main.py

import os
import time
import random
import asyncio
import json
from fastapi import FastAPI, HTTPException, Depends, File, UploadFile, Form  
from fastapi.middleware.cors import CORSMiddleware
from fastapi.security.api_key import APIKeyHeader
from pydantic import BaseModel
from typing import List, Optional
from dotenv import load_dotenv
from starlette.responses import StreamingResponse
from openai import OpenAI
from typing import List, Optional, Dict, Any
import io
import copy
from pathlib import Path
from pydub import AudioSegment

load_dotenv()

BASE_URL = "https://generativelanguage.googleapis.com/v1beta/openai/"
EXPECTED_API_KEY = os.getenv("API_HUGGINGFACE")
API_KEY_NAME = "Authorization"
API_KEYS = [
    os.getenv("API_GEMINI_1"),
    os.getenv("API_GEMINI_2"),
    os.getenv("API_GEMINI_3"), 
    os.getenv("API_GEMINI_4"), 
    os.getenv("API_GEMINI_5"), 
]
GROQ_BASE_URL  = "https://api.groq.com/openai/v1"
WHISPER_MODEL  = "whisper-large-v3-turbo"     
SEGMENT_MINUTES = 50
GROQ_API_KEYS  = [
    os.getenv("API_GROQ_1"),
    #os.getenv("API_GROQ_2"),
    #os.getenv("API_GROQ_3"),
    #os.getenv("API_GROQ_4"),
    #os.getenv("API_GROQ_5")
]
                           
# Classi Pydantic di VALIDAZIONE Body
class ChatCompletionRequest(BaseModel):
    model: str = "gemini-2.0-flash" 
    messages: Optional[Any]
    temperature: Optional[float] = 0.8
    stream: Optional[bool] = False
    stream_options: Optional[Dict[str, Any]] = None
    class Config:
        extra = "allow"

# Server FAST API 
app = FastAPI(title="OpenAI-SDK-compatible API", version="1.0.0", description="Un wrapper FastAPI compatibile con le specifiche dell'API OpenAI.")
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Client OpenAI
def get_openai_client():
    ''' Client OpenAI passando in modo RANDOM le Chiavi API. In questo modo posso aggirare i limiti "Quota Exceeded" '''
    api_key = random.choice(API_KEYS)
    return OpenAI(api_key=api_key, base_url=BASE_URL)

# Validazione API       
api_key_header = APIKeyHeader(name=API_KEY_NAME, auto_error=False)
def verify_api_key(api_key: str = Depends(api_key_header)):
    ''' Validazione Chiave API - Per ora in ENV, Token HF '''
    if not api_key:
        raise HTTPException(status_code=403, detail="API key mancante")
    if api_key != f"Bearer {EXPECTED_API_KEY}":
        raise HTTPException(status_code=403, detail="API key non valida")
    return api_key

# Correzione payload con content=None
def sanitize_messages(messages):
    """Convert None content to empty string to avoid Gemini API errors"""
    if not messages:
        return messages
    for message in messages:
        if message.get('content') is None:
            message['content'] = " "
    return messages

# Funzione per conversione Payload OpenAI to GEMINI (anomalia per ACTION) AnyOf, e property: {}
def convert_openai_schema_for_gemini(tools_schema):
    if isinstance(tools_schema, str):
        try:
            tools_schema = json.loads(tools_schema)
        except json.JSONDecodeError:
            raise ValueError("Stringa JSON non valida fornita")
    converted_schema = []
    for tool in tools_schema:
        if tool.get("type") != "function":
            converted_schema.append(tool)
            continue
        converted_tool = {"type": "function", "function": {}}
        func_def = tool.get("function", {})
        if not func_def:
            continue
        converted_tool["function"]["name"] = func_def.get("name", "")
        converted_tool["function"]["description"] = func_def.get("description", "")
        if "parameters" in func_def:
            params = func_def["parameters"]
            converted_params = {"type": "object"}
            if "properties" in params:
                converted_properties = {}
                for prop_name, prop_value in params["properties"].items():
                    cleaned = clean_schema_property(prop_value)
                    if cleaned:  
                        converted_properties[prop_name] = cleaned
                if converted_properties:
                    converted_params["properties"] = converted_properties
                else:
                    converted_params["properties"] = {"parameter": {"type": "string"}}
            else:
                converted_params["properties"] = {"parameter": {"type": "string"}}
            if "required" in params:
                converted_params["required"] = params["required"]
            converted_tool["function"]["parameters"] = converted_params
        converted_schema.append(converted_tool)
    return converted_schema

def clean_schema_property(prop):
    if not isinstance(prop, dict):
        return prop
    result = {}
    for key, value in prop.items():
        if key in ("title", "default"):
            continue
        elif key == "anyOf":
            if isinstance(value, list):
                for item in value:
                    if isinstance(item, dict) and item.get("type") != "null":
                        cleaned_item = clean_schema_property(item)
                        for k, v in cleaned_item.items():
                            if k not in result:
                                result[k] = v
                        break
        elif key == "oneOf":
            if isinstance(value, list) and len(value) > 0:
                cleaned_item = clean_schema_property(value[0])
                for k, v in cleaned_item.items():
                    if k not in result:
                        result[k] = v
            elif isinstance(value, dict):
                cleaned_item = clean_schema_property(value)
                for k, v in cleaned_item.items():
                    if k not in result:
                        result[k] = v
        elif key == "properties" and isinstance(value, dict):
            new_props = {}
            for prop_name, prop_value in value.items():
                cleaned_prop = clean_schema_property(prop_value)
                if cleaned_prop:
                    new_props[prop_name] = cleaned_prop
            if not new_props:
                new_props = {"parameter": {"type": "string"}}
            result[key] = new_props
        elif key == "items" and isinstance(value, dict):
            result[key] = clean_schema_property(value)
        elif isinstance(value, list):
            result[key] = [clean_schema_property(item) if isinstance(item, dict) else item for item in value]
        else:
            result[key] = value
    if result.get("type") == "object" and ("properties" not in result or not result["properties"]):
        result["properties"] = {"parameter": {"type": "string"}}
    return result

def convert_payload_for_gemini(payload: ChatCompletionRequest):
    if hasattr(payload, "model_dump"):
        payload_converted = json.loads(payload.model_dump_json())
    elif isinstance(payload, dict):
        payload_converted = payload.copy()
    else:
        raise ValueError("Formato payload non supportato")
    payload_converted.pop("metadata", None)
    payload_converted.pop("store", None)
    if "tools" in payload_converted:
        payload_converted["tools"] = convert_openai_schema_for_gemini(payload_converted["tools"])
    new_payload = ChatCompletionRequest.model_validate(payload_converted)
    return new_payload

# ---------------------------------- Funzioni per Chat Completion ---------------------------------------
# Chiama API (senza Streaming)
def call_api_sync(params: ChatCompletionRequest):
    ''' Chiamata API senza streaming. Se da errore 429 lo rifa'''
    try:
        client = get_openai_client()
        if params.messages:
            params.messages = sanitize_messages(params.messages)       
        params = convert_payload_for_gemini(params)
        print('------------------------------------------------------- INPUT ---------------------------------------------------------------')
        print(params)
        response_format = getattr(params, 'response_format', None)
        if response_format and getattr(response_format, 'type', None) == 'json_schema':    
            response = client.beta.chat.completions.parse(**params.model_dump())
        else:
            response = client.chat.completions.create(**params.model_dump())
        print('------------------------------------------------------- OUTPUT ---------------------------------------------------------------')
        print(response)
        print("")
        return response
    except Exception as e:
        if "429" in str(e):
            time.sleep(2)
            return call_api_sync(params)
        else:
            raise e

# Chiama API (con Streaming)
async def _resp_async_generator(params: ChatCompletionRequest):
    ''' Chiamata API con streaming. Se da errore 429 lo rifa'''
    client = get_openai_client()
    try:
        response = client.chat.completions.create(**params.model_dump())
        if params.messages:
            params.messages = sanitize_messages(params.messages)     
        params = convert_payload_for_gemini(params)
        print('------------------------------------------------------- INPUT ---------------------------------------------------------------')
        print(params.model_dump_json(indent=4))
        final_response_content = ''
        for chunk in response:
            chunk_data = chunk.to_dict() if hasattr(chunk, "to_dict") else chunk
            chunk_content = None
            if chunk.choices and chunk.choices[0].delta:
                chunk_content = chunk.choices[0].delta.content
            if chunk_content:
                final_response_content += chunk_content
            yield f"data: {json.dumps(chunk_data)}\n\n"
            await asyncio.sleep(0.01)
        yield "data: [DONE]\n\n"
        print('------------------------------------------------------- OUTPUT ---------------------------------------------------------------')
        print(final_response_content)
    except Exception as e:
        if "429" in str(e):
            await asyncio.sleep(2)
            async for item in _resp_async_generator(params):
                yield item
        else:
            error_data = {"error": str(e)}
            yield f"data: {json.dumps(error_data)}\n\n"


def get_openai_client():
    ''' Client OpenAI passando in modo RANDOM le Chiavi API. In questo modo posso aggirare i limiti "Quota Exceeded" '''
    api_key = random.choice(API_KEYS)
    return OpenAI(api_key=api_key, base_url=BASE_URL)

# API Whisper Audio:
FORMAT_ALIASES = {
    "mpeg": "mp3",
    "x-wav": "wav",
    "vnd.wave": "wav",
    "x-m4a": "m4a",
    "x-aac": "aac",
}

def _detect_format(upload_file: UploadFile) -> str:
    """Rileva il formato audio dal MIME-type o dall'estensione, con alias safe."""
    if upload_file.content_type and upload_file.content_type.startswith("audio/"):
        fmt = upload_file.content_type.split("/", 1)[1]
    else:
        fmt = Path(upload_file.filename).suffix.lstrip(".").lower()
    return FORMAT_ALIASES.get(fmt, fmt)

def _split_audio_to_mp3_chunks(audio_bytes: bytes, input_format: str, minutes: int):
    """ Converte (se serve) e splitta. Lascia che ffmpeg auto-rilevi il formato passando format=None: è più sicuro e ignora alias sbagliati. """
    try:
        audio = AudioSegment.from_file(io.BytesIO(audio_bytes)) 
    except Exception:
        audio = AudioSegment.from_file(io.BytesIO(audio_bytes), format=input_format)
    chunk_len_ms = minutes * 60 * 1000
    for start_ms in range(0, len(audio), chunk_len_ms):
        chunk = audio[start_ms : start_ms + chunk_len_ms]
        buf = io.BytesIO()
        chunk.export(buf, format="mp3")
        yield buf.getvalue()

def _transcribe_chunk(chunk_bytes: bytes,
                      model: str,
                      language: str,
                      response_format: str = "json") -> str:
    bio = io.BytesIO(chunk_bytes)
    bio.name = "chunk.mp3"
    resp = call_whisper_api(
        bio,
        model=model,
        language=language,
        response_format=response_format
    )
    if isinstance(resp, str):
        return resp
    if hasattr(resp, "text"):
        return resp.text
    return resp.get("text", "")


def get_whisper_client():
    api_key = random.choice(GROQ_API_KEYS)
    return OpenAI(api_key=api_key, base_url=GROQ_BASE_URL)

def call_whisper_api(audio_file: io.BytesIO,
                     model: str = WHISPER_MODEL,
                     language: str = "it",
                     response_format: str = "json"):
    try:
        client = get_whisper_client()
        return client.audio.transcriptions.create(
            file=audio_file,
            model=model,
            language=language,
            response_format=response_format
        )
    except Exception as e:
        if "429" in str(e):
            time.sleep(2)
            return call_whisper_api(audio_file, model, language, response_format)
        raise e

# ---------------------------------- Metodi API ---------------------------------------
@app.get("/")
def read_general(): 
    return {"response": "Benvenuto"}

@app.get("/health")
async def health_check():
    return {"message": "success"}

@app.post("/v1/chat/completions", dependencies=[Depends(verify_api_key)])
async def chat_completions(req: ChatCompletionRequest):
    try:    
        if not req.messages:
            raise HTTPException(status_code=400, detail="Nessun messaggio fornito")
        if not req.stream:
            return call_api_sync(req)
        else:
            return StreamingResponse(_resp_async_generator(req), media_type="application/x-ndjson")            
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

@app.post("/v1/audio/transcriptions", dependencies=[Depends(verify_api_key)])
async def audio_transcriptions_endpoint(
    file: UploadFile = File(...),
    model: str = Form(WHISPER_MODEL),
    language: str = Form("it"),
    response_format: str = Form("text"),
    segment_minutes: int = Form(SEGMENT_MINUTES)):
    try:
        raw_bytes = await file.read()
        input_fmt  = _detect_format(file) or "mp3"
        chunks = list(_split_audio_to_mp3_chunks(raw_bytes, input_fmt, segment_minutes))
        if not chunks:
            raise ValueError("Audio vuoto o formato non riconosciuto")
        transcripts = [_transcribe_chunk(c, model, language, response_format) for c in chunks]
        final_text = "\n\n".join(transcripts)
        return {
            "model": model,
            "language": language,
            "segments": len(transcripts),
            "segment_minutes": segment_minutes,
            "text": final_text,
        }
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

if __name__ == "__main__":
    import uvicorn
    uvicorn.run("main:app", host="0.0.0.0", port=8000, reload=True)