How spcl/knowledge-graph-of-thoughts works
Project type and purpose
• This is a modular AI-assistant framework (“Knowledge Graph of Thoughts” or KGoT) plus an evaluation pipeline (GAIA) and a standalone visualization tool.
• Key purpose: orchestrate LLM-based reasoning, dynamically build a knowledge graph, invoke external tools in secure sandboxes, and evaluate on the GAIA benchmark.
File structure
• Top level:
• kgot/ – core Python package with submodules:
– controller (pipeline orchestrator)
– knowledge_graph (graph‐store abstraction over Neo4j or networkX)
– tools (tool‐invocation interfaces, isolated code execution)
– prompts, utils, config files, CLI entrypoint (__main__.py)
• GAIA/ – dataset fetchers, baselines, evaluation script (gaia.py), plotters
• docker_instances/ – container definitions for Neo4j, Python Code Tool, KGoT dependencies
• Configuration files: Docker Compose files, .env, pyproject.toml, JSON configs (config_llms.json, config_tools.json).
README insights
• Three main components in architecture overview:
a. Controller – orchestrates reasoning loop
b. Graph Store – pluggable backends (Neo4j via Docker, or in-process networkX)
c. Integrated Tools – Python code sandbox, web crawler, math solver, etc.
• CLI entrypoint kgot, plus GAIA evaluation script and visualization subproject.
Detail
• Both supported graph backends as distinct modules
• Tool manager dispatching to multiple tool‐specific containers or HTTP clients
• GAIA dataset fetch & evaluation as a side‐pipeline
• Snapshot visualization full-stack subproject as an optional consumer of outputs