Created models config

app.py

@@ -1,7 +1,7 @@
 import gradio as gr
 
-from llm import end_interview, get_problem, read_last_message, send_request, transcribe_audio
-from options import fixed_messages, models, topics_list
+from llm import end_interview, get_problem, read_last_message, send_request, speech_to_text
+from options import fixed_messages, topics_list
 
 default_audio_params = {
     "label": "Record answer",
@@ -41,8 +41,8 @@ def hide_solution():
 
 with gr.Blocks() as demo:
     gr.Markdown("Your coding interview practice AI assistant!")
-    # TODO: add other types of interviews (e.g. system design, ML design, behavioral, etc.)
-
+    with gr.Tab("Instruction") as instruction_tab:
+        pass
     with gr.Tab("Coding") as coding_tab:
         chat_history = gr.State([])
         previous_code = gr.State("")
@@ -66,16 +66,10 @@ with gr.Blocks() as demo:
                         topic_select = gr.Dropdown(
                             label="Select topic", choices=topics_list, value="Arrays", container=False, allow_custom_value=True
                         )
-
-                    gr.Markdown("##### Assistant settings")
-                    with gr.Row():
-                        gr.Markdown("Select LLM model to use")
-                        model_select = gr.Dropdown(label="Select model", choices=models, value="gpt-3.5-turbo", container=False)
                 with gr.Column(scale=2):
                     requirements = gr.Textbox(label="Requirements", placeholder="Specify additional requirements", lines=5)
                     start_btn = gr.Button("Generate a problem")
 
-            # TODO: select LLM model
         with gr.Accordion("Problem statement", open=True) as problem_acc:
             description = gr.Markdown()
         with gr.Accordion("Solution", open=False) as solution_acc:
@@ -94,21 +88,18 @@ with gr.Blocks() as demo:
         with gr.Accordion("Feedback", open=True) as feedback_acc:
             feedback = gr.Markdown()
 
-    with gr.Tab("Instruction") as instruction_tab:
-        pass
-
     coding_tab.select(fn=add_interviewer_message(fixed_messages["intro"]), inputs=[chat], outputs=[chat])
 
     start_btn.click(fn=add_interviewer_message(fixed_messages["start"]), inputs=[chat], outputs=[chat]).then(
         fn=get_problem,
-        inputs=[requirements, difficulty_select, topic_select, model_select],
+        inputs=[requirements, difficulty_select, topic_select],
         outputs=[description, chat_history],
         scroll_to_output=True,
     ).then(fn=hide_settings, inputs=None, outputs=[init_acc, start_btn, solution_acc, end_btn, audio_input])
 
     message.submit(
         fn=send_request,
-        inputs=[code, previous_code, message, chat_history, chat, model_select],
+        inputs=[code, previous_code, message, chat_history, chat],
         outputs=[chat_history, chat, message, previous_code],
     )
 
@@ -117,14 +108,14 @@ with gr.Blocks() as demo:
         inputs=[chat],
         outputs=[chat],
     ).then(
-        fn=end_interview, inputs=[description, chat_history, model_select], outputs=feedback
+        fn=end_interview, inputs=[description, chat_history], outputs=feedback
     ).then(fn=hide_solution, inputs=None, outputs=[solution_acc, end_btn, problem_acc, audio_input])
 
-    audio_input.stop_recording(fn=transcribe_audio, inputs=[audio_input], outputs=[message]).then(
+    audio_input.stop_recording(fn=speech_to_text, inputs=[audio_input], outputs=[message]).then(
         fn=lambda: None, inputs=None, outputs=[audio_input]
     ).then(
         fn=send_request,
-        inputs=[code, previous_code, message, chat_history, chat, model_select],
+        inputs=[code, previous_code, message, chat_history, chat],
         outputs=[chat_history, chat, message, previous_code],
     )
 

config.py

@@ -0,0 +1,18 @@
+LLM_URL = f"https://api.openai.com/v1"
+LLM_KEY_TYPE = "OPENAI_API_KEY"  # there should be an environment variable with this name
+LLM_NAME = "gpt-3.5-turbo"
+# "gpt-3.5-turbo" - ~3 seconds delay with decent quality
+# "gpt-4-turbo","gpt-4", etc. 10+ seconds delay but higher quality
+
+STT_URL = f"https://api.openai.com/v1"
+STT_KEY_TYPE = "OPENAI_API_KEY"  # there should be an environment variable with this name
+STT_NAME = "whisper-1"
+# "whisper-1" - the only OpenAI STT model available
+
+
+TTS_URL = f"https://api.openai.com/v1"
+TTS_KEY_TYPE = "OPENAI_API_KEY"  # there should be an environment variable with this name
+TTS_NAME = "tts-1"
+# Recommended options
+# "tts-1" - good quality and close to real-time response. Just use this one
+# "tts-1-hd" - slightly better quality with slightly longer response time

llm.py

@@ -1,14 +1,19 @@
 import json
+import os
 
 from dotenv import load_dotenv
 from openai import OpenAI
 
 from audio import numpy_audio_to_bytes
+from config import LLM_KEY_TYPE, LLM_NAME, LLM_URL, STT_KEY_TYPE, STT_NAME, STT_URL, TTS_KEY_TYPE, TTS_NAME, TTS_URL
 from prompts import coding_interviewer_prompt, grading_feedback_prompt, problem_generation_prompt
 
 load_dotenv()
-# TODO: don't use my key
-client = OpenAI()
+
+client_LLM = OpenAI(base_url=LLM_URL, api_key=os.getenv(LLM_KEY_TYPE))
+print(client_LLM.base_url)
+client_STT = OpenAI(base_url=STT_URL, api_key=os.getenv(STT_KEY_TYPE))
+client_TTS = OpenAI(base_url=TTS_URL, api_key=os.getenv(TTS_KEY_TYPE))
 
 
 def init_bot(problem=""):
@@ -19,7 +24,7 @@ def init_bot(problem=""):
     return chat_history
 
 
-def get_problem(requirements, difficulty, topic, model, client=client):
+def get_problem(requirements, difficulty, topic, client=client_LLM):
     full_prompt = (
         f"Create a {difficulty} {topic} coding problem. "
         f"Additional requirements: {requirements}. "
@@ -27,7 +32,7 @@ def get_problem(requirements, difficulty, topic, model, client=client):
         "Ensure the problem varies each time to provide a wide range of challenges."
     )
     response = client.chat.completions.create(
-        model=model,
+        model=LLM_NAME,
         messages=[
             {"role": "system", "content": problem_generation_prompt},
             {"role": "user", "content": full_prompt},
@@ -39,7 +44,7 @@ def get_problem(requirements, difficulty, topic, model, client=client):
     return question, chat_history
 
 
-def end_interview(problem_description, chat_history, model, client=client):
+def end_interview(problem_description, chat_history, client=client_LLM):
     if not chat_history or len(chat_history) <= 2:
         return "No interview content available to review."
 
@@ -50,7 +55,7 @@ def end_interview(problem_description, chat_history, model, client=client):
         transcript.append(content)
 
     response = client.chat.completions.create(
-        model=model,
+        model=LLM_NAME,
         messages=[
             {"role": "system", "content": grading_feedback_prompt},
             {"role": "user", "content": f"The original problem to solve: {problem_description}"},
@@ -63,12 +68,12 @@ def end_interview(problem_description, chat_history, model, client=client):
     return feedback
 
 
-def send_request(code, previous_code, message, chat_history, chat_display, model, client=client):
+def send_request(code, previous_code, message, chat_history, chat_display, client=client_LLM):
     if code != previous_code:
-        chat_history.append({"role": "user", "content": f"My latest code: {code}"})
+        chat_history.append({"role": "user", "content": f"My latest code:\n{code}"})
     chat_history.append({"role": "user", "content": message})
 
-    response = client.chat.completions.create(model=model, response_format={"type": "json_object"}, messages=chat_history)
+    response = client.chat.completions.create(model=LLM_NAME, response_format={"type": "json_object"}, messages=chat_history)
 
     json_reply = response.choices[0].message.content.strip()
 
@@ -85,16 +90,15 @@ def send_request(code, previous_code, message, chat_history, chat_display, model
     return chat_history, chat_display, "", code
 
 
-def transcribe_audio(audio, client=client):
+def speech_to_text(audio, client=client_STT):
     transcription = client.audio.transcriptions.create(
-        model="whisper-1", file=("temp.wav", numpy_audio_to_bytes(audio[1]), "audio/wav"), response_format="text"
+        model=STT_NAME, file=("temp.wav", numpy_audio_to_bytes(audio[1]), "audio/wav"), response_format="text"
     )
-
     return transcription
 
 
-def text_to_speech(text, client=client):
-    response = client.audio.speech.create(model="tts-1", voice="alloy", input=text)
+def text_to_speech(text, client=client_TTS):
+    response = client.audio.speech.create(model=TTS_NAME, voice="alloy", input=text)
     return response.content
 
 

options.py

@@ -20,10 +20,6 @@ topics_list = [
     "Binary Search Trees",
     "Tries",
 ]
-models = ["gpt-3.5-turbo"]
-# TODO: add more OAI models
-# TODO: add more OS models
-
 
 fixed_messages = {
     "intro": "Welcome to the coding interview! I am your AI interview assistant. For the start select the difficulty and topic of the problem you would like to solve. Then click on the 'Generate a problem' button. Good luck!",