quivr/backend/parsers/audio.py
2023-06-28 19:39:27 +02:00

57 lines
2.0 KiB
Python

import os
import tempfile
import time
import openai
from langchain.schema import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
from models.files import File
from models.settings import CommonsDep
from utils.file import compute_sha1_from_content
async def process_audio(commons: CommonsDep, file: File, enable_summarization: bool, user, user_openai_api_key):
temp_filename = None
file_sha = ""
dateshort = time.strftime("%Y%m%d-%H%M%S")
file_meta_name = f"audiotranscript_{dateshort}.txt"
# use this for whisper
openai_api_key = os.environ.get("OPENAI_API_KEY")
if user_openai_api_key:
openai_api_key = user_openai_api_key
try:
upload_file = file.file
with tempfile.NamedTemporaryFile(delete=False, suffix=upload_file.filename) as tmp_file:
await upload_file.seek(0)
content = await upload_file.read()
tmp_file.write(content)
tmp_file.flush()
tmp_file.close()
temp_filename = tmp_file.name
with open(tmp_file.name, "rb") as audio_file:
transcript = openai.Audio.transcribe("whisper-1", audio_file)
file_sha = compute_sha1_from_content(transcript.text.encode("utf-8"))
file_size = len(transcript.text.encode("utf-8"))
chunk_size = 500
chunk_overlap = 0
text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(
chunk_size=chunk_size, chunk_overlap=chunk_overlap)
texts = text_splitter.split_text(transcript.text.encode("utf-8"))
docs_with_metadata = [Document(page_content=text, metadata={"file_sha1": file_sha, "file_size": file_size, "file_name": file_meta_name,
"chunk_size": chunk_size, "chunk_overlap": chunk_overlap, "date": dateshort}) for text in texts]
commons.documents_vector_store.add_documents(docs_with_metadata)
finally:
if temp_filename and os.path.exists(temp_filename):
os.remove(temp_filename)