Implementation:Liu00222 Open Prompt Injection Application query

Knowledge Sources	Open-Prompt-Injection
Domains	Prompt_Injection, LLM
Last Updated	2026-02-14 15:00 GMT

Overview

Concrete query method for processing data prompts through the defense pipeline and LLM, provided by the Application class.

Description

The Application.query method is the central execution path that processes a data prompt through four internal stages: `__prehand_detection` (PPL/response-based filtering), `__preprocess_data_prompt` (retokenization/paraphrasing), `__construct_prompt` (system instruction + defense wrapping + data), and `__process_response` (sandwich answer extraction). It delegates model inference to `self.model.query()`.

Usage

Call this method on an Application instance to process a data prompt (clean or attacked) through the full defense pipeline and obtain the model response.

Code Reference

Source Location

• Repository: Open-Prompt-Injection
• File: OpenPromptInjection/apps/Application.py
• Lines: L19-33

Signature

def query(self, data_prompt, verbose=1, idx=-1, total=-1):
    """
    Process a data prompt through the defense pipeline and query the model.

    Args:
        data_prompt (str): Input text to process.
        verbose (int): Print verbosity level (default 1).
        idx (int): Current sample index for logging (default -1).
        total (int): Total samples for logging (default -1).
    Returns:
        str: Model response after defense pipeline processing.
             Returns "[Potentially harmful. Request blocked.]" if blocked by pre-hand detection.
    """

Import

from OpenPromptInjection.apps import Application
# Typically accessed via create_app result:
# app = PI.create_app(task, model, defense)
# app.query(data_prompt)

I/O Contract

Inputs

Name	Type	Required	Description
data_prompt	str	Yes	Input text (clean or attacked)
verbose	int	No	Print verbosity (default 1)
idx	int	No	Current sample index for progress logging
total	int	No	Total sample count for progress logging

Outputs

Name	Type	Description
response	str	Model response after pipeline processing, or blocked message if rejected by defense

Usage Examples

Querying with Clean Data

import OpenPromptInjection as PI
from OpenPromptInjection.utils import open_config

target_task = PI.create_task(open_config("configs/task_configs/sst2_config.json"), 100)
model = PI.create_model(open_config("configs/model_configs/gpt_config.json"))
app = PI.create_app(target_task, model, defense='sandwich')

for i, (data_prompt, label) in enumerate(app):
    response = app.query(data_prompt, verbose=1, idx=i, total=len(app))
    print(f"Response: {response}")
    break

Related Pages

Implements Principle

• Principle:Liu00222_Open_Prompt_Injection_Application_Query_Pipeline

Uses Heuristic

• Heuristic:Liu00222_Open_Prompt_Injection_Defense_Strategy_Selection