KARMA

Knowledge Association and Reasoning Mapping Agent

What is KARMA?

KARMA is a large language model (LLM) internal knowledge-based fact crawler. KARMA is a dynamic system that can crawl semantically related concepts in an LLM. Facts are checked verified against existing an knowledge base, using the WikiData API. KARMA will walk through facts about a given topic, dynamically selecting new facts based on the LLMs perceived relationship to one another.

Examples

Query for Wright State University

Query for Albert Einstein

Usage

To use KARMA, just run the shell script with the topic as the first argument (do not use spaces)

# crawl LLM knowledge related to Albert Einstein
./run.sh "AlbertEinstein"
# see output image in ./visualization/graph_output.png
# bat visualization/graph_output.png

Prerequisites

Requirement	Purpose
Rust	Implementation language
Cargo	Dependency and build manager
Python	Visualization language
Pip	Visualization dependency manager
Ollama	LLM backend

# Install mistral with Ollama
ollama pull mistral

Documentation

This serves as a table of contents for much of the information related to KARMA.

Knowledge Crawling

• How does KARMA crawl knowledge?

Future Work

• Configurable Models
• Predicate Verification
• WikiData Global Verification vs Retrieved Verification
• Usability
• Ranked Predictions
• Improved Object Verification

Prompting Methodology

• Task Exposition
• Definitions
• Assurance of Relevancy
• Example
• Generation Request
• Final Prompt

Relevance

• Knowledge Verification
• Internal Knowledge Crawling

Models

• Why was Mistral Chosen for KARMA?
  • Deepseek
  • LlammaThree
  • GPT2
  • Triplex