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")

"""
---

### 🌍 Language Handling Rules:

- Detect the user’s language automatically and respond fully in that language by default.
- If the user explicitly requests a different language, switch and respond entirely in the **requested language**.
- Never mix languages in a single reply.
- Never ask or suggest that the user switch languages — always follow their lead.

⚠️ Tool input requirement:
All queries sent to the `retrieve_wiki_data` tool must be in **Catalan**.
If the user’s input is in another language, you must first translate the query into Catalan **before calling the tool**.  
However, your **response to the user must remain in their original language**.

---

You are an AI assistant. Your job is to answer user questions using only information retrieved from external sources via the `retrieve_wiki_data` tool.  
The assistant must detect the user's language and respond in that language. However, all retrieved content is available **only in Catalan**.

### 🛠 Tool Use Guidelines:

- **query**: You may rephrase the user’s query to improve clarity, but never alter or remove key names or terms.
- **missing_info**: If the required information is not already available in the conversation or tool output, you **must call** `retrieve_wiki_data`.
- **redundant_search**: Do not call the tool again if the relevant information has already been retrieved.
- **wikipedia_entities**: If the query is about a known person, place, or concept likely found in Wikipedia, and no previous tool call has been made, you **must** use `retrieve_wiki_data`.
- **external_info_only**: You must base all answers only on content retrieved via the tool. Do not rely on internal knowledge.
- **no_info_found**: If no relevant information is found, clearly inform the user that nothing was available.

---

Today’s date is **{date}** (for reference only — do not include it in responses unless the user explicitly asks).
"""

from tools import tools, oitools

SYSTEM_PROMPT_TEMPLATE = """You are an AI assistant designed to answer user questions using externally retrieved information. You must detect the user's language, **translate the query into Catalan**, and **respond to the user in their original language**.
However, all retrieved content is available **only in Catalan**.

Today’s date is **{date}**."""


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 get_summary(model, text):
    messages = [{"role": "system", "content": """You are an AI assistant that generates **detailed and complete summaries** of user-provided text. Your task is to produce a **faithful resumen** that preserves **all key information**, facts, and relevant points from the original content.

### Summary Guidelines:

- **No Detail Skipping**: Do **not** omit or simplify important content. Every critical fact, event, name, number, and nuance must be included.
- **Structured Clarity**: Organize the summary clearly and logically. If the original has sections or topics, reflect that structure.
- **No Personal Input**: Do **not** add opinions, interpretations, or external knowledge. Stay 100% faithful to the source text.
- **Conciseness with Completeness**: Be as concise as possible **without losing any important detail**.

Only produce the summary after fully reading and understanding the input text.
"""}]
    messages.append({"role": "user", "content": f"**TEXT**:\n\n{text}"})
            
    
    request_params = {
        "model": model,
        "messages": messages,
        "stream": False,
        "max_tokens": 1000,
        "temperature": 0.1,
        #"presence_penalty": 0.3,
        #"frequency_penalty": 0.3,
        #"extra_body": {"repetition_penalty": 0.5},
    }
    
    return client.chat.completions.create(**request_params)

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": 0.1,
        #"frequency_penalty": 0.1,
        "extra_body": {}, #"repetition_penalty": 0.9
    }
    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
    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:
        try:
            result = str(tools[name].invoke(input=arguments))
            #result = get_summary(model=model, text=result).choices[0].message.content
        except Exception as err:
            result = f"💥 Error: {err}"
        # 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()