tests/candidate.py

import json
import os
import random
import string
import time
from collections import defaultdict
from typing import Dict, Optional, Tuple

from openai import OpenAI
from api.llm import LLMManager
from utils.config import Config
from resources.data import fixed_messages, topic_lists
from resources.prompts import prompts
from tests.testing_prompts import candidate_prompt


def complete_interview(
    interview_type: str,
    exp_name: str,
    llm_config: Optional[Config] = None,
    requirements: str = "",
    difficulty: str = "",
    topic: str = "",
    model: str = "gpt-3.5-turbo",
    pause: int = 0,
    mode: str = "normal",
    max_messages: Optional[int] = None,
) -> Tuple[str, Dict]:
    """
    Complete an interview and record the results with additional strange use cases.

    :param interview_type: Type of interview to complete.
    :param exp_name: Experiment name for file saving.
    :param llm_config: Optional LLM configuration.
    :param requirements: Additional requirements for the interview.
    :param difficulty: Difficulty level for the interview.
    :param topic: Topic for the interview.
    :param model: Model to use for the candidate.
    :param pause: Pause duration between requests to prevent rate limits.
    :param mode: Mode of operation ("normal", "empty", "gibberish", "repeat").
    :param max_messages: Maximum number of messages in the conversation.
    :return: Tuple containing the file path and interview data.
    """
    client = OpenAI(base_url="https://api.openai.com/v1")
    config = Config()
    if llm_config:
        config.llm = llm_config
    llm = LLMManager(config, prompts)
    llm_name = config.llm.name
    print(f"Starting evaluation interviewer LLM: {llm_name}, candidate LLM: {model}, interview type: {interview_type}")
    # Select a random topic or difficulty if not provided
    topic = topic or random.choice(topic_lists[interview_type])
    difficulty = difficulty or random.choice(["easy", "medium", "hard"])

    for problem_statement_text in llm.get_problem(requirements, difficulty, topic, interview_type):
        pass

    interview_data = defaultdict(
        lambda: None,
        {
            "interviewer_llm": llm_name,
            "candidate_llm": model,
            "inputs": {
                "interview_type": interview_type,
                "difficulty": difficulty,
                "topic": topic,
                "requirements": requirements,
            },
            "problem_statement": problem_statement_text,
            "transcript": [],
            "feedback": None,
            "average_response_time_seconds": 0,
        },
    )
    # Initialize interviewer and candidate messages
    messages_interviewer = llm.init_bot(problem_statement_text, interview_type)
    chat_display = [[None, fixed_messages["start"]]]

    messages_candidate = [
        {"role": "system", "content": candidate_prompt},
        {"role": "user", "content": f"Your problem: {problem_statement_text}"},
        {"role": "user", "content": chat_display[-1][1]},
    ]

    response_times = []
    previous_code = ""

    if max_messages is None:
        max_messages = 30 if mode == "normal" else 5

    for _ in range(max_messages):
        if mode == "empty":
            response_content = ""
        elif mode == "gibberish":
            response_content = "".join(random.choices(string.ascii_letters + string.digits, k=50))
        elif mode == "repeat":
            response_content = chat_display[-1][1]
        else:
            response = client.chat.completions.create(
                model=model, messages=messages_candidate, temperature=1, response_format={"type": "json_object"}, stream=False
            )
            try:
                response_json = json.loads(response.choices[0].message.content)
                response_content = response_json.get("message", "")
            except:
                continue

        candidate_message = response_content

        if not candidate_message and mode != "empty":
            print("No message in response")
            continue

        messages_candidate.append({"role": "assistant", "content": candidate_message})

        interview_data["transcript"].append(f"CANDIDATE MESSAGE: {candidate_message}")

        chat_display.append([candidate_message, None])

        send_time = time.time()
        for messages_interviewer, chat_display, previous_code in llm.send_request(
            candidate_message, previous_code, messages_interviewer, chat_display
        ):
            pass

        response_times.append(time.time() - send_time)

        messages_candidate.append({"role": "user", "content": chat_display[-1][1]})

        message_split = messages_interviewer[-1]["content"].split("#NOTES#")
        interview_data["transcript"].append(f"INTERVIEWER MESSAGE: {message_split[0]}")

        if len(message_split) > 1:
            interview_data["transcript"].append(f"INTERVIEWER HIDDEN NOTE: {message_split[1]}")

        time.sleep(pause)  # to prevent exceeding rate limits

    for fb in llm.end_interview(problem_statement_text, messages_interviewer, interview_type):
        interview_data["feedback"] = fb

    interview_data["average_response_time_seconds"] = round(sum(response_times) / len(response_times), 2) if response_times else 0

    current_time = time.strftime("%Y%m%d-%H%M%S")
    random_suffix = "".join(random.choices(string.ascii_letters + string.digits, k=10))
    file_path = os.path.join("records", exp_name, f"{current_time}-{random_suffix}.json")

    os.makedirs(os.path.dirname(file_path), exist_ok=True)

    with open(file_path, "w") as file:
        json.dump(interview_data, file, indent=4)

    return file_path, interview_data