evoproc_procedures.prompts¶

Prompt builders for procedure creation and step execution.

These helpers produce LLM prompts used by the GA scaffold and your runners:

create_direct_prompt() — a simple “ask the question directly” prompt (handy for baselines and ARC-style multiple choice).
create_procedure_prompt() — asks the model to emit a global-state procedure JSON that validates your Pydantic schema.
create_execution_prompt() — runs a single step by showing inputs, the step action, and the required outputs, while reminding the model to return strict JSON conforming to a supplied schema.

All functions are pure string builders and have no network side-effects. They are safe to use in tests and with any LLM backend.

Example

>>> from evoproc_procedures.prompts import create_procedure_prompt
>>> p = create_procedure_prompt("Natalia sold clips to 48 friends in April...")
>>> isinstance(p, str)
True

Functions

`create_direct_prompt`(item)	Build a direct Q→A prompt (baseline / ARC-style).
`create_execution_prompt`(visible_inputs, ...)	Build a prompt to execute a single step with strict JSON output.
`create_procedure_prompt`(item[, example_prompt])	Ask the LLM to synthesize a global-state procedure JSON.

evoproc_procedures.prompts.create_direct_prompt(item)[source]¶

Build a direct Q→A prompt (baseline / ARC-style).

Parameters:: item (str) – The task text (e.g., question stem or problem statement).
Return type:: str
Returns:: A concise prompt string suitable for a direct answer call.

evoproc_procedures.prompts.create_procedure_prompt(item, example_prompt=None)[source]¶

Ask the LLM to synthesize a global-state procedure JSON.

The model is instructed to return exactly one JSON object that validates the Pydantic schema produced by uhj_procedures.models.Procedure.

Parameters:

item (str) – Natural-language task to decompose.
example_prompt (Optional[str]) – Optional extra hint or in-context example description (kept as a single string you can preformat upstream).

Return type:

str

Returns:

A prompt string suitable for a structured generation call.

Notes

The schema is injected verbatim via Procedure.model_json_schema().
Constraints match your GA scaffold: step 1 uses problem_text only, later steps may read any earlier variables (global state), and the final step emits final_answer as a description only.

evoproc_procedures.prompts.create_execution_prompt(visible_inputs, action, schema, expected_outputs, output_descriptions=None, *, is_final=False)[source]¶

Build a prompt to execute a single step with strict JSON output.

Parameters:

visible_inputs (Dict[str, Any]) – The subset of global state the step may read (already extracted variables).
action (str) – The step’s imperative instruction (e.g., “extract the two numbers a and b”).
schema (Dict[str, Any]) – JSON Schema that the output object must validate against (for regular steps use your step-output schema; for the last step you can pass your answer schema, e.g. GSM/ARC).
expected_outputs (Iterable[str]) – Names of keys the model must return in the JSON object.
output_descriptions (Optional[Dict[str, str]]) – Optional mapping of key → human description (shown to the model).
is_final (bool) – If True, annotate that these outputs constitute the final answer.

Return type:

str

Returns:

A prompt string that instructs the model to return only a JSON object matching schema and containing exactly expected_outputs.