We implement a recursive descent into the LLMs parametric knowledge store starting from a single seed question. We use a multi-agent architecture that constrains the LLM to return facts expressed as ground Prolog terms, and facts derived as generalizations of their key concepts. Iteration over prompts engineered via summarization of the generated Prolog facts and salient next question generation enables the synthesis of a focused concept map ready to be extended via logical reasoning.
Install with:
pip install -r requirements.txt
Then, run it with
python3 test.py
Make sure to have an openai key in your environment!
Alternatively, you can use ollama - see config.py for changes in CF that redirect to a local LLM by adapting:
USE_OLLAMA = True
OLLAMA_MODEL = "gemma3:12b"
OLLAMA_BASE_URL = "http://localhost:11434/v1"