Configuration

Two YAML files configure experiments. The first, config.yml, specifies maps, MCTS parameters, and which prompt compositions to use. The second, config_prompts.yml, defines the reusable fragments that make up a prompt. The Python helper llm_abstraction.utils.yaml.load_config reads both files.

config.yml. The game section lists one or more square maps using YAML block scalars (|), with whitespace‑separated cells per row. The agent starts in the top‑left, and the goal G is at the bottom‑right. The mcts_variables section sets simulation limits and depths, the number of runs, the UCT exploration constant c, the discount factor gamma, and an optional debug flag. The llm section controls prompt generation: tries specifies how many valid responses to collect per map, and compositions lists the prompt compositions by id.

compositions. Each composition picks one instruction, one necessary_context, zero or more background_contexts, a representation_key (text, json, or adj from core_rust.generate_representations_py), and an output variant. llm_abstraction.llm.prompts.generate_prompts resolves these ids against config_prompts.yml and inserts the chosen representation rendered from the current map.

config_prompts.yml. This file defines fragments in four groups: instruction, necessary_context, context, and output. Each entry has an id and a val. When you add or remove fragments here, update the ids used under llm.compositions in config.yml accordingly.

Notes. Representation keys available to the prompt builder are json, text, and adj, all supplied by core_rust.generate_representations_py. If an id is missing or misspelled, the prompt builder raises a clear error so you can fix configuration mismatches quickly.