Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,116 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gradio as gr
|
3 |
+
import whisper
|
4 |
+
from gtts import gTTS
|
5 |
+
import io
|
6 |
+
from groq import Groq
|
7 |
+
from PyPDF2 import PdfReader
|
8 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
9 |
+
import torch
|
10 |
+
|
11 |
+
# Set up environment variables
|
12 |
+
os.environ["GROQ_API_KEY"] = "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"
|
13 |
+
|
14 |
+
# Initialize the Groq client
|
15 |
+
client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
|
16 |
+
|
17 |
+
# Load the Whisper model
|
18 |
+
whisper_model = whisper.load_model("base") # You can choose other models like "small", "medium", "large"
|
19 |
+
|
20 |
+
# Initialize the tokenizer and model from the saved checkpoint for RAG
|
21 |
+
rag_tokenizer = AutoTokenizer.from_pretrained("himmeow/vi-gemma-2b-RAG")
|
22 |
+
rag_model = AutoModelForCausalLM.from_pretrained(
|
23 |
+
"himmeow/vi-gemma-2b-RAG",
|
24 |
+
device_map="auto",
|
25 |
+
torch_dtype=torch.bfloat16
|
26 |
+
)
|
27 |
+
|
28 |
+
# Use GPU if available for RAG model
|
29 |
+
if torch.cuda.is_available():
|
30 |
+
rag_model.to("cuda")
|
31 |
+
|
32 |
+
# Load PDF content
|
33 |
+
def load_pdf(pdf_path):
|
34 |
+
pdf_text = ""
|
35 |
+
with open(pdf_path, "rb") as file:
|
36 |
+
reader = PdfReader(file)
|
37 |
+
for page_num in range(len(reader.pages)):
|
38 |
+
page = reader.pages[page_num]
|
39 |
+
text = page.extract_text()
|
40 |
+
pdf_text += text + "\n"
|
41 |
+
return pdf_text
|
42 |
+
|
43 |
+
# Define the prompt format for the RAG model
|
44 |
+
prompt_template = """
|
45 |
+
### Instruction and Input:
|
46 |
+
Based on the following context/document:
|
47 |
+
{}
|
48 |
+
Please answer the question: {}
|
49 |
+
|
50 |
+
### Response:
|
51 |
+
{}
|
52 |
+
"""
|
53 |
+
|
54 |
+
# Function to process audio and generate a response using RAG and Groq
|
55 |
+
def process_audio_rag(file_path):
|
56 |
+
try:
|
57 |
+
# Load and transcribe the audio using Whisper
|
58 |
+
audio = whisper.load_audio(file_path)
|
59 |
+
result = whisper_model.transcribe(audio)
|
60 |
+
text = result["text"]
|
61 |
+
|
62 |
+
# Load the PDF content (update with your PDF path or pass it as an argument)
|
63 |
+
pdf_path = "/content/BN_Cotton.pdf"
|
64 |
+
pdf_text = load_pdf(pdf_path)
|
65 |
+
|
66 |
+
# Prepare the input data for the RAG model
|
67 |
+
query = text
|
68 |
+
input_text = prompt_template.format(pdf_text, query, " ")
|
69 |
+
|
70 |
+
# Encode the input text into input ids for RAG model
|
71 |
+
input_ids = rag_tokenizer(input_text, return_tensors="pt")
|
72 |
+
if torch.cuda.is_available():
|
73 |
+
input_ids = input_ids.to("cuda")
|
74 |
+
|
75 |
+
# Generate text using the RAG model
|
76 |
+
outputs = rag_model.generate(
|
77 |
+
**input_ids,
|
78 |
+
max_new_tokens=500,
|
79 |
+
no_repeat_ngram_size=5
|
80 |
+
)
|
81 |
+
rag_response = rag_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
82 |
+
|
83 |
+
# Generate a response using Groq if needed
|
84 |
+
chat_completion = client.chat.completions.create(
|
85 |
+
messages=[{"role": "user", "content": rag_response}],
|
86 |
+
model="llama3-8b-8192", # Replace with the correct model if necessary
|
87 |
+
)
|
88 |
+
response_message = chat_completion.choices[0].message.content.strip()
|
89 |
+
|
90 |
+
# Convert the response text to speech
|
91 |
+
tts = gTTS(response_message)
|
92 |
+
response_audio_io = io.BytesIO()
|
93 |
+
tts.write_to_fp(response_audio_io)
|
94 |
+
response_audio_io.seek(0)
|
95 |
+
|
96 |
+
# Save audio to a file to ensure it's generated correctly
|
97 |
+
with open("response.mp3", "wb") as audio_file:
|
98 |
+
audio_file.write(response_audio_io.getvalue())
|
99 |
+
|
100 |
+
# Return the response text and the path to the saved audio file
|
101 |
+
return response_message, "response.mp3"
|
102 |
+
|
103 |
+
except Exception as e:
|
104 |
+
return f"An error occurred: {e}", None
|
105 |
+
|
106 |
+
# Create a Gradio interface
|
107 |
+
iface = gr.Interface(
|
108 |
+
fn=process_audio_rag,
|
109 |
+
inputs=gr.Audio(type="filepath"),
|
110 |
+
outputs=[gr.Textbox(label="Response Text"), gr.Audio(label="Response Audio")],
|
111 |
+
live=True,
|
112 |
+
title="Agriculture Assistant"
|
113 |
+
)
|
114 |
+
|
115 |
+
# Launch the interface with the given title
|
116 |
+
iface.launch()
|