Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,83 +1,371 @@
|
|
1 |
-
import os
|
2 |
-
import time
|
3 |
-
import torch
|
4 |
import gradio as gr
|
5 |
-
|
6 |
-
from transformers import AutoTokenizer
|
7 |
-
from auto_gptq import AutoGPTQForCausalLM
|
8 |
from sentence_transformers import SentenceTransformer
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
login(token=hf_token)
|
17 |
-
|
18 |
-
# Load tokenizer and quantized model
|
19 |
-
model_id = "TheBloke/mistral-7B-GPTQ"
|
20 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
21 |
-
|
22 |
-
print("Loading tokenizer...")
|
23 |
-
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
|
24 |
-
|
25 |
-
print("Loading quantized model...")
|
26 |
-
start = time.time()
|
27 |
-
model = AutoGPTQForCausalLM.from_quantized(
|
28 |
-
model_id,
|
29 |
-
use_safetensors=True,
|
30 |
-
device=device,
|
31 |
-
use_triton=True,
|
32 |
-
quantize_config=None,
|
33 |
-
)
|
34 |
-
print(f"Model loaded in {time.time() - start:.2f} seconds on {device}")
|
35 |
-
|
36 |
-
# Load embedding model for FAISS vector store
|
37 |
-
embedder = SentenceTransformer("all-MiniLM-L6-v2")
|
38 |
-
|
39 |
-
# Sample documents to build vector index (can replace with your own)
|
40 |
-
texts = [
|
41 |
-
"Hello world",
|
42 |
-
"Mistral 7B is a powerful language model",
|
43 |
-
"Langchain and FAISS make vector search easy",
|
44 |
-
"This is a test document for vector search",
|
45 |
-
]
|
46 |
-
embeddings = embedder.encode(texts)
|
47 |
|
48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
|
58 |
-
|
59 |
-
|
60 |
-
query_emb = embedder.encode([query])
|
61 |
-
results = faiss_index.similarity_search_by_vector(query_emb[0], k=3)
|
62 |
-
return "\n\n".join(results)
|
63 |
|
64 |
-
|
65 |
-
with gr.Blocks() as demo:
|
66 |
-
gr.Markdown("# Mistral GPTQ + FAISS Vector Search Demo")
|
67 |
|
68 |
-
|
69 |
-
prompt_input = gr.Textbox(label="Enter prompt", lines=3)
|
70 |
-
generate_btn = gr.Button("Generate")
|
71 |
-
output_text = gr.Textbox(label="Output", lines=6)
|
72 |
|
73 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
74 |
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
search_output = gr.Textbox(label="Search Results", lines=6)
|
79 |
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
|
|
82 |
if __name__ == "__main__":
|
83 |
-
demo
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
+
import torch
|
3 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
|
|
|
4 |
from sentence_transformers import SentenceTransformer
|
5 |
+
import faiss
|
6 |
+
import numpy as np
|
7 |
+
import PyPDF2
|
8 |
+
import docx
|
9 |
+
import io
|
10 |
+
import os
|
11 |
+
from typing import List, Optional
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
+
class DocumentRAG:
|
14 |
+
def __init__(self):
|
15 |
+
print("π Initializing RAG System...")
|
16 |
+
|
17 |
+
# Initialize embedding model (lightweight)
|
18 |
+
self.embedder = SentenceTransformer('all-MiniLM-L6-v2')
|
19 |
+
print("β
Embedding model loaded")
|
20 |
+
|
21 |
+
# Initialize quantized LLM
|
22 |
+
self.setup_llm()
|
23 |
+
|
24 |
+
# Document storage
|
25 |
+
self.documents = []
|
26 |
+
self.index = None
|
27 |
+
self.is_indexed = False
|
28 |
+
|
29 |
+
def setup_llm(self):
|
30 |
+
"""Setup quantized Mistral model"""
|
31 |
+
try:
|
32 |
+
quantization_config = BitsAndBytesConfig(
|
33 |
+
load_in_4bit=True,
|
34 |
+
bnb_4bit_compute_dtype=torch.float16,
|
35 |
+
bnb_4bit_use_double_quant=True,
|
36 |
+
bnb_4bit_quant_type="nf4"
|
37 |
+
)
|
38 |
+
|
39 |
+
model_name = "mistralai/Mistral-7B-Instruct-v0.1"
|
40 |
+
|
41 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
42 |
+
if self.tokenizer.pad_token is None:
|
43 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
44 |
+
|
45 |
+
self.model = AutoModelForCausalLM.from_pretrained(
|
46 |
+
model_name,
|
47 |
+
quantization_config=quantization_config,
|
48 |
+
device_map="auto",
|
49 |
+
torch_dtype=torch.float16,
|
50 |
+
trust_remote_code=True
|
51 |
+
)
|
52 |
+
print("β
Quantized Mistral model loaded")
|
53 |
+
|
54 |
+
except Exception as e:
|
55 |
+
print(f"β Error loading model: {e}")
|
56 |
+
# Fallback to a smaller model if Mistral fails
|
57 |
+
self.setup_fallback_model()
|
58 |
+
|
59 |
+
def setup_fallback_model(self):
|
60 |
+
"""Fallback to smaller model if Mistral fails"""
|
61 |
+
try:
|
62 |
+
model_name = "microsoft/DialoGPT-small"
|
63 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
64 |
+
self.model = AutoModelForCausalLM.from_pretrained(model_name)
|
65 |
+
print("β
Fallback model loaded")
|
66 |
+
except Exception as e:
|
67 |
+
print(f"β Fallback model failed: {e}")
|
68 |
+
self.model = None
|
69 |
+
self.tokenizer = None
|
70 |
|
71 |
+
def extract_text_from_file(self, file_path: str) -> str:
|
72 |
+
"""Extract text from various file formats"""
|
73 |
+
try:
|
74 |
+
file_extension = os.path.splitext(file_path)[1].lower()
|
75 |
+
|
76 |
+
if file_extension == '.pdf':
|
77 |
+
return self.extract_from_pdf(file_path)
|
78 |
+
elif file_extension == '.docx':
|
79 |
+
return self.extract_from_docx(file_path)
|
80 |
+
elif file_extension == '.txt':
|
81 |
+
return self.extract_from_txt(file_path)
|
82 |
+
else:
|
83 |
+
return f"Unsupported file format: {file_extension}"
|
84 |
+
|
85 |
+
except Exception as e:
|
86 |
+
return f"Error reading file: {str(e)}"
|
87 |
+
|
88 |
+
def extract_from_pdf(self, file_path: str) -> str:
|
89 |
+
"""Extract text from PDF"""
|
90 |
+
text = ""
|
91 |
+
try:
|
92 |
+
with open(file_path, 'rb') as file:
|
93 |
+
pdf_reader = PyPDF2.PdfReader(file)
|
94 |
+
for page in pdf_reader.pages:
|
95 |
+
text += page.extract_text() + "\n"
|
96 |
+
except Exception as e:
|
97 |
+
text = f"Error reading PDF: {str(e)}"
|
98 |
+
return text
|
99 |
+
|
100 |
+
def extract_from_docx(self, file_path: str) -> str:
|
101 |
+
"""Extract text from DOCX"""
|
102 |
+
try:
|
103 |
+
doc = docx.Document(file_path)
|
104 |
+
text = ""
|
105 |
+
for paragraph in doc.paragraphs:
|
106 |
+
text += paragraph.text + "\n"
|
107 |
+
return text
|
108 |
+
except Exception as e:
|
109 |
+
return f"Error reading DOCX: {str(e)}"
|
110 |
+
|
111 |
+
def extract_from_txt(self, file_path: str) -> str:
|
112 |
+
"""Extract text from TXT"""
|
113 |
+
try:
|
114 |
+
with open(file_path, 'r', encoding='utf-8') as file:
|
115 |
+
return file.read()
|
116 |
+
except Exception as e:
|
117 |
+
try:
|
118 |
+
with open(file_path, 'r', encoding='latin-1') as file:
|
119 |
+
return file.read()
|
120 |
+
except Exception as e2:
|
121 |
+
return f"Error reading TXT: {str(e2)}"
|
122 |
+
|
123 |
+
def chunk_text(self, text: str, chunk_size: int = 500, overlap: int = 50) -> List[str]:
|
124 |
+
"""Split text into overlapping chunks"""
|
125 |
+
if not text.strip():
|
126 |
+
return []
|
127 |
+
|
128 |
+
words = text.split()
|
129 |
+
chunks = []
|
130 |
+
|
131 |
+
for i in range(0, len(words), chunk_size - overlap):
|
132 |
+
chunk = ' '.join(words[i:i + chunk_size])
|
133 |
+
if chunk.strip():
|
134 |
+
chunks.append(chunk.strip())
|
135 |
+
|
136 |
+
if i + chunk_size >= len(words):
|
137 |
+
break
|
138 |
+
|
139 |
+
return chunks
|
140 |
+
|
141 |
+
def process_documents(self, files) -> str:
|
142 |
+
"""Process uploaded files and create embeddings"""
|
143 |
+
if not files:
|
144 |
+
return "β No files uploaded!"
|
145 |
+
|
146 |
+
try:
|
147 |
+
all_text = ""
|
148 |
+
processed_files = []
|
149 |
+
|
150 |
+
# Extract text from all files
|
151 |
+
for file in files:
|
152 |
+
if file is None:
|
153 |
+
continue
|
154 |
+
|
155 |
+
file_text = self.extract_text_from_file(file.name)
|
156 |
+
if not file_text.startswith("Error") and not file_text.startswith("Unsupported"):
|
157 |
+
all_text += f"\n\n--- {os.path.basename(file.name)} ---\n\n{file_text}"
|
158 |
+
processed_files.append(os.path.basename(file.name))
|
159 |
+
else:
|
160 |
+
return f"β {file_text}"
|
161 |
+
|
162 |
+
if not all_text.strip():
|
163 |
+
return "β No text extracted from files!"
|
164 |
+
|
165 |
+
# Chunk the text
|
166 |
+
self.documents = self.chunk_text(all_text)
|
167 |
+
|
168 |
+
if not self.documents:
|
169 |
+
return "β No valid text chunks created!"
|
170 |
+
|
171 |
+
# Create embeddings
|
172 |
+
print(f"π Creating embeddings for {len(self.documents)} chunks...")
|
173 |
+
embeddings = self.embedder.encode(self.documents, show_progress_bar=True)
|
174 |
+
|
175 |
+
# Build FAISS index
|
176 |
+
dimension = embeddings.shape[1]
|
177 |
+
self.index = faiss.IndexFlatIP(dimension)
|
178 |
+
|
179 |
+
# Normalize embeddings for cosine similarity
|
180 |
+
faiss.normalize_L2(embeddings)
|
181 |
+
self.index.add(embeddings.astype('float32'))
|
182 |
+
|
183 |
+
self.is_indexed = True
|
184 |
+
|
185 |
+
return f"β
Successfully processed {len(processed_files)} files:\n" + \
|
186 |
+
f"π Files: {', '.join(processed_files)}\n" + \
|
187 |
+
f"π Created {len(self.documents)} text chunks\n" + \
|
188 |
+
f"π Ready for Q&A!"
|
189 |
+
|
190 |
+
except Exception as e:
|
191 |
+
return f"β Error processing documents: {str(e)}"
|
192 |
+
|
193 |
+
def retrieve_context(self, query: str, k: int = 3) -> str:
|
194 |
+
"""Retrieve relevant context for the query"""
|
195 |
+
if not self.is_indexed:
|
196 |
+
return ""
|
197 |
+
|
198 |
+
try:
|
199 |
+
# Get query embedding
|
200 |
+
query_embedding = self.embedder.encode([query])
|
201 |
+
faiss.normalize_L2(query_embedding)
|
202 |
+
|
203 |
+
# Search for similar chunks
|
204 |
+
scores, indices = self.index.search(query_embedding.astype('float32'), k)
|
205 |
+
|
206 |
+
# Get relevant documents
|
207 |
+
relevant_docs = []
|
208 |
+
for i, idx in enumerate(indices[0]):
|
209 |
+
if idx < len(self.documents) and scores[0][i] > 0.1: # Similarity threshold
|
210 |
+
relevant_docs.append(self.documents[idx])
|
211 |
+
|
212 |
+
return "\n\n".join(relevant_docs)
|
213 |
+
|
214 |
+
except Exception as e:
|
215 |
+
print(f"Error in retrieval: {e}")
|
216 |
+
return ""
|
217 |
+
|
218 |
+
def generate_answer(self, query: str, context: str) -> str:
|
219 |
+
"""Generate answer using the LLM"""
|
220 |
+
if self.model is None or self.tokenizer is None:
|
221 |
+
return "β Model not available. Please try again."
|
222 |
+
|
223 |
+
try:
|
224 |
+
# Create prompt
|
225 |
+
prompt = f"""<s>[INST] Based on the following context, answer the question. If the answer is not in the context, say "I don't have enough information to answer this question."
|
226 |
|
227 |
+
Context:
|
228 |
+
{context[:2000]} # Limit context length
|
|
|
|
|
|
|
229 |
|
230 |
+
Question: {query}
|
|
|
|
|
231 |
|
232 |
+
Answer: [/INST]"""
|
|
|
|
|
|
|
233 |
|
234 |
+
# Tokenize
|
235 |
+
inputs = self.tokenizer(
|
236 |
+
prompt,
|
237 |
+
return_tensors="pt",
|
238 |
+
max_length=1024,
|
239 |
+
truncation=True,
|
240 |
+
padding=True
|
241 |
+
)
|
242 |
+
|
243 |
+
# Generate
|
244 |
+
with torch.no_grad():
|
245 |
+
outputs = self.model.generate(
|
246 |
+
**inputs,
|
247 |
+
max_new_tokens=256,
|
248 |
+
temperature=0.7,
|
249 |
+
do_sample=True,
|
250 |
+
top_p=0.9,
|
251 |
+
pad_token_id=self.tokenizer.eos_token_id,
|
252 |
+
eos_token_id=self.tokenizer.eos_token_id
|
253 |
+
)
|
254 |
+
|
255 |
+
# Decode response
|
256 |
+
full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
257 |
+
|
258 |
+
# Extract answer (remove the prompt part)
|
259 |
+
if "[/INST]" in full_response:
|
260 |
+
answer = full_response.split("[/INST]")[-1].strip()
|
261 |
+
else:
|
262 |
+
answer = full_response[len(prompt):].strip()
|
263 |
+
|
264 |
+
return answer if answer else "I couldn't generate a proper response."
|
265 |
+
|
266 |
+
except Exception as e:
|
267 |
+
return f"β Error generating answer: {str(e)}"
|
268 |
+
|
269 |
+
def answer_question(self, query: str) -> str:
|
270 |
+
"""Main function to answer questions"""
|
271 |
+
if not query.strip():
|
272 |
+
return "β Please ask a question!"
|
273 |
+
|
274 |
+
if not self.is_indexed:
|
275 |
+
return "π Please upload and process documents first!"
|
276 |
+
|
277 |
+
try:
|
278 |
+
# Retrieve relevant context
|
279 |
+
context = self.retrieve_context(query)
|
280 |
+
|
281 |
+
if not context:
|
282 |
+
return "π No relevant information found in the uploaded documents."
|
283 |
+
|
284 |
+
# Generate answer
|
285 |
+
answer = self.generate_answer(query, context)
|
286 |
+
|
287 |
+
return f"π‘ **Answer:** {answer}\n\nπ **Source Context:** {context[:500]}..."
|
288 |
+
|
289 |
+
except Exception as e:
|
290 |
+
return f"β Error answering question: {str(e)}"
|
291 |
|
292 |
+
# Initialize the RAG system
|
293 |
+
print("Initializing Document RAG System...")
|
294 |
+
rag_system = DocumentRAG()
|
|
|
295 |
|
296 |
+
# Gradio Interface
|
297 |
+
def create_interface():
|
298 |
+
with gr.Blocks(title="π Document Q&A with RAG", theme=gr.themes.Soft()) as demo:
|
299 |
+
gr.Markdown("""
|
300 |
+
# π Document Q&A System
|
301 |
+
|
302 |
+
Upload your documents and ask questions about them!
|
303 |
+
|
304 |
+
**Supported formats:** PDF, DOCX, TXT
|
305 |
+
""")
|
306 |
+
|
307 |
+
with gr.Tab("π€ Upload Documents"):
|
308 |
+
with gr.Row():
|
309 |
+
with gr.Column():
|
310 |
+
file_upload = gr.File(
|
311 |
+
label="Upload Documents",
|
312 |
+
file_count="multiple",
|
313 |
+
file_types=[".pdf", ".docx", ".txt"]
|
314 |
+
)
|
315 |
+
process_btn = gr.Button("π Process Documents", variant="primary")
|
316 |
+
|
317 |
+
with gr.Column():
|
318 |
+
process_status = gr.Textbox(
|
319 |
+
label="Processing Status",
|
320 |
+
lines=8,
|
321 |
+
interactive=False
|
322 |
+
)
|
323 |
+
|
324 |
+
process_btn.click(
|
325 |
+
fn=rag_system.process_documents,
|
326 |
+
inputs=[file_upload],
|
327 |
+
outputs=[process_status]
|
328 |
+
)
|
329 |
+
|
330 |
+
with gr.Tab("β Ask Questions"):
|
331 |
+
with gr.Row():
|
332 |
+
with gr.Column():
|
333 |
+
question_input = gr.Textbox(
|
334 |
+
label="Your Question",
|
335 |
+
placeholder="What would you like to know about your documents?",
|
336 |
+
lines=3
|
337 |
+
)
|
338 |
+
ask_btn = gr.Button("π Get Answer", variant="primary")
|
339 |
+
|
340 |
+
with gr.Column():
|
341 |
+
answer_output = gr.Textbox(
|
342 |
+
label="Answer",
|
343 |
+
lines=10,
|
344 |
+
interactive=False
|
345 |
+
)
|
346 |
+
|
347 |
+
ask_btn.click(
|
348 |
+
fn=rag_system.answer_question,
|
349 |
+
inputs=[question_input],
|
350 |
+
outputs=[answer_output]
|
351 |
+
)
|
352 |
+
|
353 |
+
# Example questions
|
354 |
+
gr.Markdown("""
|
355 |
+
### π‘ Example Questions:
|
356 |
+
- What is the main topic of the document?
|
357 |
+
- Can you summarize the key points?
|
358 |
+
- What are the conclusions mentioned?
|
359 |
+
- Are there any specific numbers or statistics?
|
360 |
+
""")
|
361 |
+
|
362 |
+
return demo
|
363 |
|
364 |
+
# Launch the app
|
365 |
if __name__ == "__main__":
|
366 |
+
demo = create_interface()
|
367 |
+
demo.launch(
|
368 |
+
server_name="0.0.0.0",
|
369 |
+
server_port=7860,
|
370 |
+
share=True
|
371 |
+
)
|