import os import re import json import gradio as gr from openai import OpenAI # Initialize the OpenAI client with the API key from environment variables. client = OpenAI(api_key=os.environ["OPENAI_API_KEY"]) # In-memory storage to track submitted emails (not persistent; resets on app restart). submitted_emails = set() def get_evaluation_questions(): """ Loads evaluation questions and expected answers from environment variables. Expected environment variable names are: - TEST_QUESTION_1: a JSON array of user query strings. - TEST_EXPECTED: a JSON array of JSON-like strings representing the expected outputs. Both lists must be of equal length. """ questions_str = os.environ.get("TEST_QUESTION_1") expected_str = os.environ.get("TEST_EXPECTED_1") if not questions_str or not expected_str: return [] try: questions_list = json.loads(questions_str) expected_list = json.loads(expected_str) except Exception as e: print(f"Error parsing evaluation questions: {str(e)}") return [] if len(questions_list) != len(expected_list): print("Mismatch in length: questions list and expected answers list must have the same length.") return [] return [{"question": q, "expected": e} for q, e in zip(questions_list, expected_list)] # Load the evaluation questions once at startup. EVALUATION_QUESTIONS = get_evaluation_questions() def sanitize_input(text): """ Sanitizes input to prevent harmful content and limits its length. """ # Allow alphanumerics and some punctuation, then truncate to 500 characters. clean_text = re.sub(r"[^a-zA-Z0-9\s.,!?@:\-]", "", text) return clean_text.strip()[:500] def validate_email(email): """ Validates that the provided email is in a valid format. Returns True if valid, False otherwise. """ email_regex = r"^[a-zA-Z0-9_.+-]+@[a-zA-Z0-9-]+\.[a-zA-Z0-9-.]+$" return re.match(email_regex, email) is not None def submit_prompt(email, name, system_prompt): """ Handles user submission: - Validates email format. - Checks if the email has already been used for submission. - Evaluates the system prompt against predefined test questions. - Prevents multiple submissions from the same email. Returns the evaluation results or an error message if the submission is invalid. """ # Validate email format. if not validate_email(email): return "Invalid email address. Please enter a valid email." # Check if this email has already been used for submission. if email in submitted_emails: return f"Submission already received for {email}. You can only submit once." # Sanitize inputs. email = sanitize_input(email) name = sanitize_input(name) system_prompt = sanitize_input(system_prompt) score = 0 responses = [] for item in EVALUATION_QUESTIONS: question = item["question"] expected = item["expected"] try: # Use the new client-based API for chat completions. response = client.chat.completions.create( model="gpt-4o-mini", # Ensure this identifier matches the deployed model. messages=[ {"role": "system", "content": system_prompt}, {"role": "user", "content": question} ] ) # Extract the answer from the response object. answer = response.choices[0].message.content.strip() except Exception as e: answer = f"Error during OpenAI API call: {str(e)}" # Simple evaluation: check if the expected output is a substring of the answer (case-insensitive). if expected.lower() in answer.lower(): score += 1 verdict = "Correct" else: verdict = "Incorrect" responses.append( f"Question: {question}\n" f"Answer: {answer}\n" f"Expected: {expected}\n" f"Result: {verdict}\n" ) result_details = "\n".join(responses) # Record this email as having submitted their prompt. submitted_emails.add(email) return ( f"Thank you for your submission, {name}!\n\n" f"Your evaluation score is {score} out of {len(EVALUATION_QUESTIONS)}.\n\nDetails:\n{result_details}" ) def build_interface(): """ Constructs the Gradio interface with a submission button and single-submission mechanism. """ with gr.Blocks() as demo: gr.Markdown("# GPT-4o Mini System Prompt Submission") # General description gr.Markdown("""Classification Task: Document and Clause Level Identification Participants must create a system prompt for a language model that classifies user queries about legal documents into two specific categories: 1. **Document Level**: Determines whether the query refers to a single document or multiple documents. 2. **Clause Level**: Identifies whether the query is focused on: - A single clause, - Multiple clauses, or - General information not constrained to any specific clause. The model must return a valid JSON object with the following structure: ``` { "document_level": "single/multiple", "clause_level": "single/multiple/general" } ``` The goal is to ensure that the model's output is concise, structured, and accurate. This task is designed to evaluate the robustness of the system prompt in handling classification tasks with short, precise outputs. """) # Example Inputs and Outputs in an Accordion with gr.Accordion("Example Inputs and Expected Outputs", open=False): gr.Markdown(""" 1. **User Message Example 1:** - *"Please provide the contract for the lease agreement."* - **Expected Output:** ``` {"document_level": "single", "clause_level": "general"} ``` 2. **User Message Example 2:** - *"I need all clauses related to termination in the employment contract."* - **Expected Output:** ``` {"document_level": "single", "clause_level": "multiple"} ``` 3. **User Message Example 3:** - *"Can you send me the financial reports and the partnership agreement?"* - **Expected Output:** ``` {"document_level": "multiple", "clause_level": "general"} ``` 4. **User Message Example 4:** - *"What are the key clauses in the NDA?"* - **Expected Output:** ``` {"document_level": "single", "clause_level": "multiple"} ``` 5. **User Message Example 5:** - *"Tell me about the company’s financials."* - **Expected Output:** ``` {"document_level": "single", "clause_level": "general"} ``` 6. **User Message Example 6:** - *"Provide all contracts and their confidentiality clauses."* - **Expected Output:** ``` {"document_level": "multiple", "clause_level": "multiple"} ``` 7. **User Message Example 7:** - *"Extract the arbitration clause from this service agreement."* - **Expected Output:** ``` {"document_level": "single", "clause_level": "single"} ``` """) # Challenge instructions in another Accordion with gr.Accordion("Challenge Instructions", open=False): gr.Markdown(""" - Design a system prompt that ensures the AI generates outputs like those above when given similar user messages. The system prompt should: 1. Specify formatting requirements (e.g., *"Output must be a valid JSON object"*). Note that we are not using constrained decoding or any sort of JSON mode; if not correctly prompted, the LLM will output plain text. 2. Emphasize strict adherence to classification definitions: - *Single Document:* Refers to one document. - *Multiple Documents:* Refers to more than one document. - *Single Clause:* Refers to one specific clause. - *Multiple Clauses:* Refers to more than one specific clause. - *General Information:* Refers to general content not tied to specific clauses. You can only submit once, so test your system prompt thoroughly before submission! """) gr.Markdown("""Classification Task: Document and Clause Level Identification Challenge Description Participants must create a system prompt for a language model that classifies user queries about legal documents into two specific categories:" 1. Document Level: Determines whether the query refers to a single document or multiple documents. 2. Clause Level: Identifies whether the query is focused on: - A single clause, - Multiple clauses, or - General information not constrained to any specific clause. The model must return a valid JSON object with the following structure: ```json {"document_level": "single/multiple","clause_level": "single/multiple/general"} ``` The goal is to ensure that the model's output is concise, structured, and accurate. This task is designed to evaluate the robustness of the system prompt in handling classification tasks with short, precise outputs.
Click to see example inputs and outputs. **Example Inputs and Expected Outputs** 1. **User Message Example 1:** - *"Please provide the contract for the lease agreement."* - **Expected Output:** ```json {"document_level": "single", "clause_level": "general"} ``` 2. **User Message Example 2:** - *"I need all clauses related to termination in the employment contract."* - **Expected Output:** ```json {"document_level": "single", "clause_level": "multiple"} ``` 3. **User Message Example 3:** - *"Can you send me the financial reports and the partnership agreement?"* - **Expected Output:** ```json {"document_level": "multiple", "clause_level": "general"} ``` 4. **User Message Example 4:** - *"What are the key clauses in the NDA?"* - **Expected Output:** ```json {"document_level": "single", "clause_level": "multiple"} ``` 5. **User Message Example 5:** - *"Tell me about the company’s financials."* - **Expected Output:** ```json {"document_level": "single", "clause_level": "general"} ``` 6. **User Message Example 6:** - *"Provide all contracts and their confidentiality clauses."* - **Expected Output:** ```json {"document_level": "multiple", "clause_level": "multiple"} ``` 7. **User Message Example 7:** - *"Extract the arbitration clause from this service agreement."* - **Expected Output:** ```json {"document_level": "single", "clause_level": "single"} ```
**Challenge Instructions** - Design a system prompt that ensures the AI generates outputs like those above when given similar user messages. - The system prompt should: 1. Specify formatting requirements (e.g., "Output must be a valid JSON object"), not that we are not using constrained decoding or any sort of JSON mode, if not correctly prompted the llm will output plain text. 2. Emphasize strict adherence to classification definitions: - *Single Document:* Refers to one document. - *Multiple Documents:* Refers to more than one document. - *Single Clause:* Refers to one specific clause. - *Multiple Clauses:* Refers to more than one specific clause. - *General Information:* Refers to general content not tied to specific clauses. """) gr.Markdown( "Please enter your details and submit your system prompt below. " "You can only submit once, I suggest trying to test and build out the system prompt using the same LM being used here elsewhere before submitting." ) email_input = gr.Textbox(label="Email", placeholder="your.email@example.com") name_input = gr.Textbox(label="Name", placeholder="Your name") system_prompt_input = gr.Textbox( label="System Prompt", placeholder="Enter your system prompt here...", lines=6, ) submit_button = gr.Button("Submit") output_text = gr.Textbox(label="Results", lines=15) submit_button.click( fn=submit_prompt, inputs=[email_input, name_input, system_prompt_input], outputs=output_text, ) return demo if __name__ == "__main__": interface = build_interface() # Launch the app on 0.0.0.0 so it is accessible externally (e.g., in a container). interface.launch(server_name="0.0.0.0", server_port=7860)