editor agent to review my article and give comments produces feedback very similar to the keys I use to label my writing skills. A sentence as a key accurately expresses the condition or problem.

  • Customization: As the key-value pair is {problem: solution}, I can specify in one condition what is the best technique the agent should use to handle it.

Problems and Challenges:
Here is my workflow:

  1. Get information about writing skills.
  2. Summarize, and validate for correctness and completeness using GPT.
  3. Generate a condition (an easily triggered condition during the AI rewiew) when I need to refer to this skill.
  4. Save {condition: paragraph} into the database.
  5. Editor agent reviews; for each suggestion, find the same problem from keys, then call the most appropriate knowledge to feed a new agent to solve it.

The challenge now is at step three. I could come up with some ideas, like listing the common conditions and problems I might meet while writing a novel and categorizing my current condition into one of those. However, this approach seems not smart and comprehensive. I prefer if GPT could directly provide a proper condition when this paragraph/article of skill could be applied. Any ideas, guys?

I think this would be a great application for GPT-4, making the problem-solution relationship clearer for the multi-agent system and controlling the correct action for each given problem.

Any existing research/results I could refer to or use? Any suggestions?