Prompt Library

Prompts in this project are built from a small set of reusable fragments that can be combined to encourage consistent, structured outputs from different models. Each prompt composition selects one instruction, one necessary context, zero or more background contexts, one representation of the map, and one output specification. Table 1 summarizes these five components and where they are defined. Fragments live in config_prompts.yml, and the active compositions are listed in config.yml under llm.compositions.

Table 1: Prompt composition components and sources.

Component	Purpose	Source
instruction	Call to action / role framing	config_prompts.yml
necessary_context	Minimal domain facts	config_prompts.yml
background_contexts	Additional guidance or constraints	config_prompts.yml
representation_key	Map encoding (text/json/adj)	core_rust.generate_representations_py
output	Output contract (e.g., JSON list of clusters)	config_prompts.yml

Representation matters. The Rust core can produce a textual grid, a JSON encoding, and an adjacency representation of the map (core_rust.generate_representations_py). In our experiments (see Thesis), JSON map representations paired with explicit JSON output instructions reduce parsing errors and improve abstraction quality relative to text‑only prompts.

Example. The snippet below shows a typical JSON‑oriented composition: a role‑style instruction, a necessary domain statement, two background contexts, a JSON map representation, and an output restriction that asks for list[list[int]] only.

instruction: "role1"
necessary_context: "necessary-domain2"
background_contexts: ["domain2", "test3"]
representation_key: "json"
output: "out2"

Under the hood, llm_abstraction.llm.prompts.generate_prompts resolves the ids to their text, inserts the JSON string returned by core_rust.generate_representations_py, and joins the pieces into a single prompt. See Repo → Configuration for the fragment schema and how to add your own.