app.py

from dotenv import load_dotenv


import gradio as gr
from gradio import ChatMessage

import json
from openai import OpenAI
from datetime import datetime
import os
import re
import logging

logging.basicConfig(level=logging.INFO, format='[%(asctime)s][%(levelname)s] - %(message)s')
# logging.getLogger().setLevel(logging.INFO)


load_dotenv(".env", override=True) 
HF_TOKEN = os.environ.get("HF_TOKEN")  
BASE_URL = os.environ.get("BASE_URL")
EMBEDDINGS = os.environ.get("EMBEDDINGS_MODEL")



from tools import tools, oitools

SYSTEM_PROMPT_TEMPLATE = """You are an AI assistant. Follow these rules when responding to user questions:
- **query**: When forming a query for `get_documents`, you **can modify** the original query to improve clarity, reduce ambiguity, or make it more specific. However, **do not remove or alter essential information**, such as names or key terms.
- **redundant_search**: Only use `get_documents` when strictly necessary. Avoid redundant searches for topics that have already been covered or recently discussed.
- **external_info_only**: You must **not use internal memory or knowledge** to respond. All information provided should come solely from the documents retrieved using the `get_documents` tool.
- **no_info_found**: If `get_documents` does not return any information, you must inform the user that you are unable to find the answer.
- **missing_info**: If the information needed to answer the user's question is not already present in previous tool responses, you must call `get_documents` to retrieve the necessary content.

- **date**: Today’s date is **{date}**. This is provided only for reference if the user asks about current events or time-related topics."""


client = OpenAI(
    base_url=f"{BASE_URL}/v1",  
    api_key=HF_TOKEN
)
logging.info(f"Client initialized: {client}")

def today_date():
    return datetime.today().strftime('%A, %B %d, %Y, %I:%M %p')


def clean_json_string(json_str):
    return re.sub(r'[ ,}\s]+$', '', json_str) + '}'


def completion(history, model, system_prompt: str, tools=None):
    messages = [{"role": "system", "content": system_prompt.format(date=today_date())}]
    for msg in history:
        if isinstance(msg, dict):  
            msg = ChatMessage(**msg)
        if msg.role == "assistant" and hasattr(msg, "metadata") and msg.metadata:  
            tools_calls = json.loads(msg.metadata.get("title", "[]")) 
            # for tool_calls in tools_calls:
            #     tool_calls["function"]["arguments"] = json.loads(tool_calls["function"]["arguments"])
            messages.append({"role": "assistant", "tool_calls": tools_calls, "content": ""})
            messages.append({"role": "tool", "content": msg.content})
        else:
            messages.append({"role": msg.role, "content": msg.content})
    
    request_params = {
        "model": model,
        "messages": messages,
        "stream": True,
        "max_tokens": 1000,
        "temperature": 0.2,
        "frequency_penalty": 1,
        "extra_body": {"repetition_penalty": 1.1},
    }
    if tools:
        request_params.update({"tool_choice": "auto", "tools": tools})
    
    return client.chat.completions.create(**request_params)  

def llm_in_loop(history, system_prompt, recursive):  
    try:   
        models = client.models.list()
        model = models.data[0].id if models.data else "gpt-3.5-turbo"  
    except Exception as err:
        gr.Warning("The model is initializing. Please wait; this may take 5 to 10 minutes ⏳.", duration=20)
        raise err
    
    arguments = ""
    name = ""
    chat_completion = completion(history=history, tools=oitools, model=model, system_prompt=system_prompt)  
    appended = False
    # if chat_completion.choices and chat_completion.choices[0].message.tool_calls:
    #     call = chat_completion.choices[0].message.tool_calls[0]
    #     if hasattr(call.function, "name") and call.function.name:
    #         name = call.function.name
    #     if hasattr(call.function, "arguments") and call.function.arguments:
    #         arguments += call.function.arguments
    # elif chat_completion.choices[0].message.content:
    #     if not appended:
    #         history.append(ChatMessage(role="assistant", content=""))
    #         appended = True
    #     history[-1].content += chat_completion.choices[0].message.content
    #     yield history[recursive:]
    for chunk in chat_completion:
        if chunk.choices and chunk.choices[0].delta.tool_calls:
            call = chunk.choices[0].delta.tool_calls[0]
            if hasattr(call.function, "name") and call.function.name:
                name = call.function.name
            if hasattr(call.function, "arguments") and call.function.arguments:
                arguments += call.function.arguments
        elif chunk.choices[0].delta.content:
            if not appended:
                history.append(ChatMessage(role="assistant", content=""))
                appended = True
            history[-1].content += chunk.choices[0].delta.content
            yield history[recursive:]
    
    arguments = clean_json_string(arguments) if arguments else "{}"
    print(name, arguments)
    arguments = json.loads(arguments)
    print(name, arguments)
    print("====================")
    if appended:
        recursive -= 1
    if name:
        result = f"💥 Error using tool {name}, tool doesn't exist" if name not in tools else str(tools[name].invoke(input=arguments))
        result = json.dumps({name: result}, ensure_ascii=False)
        # msg = ChatMessage(
        #             role="assistant",
        #             content="",
        #             metadata= {"title": f"🛠️ Using tool '{name}', arguments: {json.dumps(json_arguments, ensure_ascii=False)}"},
        #             options=[{"label":"tool_calls", "value": json.dumps([{"id": "call_FthC9qRpsL5kBpwwyw6c7j4k","function": {"arguments": arguments,"name": name},"type": "function"}])}]
        #         )
        history.append(ChatMessage(role="assistant", content=result, metadata={"title": json.dumps([{"id": "call_id", "function": {"arguments": json.dumps(arguments, ensure_ascii=False), "name": name}, "type": "function"}], ensure_ascii=False)}))
        yield history[recursive:]
        yield from llm_in_loop(history, system_prompt, recursive - 1)

def respond(message, history, additional_inputs):  
    history.append(ChatMessage(role="user", content=message))
    yield from llm_in_loop(history, additional_inputs, -1)

if __name__ == "__main__":
    system_prompt = gr.Textbox(label="System prompt", value=SYSTEM_PROMPT_TEMPLATE, lines=3)  
    demo = gr.ChatInterface(respond, type="messages", additional_inputs=[system_prompt])
    demo.launch()