Spaces:
Running
Running
Update main.py
Browse files
main.py
CHANGED
@@ -3,7 +3,7 @@ import time
|
|
3 |
import random
|
4 |
import asyncio
|
5 |
import json
|
6 |
-
from fastapi import FastAPI, HTTPException, Depends
|
7 |
from fastapi.middleware.cors import CORSMiddleware
|
8 |
from fastapi.security.api_key import APIKeyHeader
|
9 |
from pydantic import BaseModel
|
@@ -12,7 +12,10 @@ from dotenv import load_dotenv
|
|
12 |
from starlette.responses import StreamingResponse
|
13 |
from openai import OpenAI
|
14 |
from typing import List, Optional, Dict, Any
|
|
|
15 |
import copy
|
|
|
|
|
16 |
|
17 |
load_dotenv()
|
18 |
|
@@ -26,7 +29,17 @@ API_KEYS = [
|
|
26 |
os.getenv("API_GEMINI_4"),
|
27 |
os.getenv("API_GEMINI_5"),
|
28 |
]
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
# Classi Pydantic di VALIDAZIONE Body
|
31 |
class ChatCompletionRequest(BaseModel):
|
32 |
model: str = "gemini-2.0-flash"
|
@@ -181,14 +194,14 @@ def call_api_sync(params: ChatCompletionRequest):
|
|
181 |
if params.messages:
|
182 |
params.messages = sanitize_messages(params.messages)
|
183 |
params = convert_payload_for_gemini(params)
|
184 |
-
print('
|
185 |
print(params)
|
186 |
response_format = getattr(params, 'response_format', None)
|
187 |
if response_format and getattr(response_format, 'type', None) == 'json_schema':
|
188 |
response = client.beta.chat.completions.parse(**params.model_dump())
|
189 |
else:
|
190 |
response = client.chat.completions.create(**params.model_dump())
|
191 |
-
print('
|
192 |
print(response)
|
193 |
print("")
|
194 |
return response
|
@@ -208,11 +221,21 @@ async def _resp_async_generator(params: ChatCompletionRequest):
|
|
208 |
if params.messages:
|
209 |
params.messages = sanitize_messages(params.messages)
|
210 |
params = convert_payload_for_gemini(params)
|
|
|
|
|
|
|
211 |
for chunk in response:
|
212 |
chunk_data = chunk.to_dict() if hasattr(chunk, "to_dict") else chunk
|
|
|
|
|
|
|
|
|
|
|
213 |
yield f"data: {json.dumps(chunk_data)}\n\n"
|
214 |
await asyncio.sleep(0.01)
|
215 |
yield "data: [DONE]\n\n"
|
|
|
|
|
216 |
except Exception as e:
|
217 |
if "429" in str(e):
|
218 |
await asyncio.sleep(2)
|
@@ -222,6 +245,83 @@ async def _resp_async_generator(params: ChatCompletionRequest):
|
|
222 |
error_data = {"error": str(e)}
|
223 |
yield f"data: {json.dumps(error_data)}\n\n"
|
224 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
# ---------------------------------- Metodi API ---------------------------------------
|
226 |
@app.get("/")
|
227 |
def read_general():
|
@@ -243,6 +343,31 @@ async def chat_completions(req: ChatCompletionRequest):
|
|
243 |
except Exception as e:
|
244 |
raise HTTPException(status_code=500, detail=str(e))
|
245 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
if __name__ == "__main__":
|
247 |
import uvicorn
|
248 |
-
uvicorn.run("main:app", host="0.0.0.0", port=8000, reload=True)
|
|
|
3 |
import random
|
4 |
import asyncio
|
5 |
import json
|
6 |
+
from fastapi import FastAPI, HTTPException, Depends, File, UploadFile, Form
|
7 |
from fastapi.middleware.cors import CORSMiddleware
|
8 |
from fastapi.security.api_key import APIKeyHeader
|
9 |
from pydantic import BaseModel
|
|
|
12 |
from starlette.responses import StreamingResponse
|
13 |
from openai import OpenAI
|
14 |
from typing import List, Optional, Dict, Any
|
15 |
+
import io
|
16 |
import copy
|
17 |
+
from pathlib import Path
|
18 |
+
from pydub import AudioSegment
|
19 |
|
20 |
load_dotenv()
|
21 |
|
|
|
29 |
os.getenv("API_GEMINI_4"),
|
30 |
os.getenv("API_GEMINI_5"),
|
31 |
]
|
32 |
+
GROQ_BASE_URL = "https://api.groq.com/openai/v1"
|
33 |
+
WHISPER_MODEL = "whisper-large-v3-turbo"
|
34 |
+
SEGMENT_MINUTES = 50
|
35 |
+
GROQ_API_KEYS = [
|
36 |
+
os.getenv("API_GROQ_1"),
|
37 |
+
#os.getenv("API_GROQ_2"),
|
38 |
+
#os.getenv("API_GROQ_3"),
|
39 |
+
#os.getenv("API_GROQ_4"),
|
40 |
+
#os.getenv("API_GROQ_5")
|
41 |
+
]
|
42 |
+
|
43 |
# Classi Pydantic di VALIDAZIONE Body
|
44 |
class ChatCompletionRequest(BaseModel):
|
45 |
model: str = "gemini-2.0-flash"
|
|
|
194 |
if params.messages:
|
195 |
params.messages = sanitize_messages(params.messages)
|
196 |
params = convert_payload_for_gemini(params)
|
197 |
+
print('------------------------------------------------------- INPUT ---------------------------------------------------------------')
|
198 |
print(params)
|
199 |
response_format = getattr(params, 'response_format', None)
|
200 |
if response_format and getattr(response_format, 'type', None) == 'json_schema':
|
201 |
response = client.beta.chat.completions.parse(**params.model_dump())
|
202 |
else:
|
203 |
response = client.chat.completions.create(**params.model_dump())
|
204 |
+
print('------------------------------------------------------- OUTPUT ---------------------------------------------------------------')
|
205 |
print(response)
|
206 |
print("")
|
207 |
return response
|
|
|
221 |
if params.messages:
|
222 |
params.messages = sanitize_messages(params.messages)
|
223 |
params = convert_payload_for_gemini(params)
|
224 |
+
print('------------------------------------------------------- INPUT ---------------------------------------------------------------')
|
225 |
+
print(params.model_dump_json(indent=4))
|
226 |
+
final_response_content = ''
|
227 |
for chunk in response:
|
228 |
chunk_data = chunk.to_dict() if hasattr(chunk, "to_dict") else chunk
|
229 |
+
chunk_content = None
|
230 |
+
if chunk.choices and chunk.choices[0].delta:
|
231 |
+
chunk_content = chunk.choices[0].delta.content
|
232 |
+
if chunk_content:
|
233 |
+
final_response_content += chunk_content
|
234 |
yield f"data: {json.dumps(chunk_data)}\n\n"
|
235 |
await asyncio.sleep(0.01)
|
236 |
yield "data: [DONE]\n\n"
|
237 |
+
print('------------------------------------------------------- OUTPUT ---------------------------------------------------------------')
|
238 |
+
print(final_response_content)
|
239 |
except Exception as e:
|
240 |
if "429" in str(e):
|
241 |
await asyncio.sleep(2)
|
|
|
245 |
error_data = {"error": str(e)}
|
246 |
yield f"data: {json.dumps(error_data)}\n\n"
|
247 |
|
248 |
+
|
249 |
+
def get_openai_client():
|
250 |
+
''' Client OpenAI passando in modo RANDOM le Chiavi API. In questo modo posso aggirare i limiti "Quota Exceeded" '''
|
251 |
+
api_key = random.choice(API_KEYS)
|
252 |
+
return OpenAI(api_key=api_key, base_url=BASE_URL)
|
253 |
+
|
254 |
+
# API Whisper Audio:
|
255 |
+
FORMAT_ALIASES = {
|
256 |
+
"mpeg": "mp3",
|
257 |
+
"x-wav": "wav",
|
258 |
+
"vnd.wave": "wav",
|
259 |
+
"x-m4a": "m4a",
|
260 |
+
"x-aac": "aac",
|
261 |
+
}
|
262 |
+
|
263 |
+
def _detect_format(upload_file: UploadFile) -> str:
|
264 |
+
"""Rileva il formato audio dal MIME-type o dall'estensione, con alias safe."""
|
265 |
+
if upload_file.content_type and upload_file.content_type.startswith("audio/"):
|
266 |
+
fmt = upload_file.content_type.split("/", 1)[1]
|
267 |
+
else:
|
268 |
+
fmt = Path(upload_file.filename).suffix.lstrip(".").lower()
|
269 |
+
return FORMAT_ALIASES.get(fmt, fmt)
|
270 |
+
|
271 |
+
def _split_audio_to_mp3_chunks(audio_bytes: bytes, input_format: str, minutes: int):
|
272 |
+
""" Converte (se serve) e splitta. Lascia che ffmpeg auto-rilevi il formato passando format=None: è più sicuro e ignora alias sbagliati. """
|
273 |
+
try:
|
274 |
+
audio = AudioSegment.from_file(io.BytesIO(audio_bytes))
|
275 |
+
except Exception:
|
276 |
+
audio = AudioSegment.from_file(io.BytesIO(audio_bytes), format=input_format)
|
277 |
+
chunk_len_ms = minutes * 60 * 1000
|
278 |
+
for start_ms in range(0, len(audio), chunk_len_ms):
|
279 |
+
chunk = audio[start_ms : start_ms + chunk_len_ms]
|
280 |
+
buf = io.BytesIO()
|
281 |
+
chunk.export(buf, format="mp3")
|
282 |
+
yield buf.getvalue()
|
283 |
+
|
284 |
+
def _transcribe_chunk(chunk_bytes: bytes,
|
285 |
+
model: str,
|
286 |
+
language: str,
|
287 |
+
response_format: str = "json") -> str:
|
288 |
+
bio = io.BytesIO(chunk_bytes)
|
289 |
+
bio.name = "chunk.mp3"
|
290 |
+
resp = call_whisper_api(
|
291 |
+
bio,
|
292 |
+
model=model,
|
293 |
+
language=language,
|
294 |
+
response_format=response_format
|
295 |
+
)
|
296 |
+
if isinstance(resp, str):
|
297 |
+
return resp
|
298 |
+
if hasattr(resp, "text"):
|
299 |
+
return resp.text
|
300 |
+
return resp.get("text", "")
|
301 |
+
|
302 |
+
|
303 |
+
def get_whisper_client():
|
304 |
+
api_key = random.choice(GROQ_API_KEYS)
|
305 |
+
return OpenAI(api_key=api_key, base_url=GROQ_BASE_URL)
|
306 |
+
|
307 |
+
def call_whisper_api(audio_file: io.BytesIO,
|
308 |
+
model: str = WHISPER_MODEL,
|
309 |
+
language: str = "it",
|
310 |
+
response_format: str = "json"):
|
311 |
+
try:
|
312 |
+
client = get_whisper_client()
|
313 |
+
return client.audio.transcriptions.create(
|
314 |
+
file=audio_file,
|
315 |
+
model=model,
|
316 |
+
language=language,
|
317 |
+
response_format=response_format
|
318 |
+
)
|
319 |
+
except Exception as e:
|
320 |
+
if "429" in str(e):
|
321 |
+
time.sleep(2)
|
322 |
+
return call_whisper_api(audio_file, model, language, response_format)
|
323 |
+
raise e
|
324 |
+
|
325 |
# ---------------------------------- Metodi API ---------------------------------------
|
326 |
@app.get("/")
|
327 |
def read_general():
|
|
|
343 |
except Exception as e:
|
344 |
raise HTTPException(status_code=500, detail=str(e))
|
345 |
|
346 |
+
@app.post("/v1/audio/transcriptions", dependencies=[Depends(verify_api_key)])
|
347 |
+
async def audio_transcriptions_endpoint(
|
348 |
+
file: UploadFile = File(...),
|
349 |
+
model: str = Form(WHISPER_MODEL),
|
350 |
+
language: str = Form("it"),
|
351 |
+
response_format: str = Form("text"),
|
352 |
+
segment_minutes: int = Form(SEGMENT_MINUTES)):
|
353 |
+
try:
|
354 |
+
raw_bytes = await file.read()
|
355 |
+
input_fmt = _detect_format(file) or "mp3"
|
356 |
+
chunks = list(_split_audio_to_mp3_chunks(raw_bytes, input_fmt, segment_minutes))
|
357 |
+
if not chunks:
|
358 |
+
raise ValueError("Audio vuoto o formato non riconosciuto")
|
359 |
+
transcripts = [_transcribe_chunk(c, model, language, response_format) for c in chunks]
|
360 |
+
final_text = "\n\n".join(transcripts)
|
361 |
+
return {
|
362 |
+
"model": model,
|
363 |
+
"language": language,
|
364 |
+
"segments": len(transcripts),
|
365 |
+
"segment_minutes": segment_minutes,
|
366 |
+
"text": final_text,
|
367 |
+
}
|
368 |
+
except Exception as e:
|
369 |
+
raise HTTPException(status_code=500, detail=str(e))
|
370 |
+
|
371 |
if __name__ == "__main__":
|
372 |
import uvicorn
|
373 |
+
uvicorn.run("main:app", host="0.0.0.0", port=8000, reload=True)
|