prompting_json.md

Function Calling via JSON Prompt Injection

I implement function selection (tool calling) by injecting tool definitions from exercise_input/functions_definition.json directly into the LLM system prompt. This allows my pipeline to select functions dynamically and extract target arguments from user queries.

Pipeline Workflow

1. Load Tool Schemas: I read the function definitions dynamically from the JSON configuration file using pathlib:

import json
from pathlib import Path

functions_path = Path("exercise_input/functions_definition.json")
with functions_path.open("r", encoding="utf-8") as f:
    functions = json.load(f)

2. Construct System Context: I inject the serialized tool definitions into my system prompt template along with the incoming user prompt:

Available functions:
{functions_json}

User query: 2 + 2

Select the appropriate function from the list above to answer the query. Return the result strictly as a JSON object matching the required schema.

3. Structured Model Output: Instead of outputting a plain-text answer directly, the model returns a JSON payload containing the function name (fn_name, e.g., "fn_add_numbers") and required arguments (args).
4. Parsing & Execution: My extract_json_from_response utility captures the JSON block, validates it against my SelectedFunction Pydantic model, enforces strict type coercion on parameters, and routes execution to the targeted Python function.