cleaned version
This commit is contained in:
@@ -5,7 +5,7 @@ Nepoužíváme těžké SDK — voláme REST endpoint přímo, aby nebyl závisl
|
||||
import httpx
|
||||
|
||||
from ..config import DEEPGRAM_BASE_URL, UPSTREAM_TIMEOUT_SECONDS
|
||||
from ..errors import UpstreamError
|
||||
from ..errors import UpstreamError, raise_for_upstream
|
||||
from ..logging_config import get_logger
|
||||
|
||||
log = get_logger("audio-transcription.deepgram")
|
||||
@@ -43,12 +43,8 @@ async def transcribe(
|
||||
|
||||
if resp.status_code >= 400:
|
||||
# Deepgram vrací chybu v JSON; nikdy nelogujeme klíč (ten je jen v hlavičce).
|
||||
snippet = resp.text[:500]
|
||||
log.warning("Deepgram returned %s: %s", resp.status_code, snippet)
|
||||
raise UpstreamError(
|
||||
f"Deepgram vrátil chybu {resp.status_code}.",
|
||||
detail=snippet,
|
||||
)
|
||||
log.warning("Deepgram returned %s: %s", resp.status_code, resp.text[:500])
|
||||
raise_for_upstream("Deepgram", resp.status_code, resp.text)
|
||||
|
||||
data = resp.json()
|
||||
try:
|
||||
|
||||
@@ -2,7 +2,7 @@
|
||||
import httpx
|
||||
|
||||
from ..config import OPENAI_BASE_URL, UPSTREAM_TIMEOUT_SECONDS
|
||||
from ..errors import UpstreamError
|
||||
from ..errors import UpstreamError, raise_for_upstream
|
||||
from ..logging_config import get_logger
|
||||
|
||||
log = get_logger("audio-transcription.openai")
|
||||
@@ -33,12 +33,8 @@ async def transcribe(
|
||||
raise UpstreamError(f"OpenAI Whisper request se nezdařil: {exc}") from exc
|
||||
|
||||
if resp.status_code >= 400:
|
||||
snippet = resp.text[:500]
|
||||
log.warning("OpenAI transcription returned %s: %s", resp.status_code, snippet)
|
||||
raise UpstreamError(
|
||||
f"OpenAI Whisper vrátil chybu {resp.status_code}.",
|
||||
detail=snippet,
|
||||
)
|
||||
log.warning("OpenAI transcription returned %s: %s", resp.status_code, resp.text[:500])
|
||||
raise_for_upstream("OpenAI Whisper", resp.status_code, resp.text)
|
||||
|
||||
data = resp.json()
|
||||
return data.get("text", "") or ""
|
||||
@@ -72,12 +68,8 @@ async def merge_transcripts(
|
||||
raise UpstreamError(f"OpenAI chat (merge) request se nezdařil: {exc}") from exc
|
||||
|
||||
if resp.status_code >= 400:
|
||||
snippet = resp.text[:500]
|
||||
log.warning("OpenAI chat returned %s: %s", resp.status_code, snippet)
|
||||
raise UpstreamError(
|
||||
f"OpenAI chat (merge) vrátil chybu {resp.status_code}.",
|
||||
detail=snippet,
|
||||
)
|
||||
log.warning("OpenAI chat returned %s: %s", resp.status_code, resp.text[:500])
|
||||
raise_for_upstream("OpenAI chat (merge)", resp.status_code, resp.text)
|
||||
|
||||
data = resp.json()
|
||||
try:
|
||||
|
||||
+32
-1
@@ -35,12 +35,43 @@ class BadRequestError(TranscriptionError):
|
||||
status_code = 400
|
||||
|
||||
|
||||
class RateLimitError(TranscriptionError):
|
||||
"""Upstream rate limit (HTTP 429)."""
|
||||
|
||||
status_code = 429
|
||||
|
||||
|
||||
class UpstreamError(TranscriptionError):
|
||||
"""Chyba při volání upstream API (Deepgram / OpenAI / stažení audia)."""
|
||||
"""Chyba při volání upstream API (Deepgram / OpenAI)."""
|
||||
|
||||
status_code = 502
|
||||
|
||||
|
||||
def raise_for_upstream(upstream: str, status: int, body: str) -> None:
|
||||
"""Zmapuje HTTP status z upstreamu (Deepgram/OpenAI) na správnou chybu služby.
|
||||
|
||||
- 400/415/422 → 400 (špatný / nepodporovaný audio soubor = chyba klienta)
|
||||
- 401/403 → 401 (upstream odmítl API klíč volajícího)
|
||||
- 429 → 429 (rate limit)
|
||||
- jinak → 502 (výpadek / neočekávaná chyba upstreamu)
|
||||
"""
|
||||
snippet = (body or "")[:500]
|
||||
if status in (400, 415, 422):
|
||||
raise BadRequestError(
|
||||
f"{upstream} odmítl audio (HTTP {status}) — pravděpodobně nepodporovaný "
|
||||
f"formát nebo poškozený soubor.",
|
||||
detail=snippet,
|
||||
)
|
||||
if status in (401, 403):
|
||||
raise CredentialsError(
|
||||
f"{upstream} odmítl API klíč (HTTP {status}).",
|
||||
detail=snippet,
|
||||
)
|
||||
if status == 429:
|
||||
raise RateLimitError(f"{upstream} rate limit (HTTP 429).", detail=snippet)
|
||||
raise UpstreamError(f"{upstream} vrátil chybu {status}.", detail=snippet)
|
||||
|
||||
|
||||
def register_exception_handlers(app: FastAPI) -> None:
|
||||
@app.exception_handler(TranscriptionError)
|
||||
async def _handle_transcription_error(request: Request, exc: TranscriptionError):
|
||||
|
||||
+40
-101
@@ -1,11 +1,12 @@
|
||||
"""Endpointy pro přepis audia.
|
||||
"""Endpoint pro přepis audia.
|
||||
|
||||
- POST /transcribe/dual → paralelní přepis Deepgram + OpenAI Whisper (dva texty)
|
||||
- POST /transcribe/combined → dual + sloučení do jednoho co nejlepšího přepisu (merge)
|
||||
POST /transcribe — udělá vše najednou: audio přepíše paralelně Deepgramem i OpenAI
|
||||
(whisper-1) a oba texty sloučí přes OpenAI Chat podle `combine_prompt` do jednoho
|
||||
co nejlepšího přepisu.
|
||||
|
||||
Audio přichází jako proměnná (upload `file` ve form-data). Všechny ostatní parametry
|
||||
jsou v POST těle (form fields). API klíče jsou v X- hlavičkách (viz credentials.py).
|
||||
Výstup je vždy JSON.
|
||||
Výstup je vždy JSON: {"gpt": ..., "deepgram": ..., "merge": ...}.
|
||||
"""
|
||||
import asyncio
|
||||
|
||||
@@ -29,17 +30,10 @@ log = get_logger("audio-transcription.transcribe")
|
||||
router = APIRouter(tags=["transcribe"])
|
||||
|
||||
|
||||
class DualResult(BaseModel):
|
||||
text1: str = Field(..., description="Přepis z Deepgramu.")
|
||||
text2: str = Field(..., description="Přepis z OpenAI Whisper.")
|
||||
deepgram_model: str
|
||||
whisper_model: str
|
||||
language: str
|
||||
|
||||
|
||||
class CombinedResult(DualResult):
|
||||
merged: str = Field(..., description="Sloučený, co nejpřesnější výsledný přepis.")
|
||||
chat_model: str
|
||||
class TranscribeResult(BaseModel):
|
||||
gpt: str = Field(..., description="Přepis z OpenAI (model whisper-1).")
|
||||
deepgram: str = Field(..., description="Přepis z Deepgramu.")
|
||||
merge: str = Field(..., description="Sloučený, co nejpřesnější výsledný přepis.")
|
||||
|
||||
|
||||
async def _read_upload(file: UploadFile) -> tuple[bytes, str, str]:
|
||||
@@ -52,22 +46,36 @@ async def _read_upload(file: UploadFile) -> tuple[bytes, str, str]:
|
||||
return audio, filename, content_type
|
||||
|
||||
|
||||
async def _transcribe_dual(
|
||||
creds: Credentials,
|
||||
audio: bytes,
|
||||
filename: str,
|
||||
content_type: str,
|
||||
deepgram_model: str,
|
||||
whisper_model: str,
|
||||
language: str,
|
||||
diarize: bool,
|
||||
smart_format: bool,
|
||||
) -> tuple[str, str]:
|
||||
"""Spustí oba přepisy paralelně; vrátí (deepgram_text, whisper_text)."""
|
||||
@router.post(
|
||||
"/dual-with-merge",
|
||||
response_model=TranscribeResult,
|
||||
summary="Přepis audia (Deepgram + OpenAI) a AI sloučení — vše najednou",
|
||||
description="Přijme audio soubor, vytvoří paralelně přepis z Deepgramu i z OpenAI "
|
||||
"(whisper-1) a poté je sloučí přes OpenAI Chat podle `combine_prompt` do jediného, "
|
||||
"co nejpřesnějšího českého přepisu. Vrací JSON `{gpt, deepgram, merge}`.",
|
||||
)
|
||||
async def transcribe(
|
||||
file: UploadFile = File(..., description="Audio soubor k přepisu (proměnná)."),
|
||||
combine_prompt: str = Form(
|
||||
DEFAULT_COMBINE_PROMPT,
|
||||
description="System prompt pro sloučení. Nezadáš-li, použije se výchozí "
|
||||
"'Czech Transcript Merger'.",
|
||||
),
|
||||
deepgram_model: str = Form(DEFAULT_DEEPGRAM_MODEL, description="Deepgram model."),
|
||||
whisper_model: str = Form(DEFAULT_WHISPER_MODEL, description="OpenAI přepisový model."),
|
||||
chat_model: str = Form(DEFAULT_CHAT_MODEL, description="OpenAI chat model pro sloučení."),
|
||||
language: str = Form(DEFAULT_LANGUAGE, description="Jazyk audia (ISO kód, např. cs)."),
|
||||
diarize: bool = Form(True, description="Deepgram diarizace (rozlišení mluvčích)."),
|
||||
smart_format: bool = Form(True, description="Deepgram smart formatting."),
|
||||
creds: Credentials = Depends(get_credentials),
|
||||
) -> TranscribeResult:
|
||||
deepgram_key = creds.require_deepgram()
|
||||
openai_key = creds.require_openai()
|
||||
|
||||
text1, text2 = await asyncio.gather(
|
||||
audio, filename, content_type = await _read_upload(file)
|
||||
|
||||
# Oba přepisy paralelně.
|
||||
text_deepgram, text_gpt = await asyncio.gather(
|
||||
deepgram_client.transcribe(
|
||||
api_key=deepgram_key,
|
||||
audio=audio,
|
||||
@@ -86,83 +94,14 @@ async def _transcribe_dual(
|
||||
language=language,
|
||||
),
|
||||
)
|
||||
return text1, text2
|
||||
|
||||
|
||||
@router.post(
|
||||
"/transcribe/dual",
|
||||
response_model=DualResult,
|
||||
summary="Paralelní přepis Deepgram + OpenAI Whisper",
|
||||
description="Přijme audio soubor a vrátí dva nezávislé přepisy: "
|
||||
"`text1` z Deepgramu a `text2` z OpenAI Whisper.",
|
||||
)
|
||||
async def transcribe_dual(
|
||||
file: UploadFile = File(..., description="Audio soubor k přepisu (proměnná)."),
|
||||
deepgram_model: str = Form(DEFAULT_DEEPGRAM_MODEL, description="Deepgram model."),
|
||||
whisper_model: str = Form(DEFAULT_WHISPER_MODEL, description="OpenAI přepisový model."),
|
||||
language: str = Form(DEFAULT_LANGUAGE, description="Jazyk audia (ISO kód, např. cs)."),
|
||||
diarize: bool = Form(True, description="Deepgram diarizace (rozlišení mluvčích)."),
|
||||
smart_format: bool = Form(True, description="Deepgram smart formatting."),
|
||||
creds: Credentials = Depends(get_credentials),
|
||||
) -> DualResult:
|
||||
audio, filename, content_type = await _read_upload(file)
|
||||
text1, text2 = await _transcribe_dual(
|
||||
creds, audio, filename, content_type,
|
||||
deepgram_model, whisper_model, language, diarize, smart_format,
|
||||
)
|
||||
return DualResult(
|
||||
text1=text1,
|
||||
text2=text2,
|
||||
deepgram_model=deepgram_model,
|
||||
whisper_model=whisper_model,
|
||||
language=language,
|
||||
)
|
||||
|
||||
|
||||
@router.post(
|
||||
"/transcribe/combined",
|
||||
response_model=CombinedResult,
|
||||
summary="Dual přepis + AI sloučení do nejlepšího přepisu",
|
||||
description="Přijme audio soubor, vytvoří přepis z Deepgramu i Whisperu a poté je "
|
||||
"sloučí přes OpenAI Chat podle `combine_prompt` do jediného, co nejpřesnějšího "
|
||||
"českého přepisu. Vrací `text1`, `text2` i výsledný `merged`.",
|
||||
)
|
||||
async def transcribe_combined(
|
||||
file: UploadFile = File(..., description="Audio soubor k přepisu (proměnná)."),
|
||||
combine_prompt: str = Form(
|
||||
DEFAULT_COMBINE_PROMPT,
|
||||
description="System prompt pro sloučení. Nezadáš-li, použije se výchozí "
|
||||
"'Czech Transcript Merger'.",
|
||||
),
|
||||
deepgram_model: str = Form(DEFAULT_DEEPGRAM_MODEL, description="Deepgram model."),
|
||||
whisper_model: str = Form(DEFAULT_WHISPER_MODEL, description="OpenAI přepisový model."),
|
||||
chat_model: str = Form(DEFAULT_CHAT_MODEL, description="OpenAI chat model pro sloučení."),
|
||||
language: str = Form(DEFAULT_LANGUAGE, description="Jazyk audia (ISO kód, např. cs)."),
|
||||
diarize: bool = Form(True, description="Deepgram diarizace (rozlišení mluvčích)."),
|
||||
smart_format: bool = Form(True, description="Deepgram smart formatting."),
|
||||
creds: Credentials = Depends(get_credentials),
|
||||
) -> CombinedResult:
|
||||
audio, filename, content_type = await _read_upload(file)
|
||||
text1, text2 = await _transcribe_dual(
|
||||
creds, audio, filename, content_type,
|
||||
deepgram_model, whisper_model, language, diarize, smart_format,
|
||||
)
|
||||
|
||||
prompt = combine_prompt.strip() if combine_prompt and combine_prompt.strip() else DEFAULT_COMBINE_PROMPT
|
||||
merged = await openai_client.merge_transcripts(
|
||||
api_key=creds.require_openai(),
|
||||
api_key=openai_key,
|
||||
model=chat_model,
|
||||
system_prompt=prompt,
|
||||
text_deepgram=text1,
|
||||
text_whisper=text2,
|
||||
text_deepgram=text_deepgram,
|
||||
text_whisper=text_gpt,
|
||||
)
|
||||
|
||||
return CombinedResult(
|
||||
text1=text1,
|
||||
text2=text2,
|
||||
merged=merged,
|
||||
deepgram_model=deepgram_model,
|
||||
whisper_model=whisper_model,
|
||||
chat_model=chat_model,
|
||||
language=language,
|
||||
)
|
||||
return TranscribeResult(gpt=text_gpt, deepgram=text_deepgram, merge=merged)
|
||||
|
||||
Reference in New Issue
Block a user