A Python project to implement exploration guidance. OpenAI Codex was used in the coding of the project.
The proposed ontology for exploration for exploration guidance, that defines our operation semantics is available here.
The ontology used to represent the maze environment is available here.
The file exploration_guidance_general.py provides the classes (Model, ExplorationGuidanceServer) used to define exploration guidance.
The file maze_app.py defines the Maze HTTP server without exploration guidance (it needs to be wrapped within an ExplorationGuidanceServer and a model to do so). The model for the maze is provided in the file maze_model.py.
You need to have uv installed.
Set an OpenAI API key in API_KEY.txt to use an OpenAI model. The model to use can be configured in config.json.
The server can be run with:
uv run test_guidance.py
This opens a server at: http://localhost:5001/.
The code of MCP server implementing exploration guidance is available here. The server is running at http://localhost:8100/mcp. You can use this script to interact with it.
The code of the exploration agent is available here.
The evaluation can be run with:
uv run evaluation.py
This opens a server at: http://localhost:8765/, which can be used to start, pause, or restart an evaluation.
Before running the script, you can configure the guidance policies that will be applied in this list variable: GUIDANCE_ALLOWED_POLICIES. The possible values are: "all", "none", "outcome", "danger", "explorability", and "outcome_only"