Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -9,43 +9,42 @@ import os
|
|
9 |
import subprocess
|
10 |
import numpy as np
|
11 |
from typing import List, Dict, Tuple, Any
|
12 |
-
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
try:
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
print("Warning: Could not install espeak or espeak-ng. TTS functionality may be limited.")
|
29 |
-
|
30 |
except Exception as e:
|
31 |
-
print(f"
|
32 |
-
|
33 |
-
|
34 |
-
# ---
|
35 |
-
|
36 |
-
model_name = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
|
37 |
-
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
38 |
-
tokenizer.pad_token = tokenizer.eos_token
|
39 |
-
# Initialize DeepSeek model
|
40 |
-
model = AutoModelForCausalLM.from_pretrained(
|
41 |
-
model_name,
|
42 |
-
device_map="auto",
|
43 |
-
offload_folder="offload",
|
44 |
-
low_cpu_mem_usage=True,
|
45 |
-
torch_dtype=torch.float16
|
46 |
-
)
|
47 |
-
|
48 |
-
# Initialize Kokoro TTS (with error handling)
|
49 |
VOICE_CHOICES = {
|
50 |
'🇺🇸 Female (Default)': 'af',
|
51 |
'🇺🇸 Bella': 'af_bella',
|
@@ -54,41 +53,81 @@ VOICE_CHOICES = {
|
|
54 |
}
|
55 |
TTS_ENABLED = False
|
56 |
TTS_MODEL = None
|
57 |
-
|
58 |
-
|
59 |
-
try:
|
60 |
-
if os.path.exists('Kokoro-82M'):
|
61 |
-
import sys
|
62 |
-
sys.path.append('Kokoro-82M')
|
63 |
-
from models import build_model # type: ignore
|
64 |
-
from kokoro import generate # type: ignore
|
65 |
-
|
66 |
-
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
67 |
-
TTS_MODEL = build_model('Kokoro-82M/kokoro-v0_19.pth', device)
|
68 |
|
69 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
try:
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
82 |
|
83 |
-
|
84 |
-
|
|
|
|
|
|
|
|
|
|
|
85 |
try:
|
86 |
with DDGS() as ddgs:
|
87 |
results = list(ddgs.text(query, max_results=max_results))
|
88 |
return [{
|
89 |
"title": result.get("title", ""),
|
90 |
-
"snippet": result
|
91 |
-
"url": result
|
92 |
"date": result.get("published", "")
|
93 |
} for result in results]
|
94 |
except Exception as e:
|
@@ -116,23 +155,24 @@ def format_sources(web_results: List[Dict[str, str]]) -> str:
|
|
116 |
sources_html = "<div class='sources-container'>"
|
117 |
for i, res in enumerate(web_results, 1):
|
118 |
title = res["title"] or "Source"
|
119 |
-
date = f"<span class='source-date'>{res['date']}</span>" if res
|
|
|
120 |
sources_html += f"""
|
121 |
<div class='source-item'>
|
122 |
<div class='source-number'>[{i}]</div>
|
123 |
<div class='source-content'>
|
124 |
<a href="{res['url']}" target="_blank" class='source-title'>{title}</a>
|
125 |
{date}
|
126 |
-
<div class='source-snippet'>{
|
127 |
</div>
|
128 |
</div>
|
129 |
"""
|
130 |
sources_html += "</div>"
|
131 |
return sources_html
|
132 |
|
133 |
-
@spaces.GPU(duration=
|
134 |
def generate_answer(prompt: str) -> str:
|
135 |
-
"""Generate answer using the DeepSeek model"""
|
136 |
inputs = tokenizer(
|
137 |
prompt,
|
138 |
return_tensors="pt",
|
@@ -142,52 +182,56 @@ def generate_answer(prompt: str) -> str:
|
|
142 |
return_attention_mask=True
|
143 |
).to(model.device)
|
144 |
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
|
|
155 |
return tokenizer.decode(outputs[0], skip_special_tokens=True)
|
156 |
|
157 |
-
@spaces.GPU(duration=
|
158 |
-
def
|
159 |
-
"""Generate speech from text using Kokoro TTS model."""
|
160 |
-
|
161 |
-
|
|
|
162 |
return None
|
163 |
|
164 |
try:
|
|
|
165 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
166 |
|
167 |
-
#
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
|
|
|
|
181 |
else:
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
# Clean the text
|
186 |
clean_text = ' '.join([line for line in text.split('\n') if not line.startswith('#')])
|
187 |
clean_text = clean_text.replace('[', '').replace(']', '').replace('*', '')
|
188 |
|
189 |
# Split long text into chunks
|
190 |
-
max_chars =
|
191 |
chunks = []
|
192 |
if len(clean_text) > max_chars:
|
193 |
sentences = clean_text.split('.')
|
@@ -207,7 +251,7 @@ def generate_speech_with_gpu(text: str, voice_name: str = 'af', tts_model=TTS_MO
|
|
207 |
audio_chunks = []
|
208 |
for chunk in chunks:
|
209 |
if chunk.strip():
|
210 |
-
chunk_audio, _ =
|
211 |
if isinstance(chunk_audio, torch.Tensor):
|
212 |
chunk_audio = chunk_audio.cpu().numpy()
|
213 |
audio_chunks.append(chunk_audio)
|
@@ -215,35 +259,61 @@ def generate_speech_with_gpu(text: str, voice_name: str = 'af', tts_model=TTS_MO
|
|
215 |
# Concatenate chunks
|
216 |
if audio_chunks:
|
217 |
final_audio = np.concatenate(audio_chunks) if len(audio_chunks) > 1 else audio_chunks[0]
|
218 |
-
return (
|
219 |
-
|
220 |
-
|
221 |
|
222 |
except Exception as e:
|
223 |
print(f"Error generating speech: {str(e)}")
|
224 |
return None
|
225 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
226 |
def process_query(query: str, history: List[List[str]], selected_voice: str = 'af'):
|
227 |
-
"""Process user query with streaming effect"""
|
228 |
try:
|
229 |
if history is None:
|
230 |
history = []
|
231 |
|
232 |
-
#
|
233 |
-
web_results = get_web_results(query)
|
234 |
-
sources_html = format_sources(web_results)
|
235 |
-
|
236 |
current_history = history + [[query, "*Searching...*"]]
|
237 |
-
|
238 |
# Yield initial searching state
|
239 |
yield (
|
240 |
"*Searching & Thinking...*", # answer_output (Markdown)
|
241 |
-
|
242 |
"Searching...", # search_btn (Button)
|
243 |
current_history, # chat_history_display (Chatbot)
|
244 |
None # audio_output (Audio)
|
245 |
)
|
246 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
247 |
# Generate answer
|
248 |
prompt = format_prompt(query, web_results)
|
249 |
answer = generate_answer(prompt)
|
@@ -251,26 +321,27 @@ def process_query(query: str, history: List[List[str]], selected_voice: str = 'a
|
|
251 |
|
252 |
# Update history before TTS
|
253 |
updated_history = history + [[query, final_answer]]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
|
255 |
-
# Generate speech
|
256 |
-
|
257 |
-
|
258 |
-
final_answer, # answer_output
|
259 |
-
sources_html, # sources_output
|
260 |
-
"Generating audio...", # search_btn
|
261 |
-
updated_history, # chat_history_display
|
262 |
-
None # audio_output
|
263 |
-
)
|
264 |
try:
|
265 |
-
audio =
|
266 |
if audio is None:
|
267 |
final_answer += "\n\n*Audio generation failed. The voicepack may be missing or incompatible.*"
|
268 |
except Exception as e:
|
269 |
final_answer += f"\n\n*Error generating audio: {str(e)}*"
|
270 |
-
audio = None
|
271 |
else:
|
272 |
-
final_answer += "\n\n*TTS is disabled.
|
273 |
-
audio = None
|
274 |
|
275 |
# Yield final result
|
276 |
yield (
|
@@ -278,7 +349,7 @@ def process_query(query: str, history: List[List[str]], selected_voice: str = 'a
|
|
278 |
sources_html, # sources_output
|
279 |
"Search", # search_btn
|
280 |
updated_history, # chat_history_display
|
281 |
-
audio
|
282 |
)
|
283 |
|
284 |
except Exception as e:
|
@@ -287,13 +358,13 @@ def process_query(query: str, history: List[List[str]], selected_voice: str = 'a
|
|
287 |
error_message = "⚠️ GPU quota exceeded. Please try again later when the daily quota resets."
|
288 |
yield (
|
289 |
f"Error: {error_message}", # answer_output
|
290 |
-
|
291 |
"Search", # search_btn
|
292 |
history + [[query, f"*Error: {error_message}*"]], # chat_history_display
|
293 |
None # audio_output
|
294 |
)
|
295 |
|
296 |
-
#
|
297 |
css = """
|
298 |
.gradio-container {
|
299 |
max-width: 1200px !important;
|
@@ -303,36 +374,44 @@ css = """
|
|
303 |
text-align: center;
|
304 |
margin-bottom: 2rem;
|
305 |
padding: 2rem 0;
|
306 |
-
background: #1a1b1e;
|
307 |
border-radius: 12px;
|
308 |
color: white;
|
|
|
309 |
}
|
310 |
#header h1 {
|
311 |
color: white;
|
312 |
font-size: 2.5rem;
|
313 |
margin-bottom: 0.5rem;
|
|
|
314 |
}
|
315 |
#header h3 {
|
316 |
color: #a8a9ab;
|
317 |
}
|
318 |
.search-container {
|
319 |
-
background: #1a1b1e;
|
320 |
border-radius: 12px;
|
321 |
-
box-shadow: 0 4px
|
322 |
-
padding:
|
323 |
-
margin-bottom:
|
324 |
}
|
325 |
.search-box {
|
326 |
padding: 1rem;
|
327 |
background: #2c2d30;
|
328 |
-
border-radius:
|
329 |
margin-bottom: 1rem;
|
|
|
330 |
}
|
331 |
.search-box input[type="text"] {
|
332 |
background: #3a3b3e !important;
|
333 |
border: 1px solid #4a4b4e !important;
|
334 |
color: white !important;
|
335 |
border-radius: 8px !important;
|
|
|
|
|
|
|
|
|
|
|
336 |
}
|
337 |
.search-box input[type="text"]::placeholder {
|
338 |
color: #a8a9ab !important;
|
@@ -340,23 +419,43 @@ css = """
|
|
340 |
.search-box button {
|
341 |
background: #2563eb !important;
|
342 |
border: none !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
343 |
}
|
344 |
.results-container {
|
345 |
background: #2c2d30;
|
346 |
-
border-radius:
|
347 |
-
padding:
|
348 |
-
margin-top:
|
|
|
349 |
}
|
350 |
.answer-box {
|
351 |
background: #3a3b3e;
|
352 |
-
border-radius:
|
353 |
padding: 1.5rem;
|
354 |
color: white;
|
355 |
-
margin-bottom:
|
|
|
|
|
|
|
|
|
|
|
356 |
}
|
357 |
.answer-box p {
|
358 |
color: #e5e7eb;
|
359 |
-
line-height: 1.
|
|
|
|
|
|
|
|
|
|
|
360 |
}
|
361 |
.sources-container {
|
362 |
margin-top: 1rem;
|
@@ -367,13 +466,16 @@ css = """
|
|
367 |
.source-item {
|
368 |
display: flex;
|
369 |
padding: 12px;
|
370 |
-
margin:
|
371 |
background: #3a3b3e;
|
372 |
border-radius: 8px;
|
373 |
transition: all 0.2s;
|
|
|
374 |
}
|
375 |
.source-item:hover {
|
376 |
background: #4a4b4e;
|
|
|
|
|
377 |
}
|
378 |
.source-number {
|
379 |
font-weight: bold;
|
@@ -388,7 +490,12 @@ css = """
|
|
388 |
font-weight: 500;
|
389 |
text-decoration: none;
|
390 |
display: block;
|
391 |
-
margin-bottom:
|
|
|
|
|
|
|
|
|
|
|
392 |
}
|
393 |
.source-date {
|
394 |
color: #a8a9ab;
|
@@ -398,7 +505,7 @@ css = """
|
|
398 |
.source-snippet {
|
399 |
color: #e5e7eb;
|
400 |
font-size: 0.9em;
|
401 |
-
line-height: 1.
|
402 |
}
|
403 |
.chat-history {
|
404 |
max-height: 400px;
|
@@ -407,6 +514,18 @@ css = """
|
|
407 |
background: #2c2d30;
|
408 |
border-radius: 8px;
|
409 |
margin-top: 1rem;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
410 |
}
|
411 |
.examples-container {
|
412 |
background: #2c2d30;
|
@@ -418,20 +537,73 @@ css = """
|
|
418 |
background: #3a3b3e !important;
|
419 |
border: 1px solid #4a4b4e !important;
|
420 |
color: #e5e7eb !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
}
|
422 |
.markdown-content {
|
423 |
color: #e5e7eb !important;
|
424 |
}
|
425 |
.markdown-content h1, .markdown-content h2, .markdown-content h3 {
|
426 |
color: white !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
427 |
}
|
428 |
.markdown-content a {
|
429 |
color: #60a5fa !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
430 |
}
|
431 |
.accordion {
|
432 |
background: #2c2d30 !important;
|
433 |
border-radius: 8px !important;
|
434 |
margin-top: 1rem !important;
|
|
|
435 |
}
|
436 |
.voice-selector {
|
437 |
margin-top: 1rem;
|
@@ -443,10 +615,54 @@ css = """
|
|
443 |
background: #3a3b3e !important;
|
444 |
color: white !important;
|
445 |
border: 1px solid #4a4b4e !important;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
446 |
}
|
447 |
"""
|
448 |
|
449 |
-
#
|
450 |
with gr.Blocks(title="AI Search Assistant", css=css, theme="dark") as demo:
|
451 |
chat_history = gr.State([])
|
452 |
|
@@ -462,13 +678,14 @@ with gr.Blocks(title="AI Search Assistant", css=css, theme="dark") as demo:
|
|
462 |
scale=5,
|
463 |
container=False
|
464 |
)
|
465 |
-
search_btn = gr.Button("Search", variant="primary", scale=1)
|
466 |
voice_select = gr.Dropdown(
|
467 |
-
choices=list(VOICE_CHOICES.
|
468 |
-
value=
|
469 |
-
label="
|
470 |
-
elem_classes="voice-selector"
|
|
|
471 |
)
|
|
|
472 |
|
473 |
with gr.Row(elem_classes="results-container"):
|
474 |
with gr.Column(scale=2):
|
@@ -486,28 +703,33 @@ with gr.Blocks(title="AI Search Assistant", css=css, theme="dark") as demo:
|
|
486 |
with gr.Row(elem_classes="examples-container"):
|
487 |
gr.Examples(
|
488 |
examples=[
|
489 |
-
"
|
490 |
-
"
|
491 |
"What are the best practices for sustainable living?",
|
492 |
-
"
|
493 |
],
|
494 |
inputs=search_input,
|
495 |
label="Try these examples"
|
496 |
)
|
497 |
|
|
|
|
|
|
|
|
|
498 |
# Handle interactions
|
499 |
search_btn.click(
|
500 |
fn=process_query,
|
501 |
-
inputs=[search_input, chat_history, voice_select],
|
502 |
outputs=[answer_output, sources_output, search_btn, chat_history_display, audio_output]
|
503 |
)
|
504 |
|
505 |
# Also trigger search on Enter key
|
506 |
search_input.submit(
|
507 |
fn=process_query,
|
508 |
-
inputs=[search_input, chat_history, voice_select],
|
509 |
outputs=[answer_output, sources_output, search_btn, chat_history_display, audio_output]
|
510 |
)
|
511 |
|
512 |
if __name__ == "__main__":
|
513 |
-
|
|
|
|
9 |
import subprocess
|
10 |
import numpy as np
|
11 |
from typing import List, Dict, Tuple, Any
|
12 |
+
from functools import lru_cache
|
13 |
+
import asyncio
|
14 |
+
import threading
|
15 |
+
from concurrent.futures import ThreadPoolExecutor
|
16 |
+
|
17 |
+
# --- Configuration ---
|
18 |
+
MODEL_NAME = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"
|
19 |
+
MAX_SEARCH_RESULTS = 5
|
20 |
+
TTS_SAMPLE_RATE = 24000
|
21 |
+
MAX_TTS_CHARS = 1000
|
22 |
+
GPU_DURATION = 30 # for spaces.GPU decorator
|
23 |
+
MAX_NEW_TOKENS = 256
|
24 |
+
TEMPERATURE = 0.7
|
25 |
+
TOP_P = 0.95
|
26 |
+
|
27 |
+
# --- Initialization ---
|
28 |
+
# Initialize model and tokenizer with better error handling
|
29 |
try:
|
30 |
+
print("Loading tokenizer...")
|
31 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
32 |
+
tokenizer.pad_token = tokenizer.eos_token
|
33 |
+
|
34 |
+
print("Loading model...")
|
35 |
+
model = AutoModelForCausalLM.from_pretrained(
|
36 |
+
MODEL_NAME,
|
37 |
+
device_map="auto",
|
38 |
+
offload_folder="offload",
|
39 |
+
low_cpu_mem_usage=True,
|
40 |
+
torch_dtype=torch.float16
|
41 |
+
)
|
42 |
+
print("Model and tokenizer loaded successfully")
|
|
|
|
|
43 |
except Exception as e:
|
44 |
+
print(f"Error initializing model: {str(e)}")
|
45 |
+
raise
|
46 |
+
|
47 |
+
# --- TTS Setup ---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
VOICE_CHOICES = {
|
49 |
'🇺🇸 Female (Default)': 'af',
|
50 |
'🇺🇸 Bella': 'af_bella',
|
|
|
53 |
}
|
54 |
TTS_ENABLED = False
|
55 |
TTS_MODEL = None
|
56 |
+
VOICEPACKS = {} # Cache voice packs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
|
58 |
+
# Initialize Kokoro TTS in a separate thread to avoid blocking startup
|
59 |
+
def setup_tts():
|
60 |
+
global TTS_ENABLED, TTS_MODEL, VOICEPACKS
|
61 |
+
|
62 |
+
try:
|
63 |
+
# Install dependencies first
|
64 |
+
subprocess.run(['git', 'lfs', 'install'], check=True)
|
65 |
+
if not os.path.exists('Kokoro-82M'):
|
66 |
+
subprocess.run(['git', 'clone', 'https://huggingface.co/hexgrad/Kokoro-82M'], check=True)
|
67 |
+
|
68 |
+
# Install espeak
|
69 |
try:
|
70 |
+
subprocess.run(['apt-get', 'update'], check=True)
|
71 |
+
subprocess.run(['apt-get', 'install', '-y', 'espeak'], check=True)
|
72 |
+
except subprocess.CalledProcessError:
|
73 |
+
try:
|
74 |
+
subprocess.run(['apt-get', 'install', '-y', 'espeak-ng'], check=True)
|
75 |
+
except subprocess.CalledProcessError:
|
76 |
+
print("Warning: Could not install espeak or espeak-ng. TTS functionality may be limited.")
|
77 |
+
|
78 |
+
# Set up Kokoro TTS
|
79 |
+
if os.path.exists('Kokoro-82M'):
|
80 |
+
import sys
|
81 |
+
sys.path.append('Kokoro-82M')
|
82 |
+
from models import build_model
|
83 |
+
from kokoro import generate
|
84 |
+
|
85 |
+
# Make these functions accessible globally
|
86 |
+
globals()['build_model'] = build_model
|
87 |
+
globals()['generate_tts'] = generate
|
88 |
+
|
89 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
90 |
+
TTS_MODEL = build_model('Kokoro-82M/kokoro-v0_19.pth', device)
|
91 |
+
|
92 |
+
# Preload default voice
|
93 |
+
default_voice = 'af'
|
94 |
+
VOICEPACKS[default_voice] = torch.load(f'Kokoro-82M/voices/{default_voice}.pt',
|
95 |
+
map_location=device,
|
96 |
+
weights_only=True)
|
97 |
+
|
98 |
+
# Preload other common voices to reduce latency
|
99 |
+
for voice_name in ['af_bella', 'af_sarah', 'af_nicole']:
|
100 |
+
try:
|
101 |
+
voice_path = f'Kokoro-82M/voices/{voice_name}.pt'
|
102 |
+
if os.path.exists(voice_path):
|
103 |
+
VOICEPACKS[voice_name] = torch.load(voice_path,
|
104 |
+
map_location=device,
|
105 |
+
weights_only=True)
|
106 |
+
except Exception as e:
|
107 |
+
print(f"Warning: Could not preload voice {voice_name}: {str(e)}")
|
108 |
+
|
109 |
+
TTS_ENABLED = True
|
110 |
+
print("TTS setup completed successfully")
|
111 |
+
else:
|
112 |
+
print("Warning: Kokoro-82M directory not found. TTS disabled.")
|
113 |
+
except Exception as e:
|
114 |
+
print(f"Warning: Could not initialize Kokoro TTS: {str(e)}")
|
115 |
+
TTS_ENABLED = False
|
116 |
|
117 |
+
# Start TTS setup in a separate thread
|
118 |
+
threading.Thread(target=setup_tts, daemon=True).start()
|
119 |
+
|
120 |
+
# --- Search and Generation Functions ---
|
121 |
+
@lru_cache(maxsize=128)
|
122 |
+
def get_web_results(query: str, max_results: int = MAX_SEARCH_RESULTS) -> List[Dict[str, str]]:
|
123 |
+
"""Get web search results using DuckDuckGo with caching for improved performance"""
|
124 |
try:
|
125 |
with DDGS() as ddgs:
|
126 |
results = list(ddgs.text(query, max_results=max_results))
|
127 |
return [{
|
128 |
"title": result.get("title", ""),
|
129 |
+
"snippet": result.get("body", ""),
|
130 |
+
"url": result.get("href", ""),
|
131 |
"date": result.get("published", "")
|
132 |
} for result in results]
|
133 |
except Exception as e:
|
|
|
155 |
sources_html = "<div class='sources-container'>"
|
156 |
for i, res in enumerate(web_results, 1):
|
157 |
title = res["title"] or "Source"
|
158 |
+
date = f"<span class='source-date'>{res['date']}</span>" if res.get('date') else ""
|
159 |
+
snippet = res.get("snippet", "")[:150] + "..." if res.get("snippet") else ""
|
160 |
sources_html += f"""
|
161 |
<div class='source-item'>
|
162 |
<div class='source-number'>[{i}]</div>
|
163 |
<div class='source-content'>
|
164 |
<a href="{res['url']}" target="_blank" class='source-title'>{title}</a>
|
165 |
{date}
|
166 |
+
<div class='source-snippet'>{snippet}</div>
|
167 |
</div>
|
168 |
</div>
|
169 |
"""
|
170 |
sources_html += "</div>"
|
171 |
return sources_html
|
172 |
|
173 |
+
@spaces.GPU(duration=GPU_DURATION)
|
174 |
def generate_answer(prompt: str) -> str:
|
175 |
+
"""Generate answer using the DeepSeek model with optimized settings"""
|
176 |
inputs = tokenizer(
|
177 |
prompt,
|
178 |
return_tensors="pt",
|
|
|
182 |
return_attention_mask=True
|
183 |
).to(model.device)
|
184 |
|
185 |
+
with torch.no_grad(): # Disable gradient calculation for inference
|
186 |
+
outputs = model.generate(
|
187 |
+
inputs.input_ids,
|
188 |
+
attention_mask=inputs.attention_mask,
|
189 |
+
max_new_tokens=MAX_NEW_TOKENS,
|
190 |
+
temperature=TEMPERATURE,
|
191 |
+
top_p=TOP_P,
|
192 |
+
pad_token_id=tokenizer.eos_token_id,
|
193 |
+
do_sample=True,
|
194 |
+
early_stopping=True
|
195 |
+
)
|
196 |
return tokenizer.decode(outputs[0], skip_special_tokens=True)
|
197 |
|
198 |
+
@spaces.GPU(duration=GPU_DURATION)
|
199 |
+
def generate_speech(text: str, voice_name: str = 'af') -> Tuple[int, np.ndarray] | None:
|
200 |
+
"""Generate speech from text using Kokoro TTS model with improved error handling and caching."""
|
201 |
+
global VOICEPACKS, TTS_MODEL, TTS_ENABLED
|
202 |
+
|
203 |
+
if not TTS_ENABLED or TTS_MODEL is None:
|
204 |
return None
|
205 |
|
206 |
try:
|
207 |
+
from kokoro import generate as generate_tts
|
208 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
209 |
|
210 |
+
# Load voicepack if needed
|
211 |
+
if voice_name not in VOICEPACKS:
|
212 |
+
voice_file = f'Kokoro-82M/voices/{voice_name}.pt'
|
213 |
+
|
214 |
+
if not os.path.exists(voice_file):
|
215 |
+
print(f"Voicepack {voice_name}.pt not found. Falling back to default 'af'.")
|
216 |
+
voice_name = 'af'
|
217 |
+
|
218 |
+
# Check if default is already loaded
|
219 |
+
if voice_name not in VOICEPACKS:
|
220 |
+
voice_file = f'Kokoro-82M/voices/{voice_name}.pt'
|
221 |
+
if os.path.exists(voice_file):
|
222 |
+
VOICEPACKS[voice_name] = torch.load(voice_file, map_location=device, weights_only=True)
|
223 |
+
else:
|
224 |
+
print("Default voicepack 'af.pt' not found. Cannot generate audio.")
|
225 |
+
return None
|
226 |
else:
|
227 |
+
VOICEPACKS[voice_name] = torch.load(voice_file, map_location=device, weights_only=True)
|
228 |
+
|
|
|
229 |
# Clean the text
|
230 |
clean_text = ' '.join([line for line in text.split('\n') if not line.startswith('#')])
|
231 |
clean_text = clean_text.replace('[', '').replace(']', '').replace('*', '')
|
232 |
|
233 |
# Split long text into chunks
|
234 |
+
max_chars = MAX_TTS_CHARS
|
235 |
chunks = []
|
236 |
if len(clean_text) > max_chars:
|
237 |
sentences = clean_text.split('.')
|
|
|
251 |
audio_chunks = []
|
252 |
for chunk in chunks:
|
253 |
if chunk.strip():
|
254 |
+
chunk_audio, _ = generate_tts(TTS_MODEL, chunk, VOICEPACKS[voice_name], lang='a')
|
255 |
if isinstance(chunk_audio, torch.Tensor):
|
256 |
chunk_audio = chunk_audio.cpu().numpy()
|
257 |
audio_chunks.append(chunk_audio)
|
|
|
259 |
# Concatenate chunks
|
260 |
if audio_chunks:
|
261 |
final_audio = np.concatenate(audio_chunks) if len(audio_chunks) > 1 else audio_chunks[0]
|
262 |
+
return (TTS_SAMPLE_RATE, final_audio)
|
263 |
+
|
264 |
+
return None
|
265 |
|
266 |
except Exception as e:
|
267 |
print(f"Error generating speech: {str(e)}")
|
268 |
return None
|
269 |
|
270 |
+
# --- Asynchronous Processing ---
|
271 |
+
async def async_web_search(query: str) -> List[Dict[str, str]]:
|
272 |
+
"""Run web search in a non-blocking way"""
|
273 |
+
loop = asyncio.get_event_loop()
|
274 |
+
return await loop.run_in_executor(None, get_web_results, query)
|
275 |
+
|
276 |
+
async def async_answer_generation(prompt: str) -> str:
|
277 |
+
"""Run answer generation in a non-blocking way"""
|
278 |
+
loop = asyncio.get_event_loop()
|
279 |
+
return await loop.run_in_executor(None, generate_answer, prompt)
|
280 |
+
|
281 |
+
async def async_speech_generation(text: str, voice_name: str) -> Tuple[int, np.ndarray] | None:
|
282 |
+
"""Run speech generation in a non-blocking way"""
|
283 |
+
loop = asyncio.get_event_loop()
|
284 |
+
return await loop.run_in_executor(None, generate_speech, text, voice_name)
|
285 |
+
|
286 |
def process_query(query: str, history: List[List[str]], selected_voice: str = 'af'):
|
287 |
+
"""Process user query with streaming effect and non-blocking operations"""
|
288 |
try:
|
289 |
if history is None:
|
290 |
history = []
|
291 |
|
292 |
+
# Start the search task
|
|
|
|
|
|
|
293 |
current_history = history + [[query, "*Searching...*"]]
|
294 |
+
|
295 |
# Yield initial searching state
|
296 |
yield (
|
297 |
"*Searching & Thinking...*", # answer_output (Markdown)
|
298 |
+
"<div class='searching'>Searching for results...</div>", # sources_output (HTML)
|
299 |
"Searching...", # search_btn (Button)
|
300 |
current_history, # chat_history_display (Chatbot)
|
301 |
None # audio_output (Audio)
|
302 |
)
|
303 |
|
304 |
+
# Get web results
|
305 |
+
web_results = get_web_results(query)
|
306 |
+
sources_html = format_sources(web_results)
|
307 |
+
|
308 |
+
# Update with the search results obtained
|
309 |
+
yield (
|
310 |
+
"*Analyzing search results...*", # answer_output
|
311 |
+
sources_html, # sources_output
|
312 |
+
"Generating answer...", # search_btn
|
313 |
+
current_history, # chat_history_display
|
314 |
+
None # audio_output
|
315 |
+
)
|
316 |
+
|
317 |
# Generate answer
|
318 |
prompt = format_prompt(query, web_results)
|
319 |
answer = generate_answer(prompt)
|
|
|
321 |
|
322 |
# Update history before TTS
|
323 |
updated_history = history + [[query, final_answer]]
|
324 |
+
|
325 |
+
# Update with the answer before generating speech
|
326 |
+
yield (
|
327 |
+
final_answer, # answer_output
|
328 |
+
sources_html, # sources_output
|
329 |
+
"Generating audio...", # search_btn
|
330 |
+
updated_history, # chat_history_display
|
331 |
+
None # audio_output
|
332 |
+
)
|
333 |
|
334 |
+
# Generate speech (but don't block if TTS is still initializing)
|
335 |
+
audio = None
|
336 |
+
if TTS_ENABLED and TTS_MODEL is not None:
|
|
|
|
|
|
|
|
|
|
|
|
|
337 |
try:
|
338 |
+
audio = generate_speech(final_answer, selected_voice)
|
339 |
if audio is None:
|
340 |
final_answer += "\n\n*Audio generation failed. The voicepack may be missing or incompatible.*"
|
341 |
except Exception as e:
|
342 |
final_answer += f"\n\n*Error generating audio: {str(e)}*"
|
|
|
343 |
else:
|
344 |
+
final_answer += "\n\n*TTS is still initializing or is disabled. Try again in a moment.*"
|
|
|
345 |
|
346 |
# Yield final result
|
347 |
yield (
|
|
|
349 |
sources_html, # sources_output
|
350 |
"Search", # search_btn
|
351 |
updated_history, # chat_history_display
|
352 |
+
audio # audio_output
|
353 |
)
|
354 |
|
355 |
except Exception as e:
|
|
|
358 |
error_message = "⚠️ GPU quota exceeded. Please try again later when the daily quota resets."
|
359 |
yield (
|
360 |
f"Error: {error_message}", # answer_output
|
361 |
+
"<div class='error'>An error occurred during search</div>", # sources_output
|
362 |
"Search", # search_btn
|
363 |
history + [[query, f"*Error: {error_message}*"]], # chat_history_display
|
364 |
None # audio_output
|
365 |
)
|
366 |
|
367 |
+
# --- Improved UI ---
|
368 |
css = """
|
369 |
.gradio-container {
|
370 |
max-width: 1200px !important;
|
|
|
374 |
text-align: center;
|
375 |
margin-bottom: 2rem;
|
376 |
padding: 2rem 0;
|
377 |
+
background: linear-gradient(135deg, #1a1b1e, #2d2e32);
|
378 |
border-radius: 12px;
|
379 |
color: white;
|
380 |
+
box-shadow: 0 8px 32px rgba(0,0,0,0.2);
|
381 |
}
|
382 |
#header h1 {
|
383 |
color: white;
|
384 |
font-size: 2.5rem;
|
385 |
margin-bottom: 0.5rem;
|
386 |
+
text-shadow: 0 2px 4px rgba(0,0,0,0.3);
|
387 |
}
|
388 |
#header h3 {
|
389 |
color: #a8a9ab;
|
390 |
}
|
391 |
.search-container {
|
392 |
+
background: linear-gradient(135deg, #1a1b1e, #2d2e32);
|
393 |
border-radius: 12px;
|
394 |
+
box-shadow: 0 4px 16px rgba(0,0,0,0.15);
|
395 |
+
padding: 1.5rem;
|
396 |
+
margin-bottom: 1.5rem;
|
397 |
}
|
398 |
.search-box {
|
399 |
padding: 1rem;
|
400 |
background: #2c2d30;
|
401 |
+
border-radius: 10px;
|
402 |
margin-bottom: 1rem;
|
403 |
+
box-shadow: inset 0 2px 4px rgba(0,0,0,0.1);
|
404 |
}
|
405 |
.search-box input[type="text"] {
|
406 |
background: #3a3b3e !important;
|
407 |
border: 1px solid #4a4b4e !important;
|
408 |
color: white !important;
|
409 |
border-radius: 8px !important;
|
410 |
+
transition: all 0.3s ease;
|
411 |
+
}
|
412 |
+
.search-box input[type="text"]:focus {
|
413 |
+
border-color: #60a5fa !important;
|
414 |
+
box-shadow: 0 0 0 2px rgba(96, 165, 250, 0.3) !important;
|
415 |
}
|
416 |
.search-box input[type="text"]::placeholder {
|
417 |
color: #a8a9ab !important;
|
|
|
419 |
.search-box button {
|
420 |
background: #2563eb !important;
|
421 |
border: none !important;
|
422 |
+
box-shadow: 0 2px 4px rgba(0,0,0,0.2) !important;
|
423 |
+
transition: all 0.3s ease !important;
|
424 |
+
}
|
425 |
+
.search-box button:hover {
|
426 |
+
background: #1d4ed8 !important;
|
427 |
+
transform: translateY(-1px) !important;
|
428 |
+
}
|
429 |
+
.search-box button:active {
|
430 |
+
transform: translateY(1px) !important;
|
431 |
}
|
432 |
.results-container {
|
433 |
background: #2c2d30;
|
434 |
+
border-radius: 10px;
|
435 |
+
padding: 1.5rem;
|
436 |
+
margin-top: 1.5rem;
|
437 |
+
box-shadow: 0 4px 12px rgba(0,0,0,0.1);
|
438 |
}
|
439 |
.answer-box {
|
440 |
background: #3a3b3e;
|
441 |
+
border-radius: 10px;
|
442 |
padding: 1.5rem;
|
443 |
color: white;
|
444 |
+
margin-bottom: 1.5rem;
|
445 |
+
box-shadow: 0 2px 8px rgba(0,0,0,0.15);
|
446 |
+
transition: all 0.3s ease;
|
447 |
+
}
|
448 |
+
.answer-box:hover {
|
449 |
+
box-shadow: 0 4px 16px rgba(0,0,0,0.2);
|
450 |
}
|
451 |
.answer-box p {
|
452 |
color: #e5e7eb;
|
453 |
+
line-height: 1.7;
|
454 |
+
}
|
455 |
+
.answer-box code {
|
456 |
+
background: #2c2d30;
|
457 |
+
border-radius: 4px;
|
458 |
+
padding: 2px 4px;
|
459 |
}
|
460 |
.sources-container {
|
461 |
margin-top: 1rem;
|
|
|
466 |
.source-item {
|
467 |
display: flex;
|
468 |
padding: 12px;
|
469 |
+
margin: 12px 0;
|
470 |
background: #3a3b3e;
|
471 |
border-radius: 8px;
|
472 |
transition: all 0.2s;
|
473 |
+
box-shadow: 0 2px 4px rgba(0,0,0,0.1);
|
474 |
}
|
475 |
.source-item:hover {
|
476 |
background: #4a4b4e;
|
477 |
+
transform: translateY(-2px);
|
478 |
+
box-shadow: 0 4px 8px rgba(0,0,0,0.15);
|
479 |
}
|
480 |
.source-number {
|
481 |
font-weight: bold;
|
|
|
490 |
font-weight: 500;
|
491 |
text-decoration: none;
|
492 |
display: block;
|
493 |
+
margin-bottom: 6px;
|
494 |
+
transition: all 0.2s;
|
495 |
+
}
|
496 |
+
.source-title:hover {
|
497 |
+
color: #93c5fd;
|
498 |
+
text-decoration: underline;
|
499 |
}
|
500 |
.source-date {
|
501 |
color: #a8a9ab;
|
|
|
505 |
.source-snippet {
|
506 |
color: #e5e7eb;
|
507 |
font-size: 0.9em;
|
508 |
+
line-height: 1.5;
|
509 |
}
|
510 |
.chat-history {
|
511 |
max-height: 400px;
|
|
|
514 |
background: #2c2d30;
|
515 |
border-radius: 8px;
|
516 |
margin-top: 1rem;
|
517 |
+
scrollbar-width: thin;
|
518 |
+
scrollbar-color: #4a4b4e #2c2d30;
|
519 |
+
}
|
520 |
+
.chat-history::-webkit-scrollbar {
|
521 |
+
width: 8px;
|
522 |
+
}
|
523 |
+
.chat-history::-webkit-scrollbar-track {
|
524 |
+
background: #2c2d30;
|
525 |
+
}
|
526 |
+
.chat-history::-webkit-scrollbar-thumb {
|
527 |
+
background-color: #4a4b4e;
|
528 |
+
border-radius: 20px;
|
529 |
}
|
530 |
.examples-container {
|
531 |
background: #2c2d30;
|
|
|
537 |
background: #3a3b3e !important;
|
538 |
border: 1px solid #4a4b4e !important;
|
539 |
color: #e5e7eb !important;
|
540 |
+
transition: all 0.2s;
|
541 |
+
margin: 4px !important;
|
542 |
+
}
|
543 |
+
.examples-container button:hover {
|
544 |
+
background: #4a4b4e !important;
|
545 |
+
transform: translateY(-1px);
|
546 |
}
|
547 |
.markdown-content {
|
548 |
color: #e5e7eb !important;
|
549 |
}
|
550 |
.markdown-content h1, .markdown-content h2, .markdown-content h3 {
|
551 |
color: white !important;
|
552 |
+
margin-top: 1.2em !important;
|
553 |
+
margin-bottom: 0.8em !important;
|
554 |
+
}
|
555 |
+
.markdown-content h1 {
|
556 |
+
font-size: 1.7em !important;
|
557 |
+
}
|
558 |
+
.markdown-content h2 {
|
559 |
+
font-size: 1.5em !important;
|
560 |
+
}
|
561 |
+
.markdown-content h3 {
|
562 |
+
font-size: 1.3em !important;
|
563 |
}
|
564 |
.markdown-content a {
|
565 |
color: #60a5fa !important;
|
566 |
+
text-decoration: none !important;
|
567 |
+
transition: all 0.2s;
|
568 |
+
}
|
569 |
+
.markdown-content a:hover {
|
570 |
+
color: #93c5fd !important;
|
571 |
+
text-decoration: underline !important;
|
572 |
+
}
|
573 |
+
.markdown-content code {
|
574 |
+
background: #2c2d30 !important;
|
575 |
+
padding: 2px 6px !important;
|
576 |
+
border-radius: 4px !important;
|
577 |
+
font-family: monospace !important;
|
578 |
+
}
|
579 |
+
.markdown-content pre {
|
580 |
+
background: #2c2d30 !important;
|
581 |
+
padding: 12px !important;
|
582 |
+
border-radius: 8px !important;
|
583 |
+
overflow-x: auto !important;
|
584 |
+
}
|
585 |
+
.markdown-content blockquote {
|
586 |
+
border-left: 4px solid #60a5fa !important;
|
587 |
+
padding-left: 1em !important;
|
588 |
+
margin-left: 0 !important;
|
589 |
+
color: #a8a9ab !important;
|
590 |
+
}
|
591 |
+
.markdown-content table {
|
592 |
+
border-collapse: collapse !important;
|
593 |
+
width: 100% !important;
|
594 |
+
}
|
595 |
+
.markdown-content th, .markdown-content td {
|
596 |
+
padding: 8px 12px !important;
|
597 |
+
border: 1px solid #4a4b4e !important;
|
598 |
+
}
|
599 |
+
.markdown-content th {
|
600 |
+
background: #2c2d30 !important;
|
601 |
}
|
602 |
.accordion {
|
603 |
background: #2c2d30 !important;
|
604 |
border-radius: 8px !important;
|
605 |
margin-top: 1rem !important;
|
606 |
+
box-shadow: 0 2px 8px rgba(0,0,0,0.1) !important;
|
607 |
}
|
608 |
.voice-selector {
|
609 |
margin-top: 1rem;
|
|
|
615 |
background: #3a3b3e !important;
|
616 |
color: white !important;
|
617 |
border: 1px solid #4a4b4e !important;
|
618 |
+
border-radius: 4px !important;
|
619 |
+
padding: 8px !important;
|
620 |
+
transition: all 0.2s;
|
621 |
+
}
|
622 |
+
.voice-selector select:focus {
|
623 |
+
border-color: #60a5fa !important;
|
624 |
+
}
|
625 |
+
.audio-player {
|
626 |
+
margin-top: 1rem;
|
627 |
+
background: #2c2d30 !important;
|
628 |
+
border-radius: 8px !important;
|
629 |
+
padding: 0.5rem !important;
|
630 |
+
}
|
631 |
+
.audio-player audio {
|
632 |
+
width: 100% !important;
|
633 |
+
}
|
634 |
+
.searching, .error {
|
635 |
+
padding: 1rem;
|
636 |
+
border-radius: 8px;
|
637 |
+
text-align: center;
|
638 |
+
margin: 1rem 0;
|
639 |
+
}
|
640 |
+
.searching {
|
641 |
+
background: rgba(96, 165, 250, 0.1);
|
642 |
+
color: #60a5fa;
|
643 |
+
}
|
644 |
+
.error {
|
645 |
+
background: rgba(239, 68, 68, 0.1);
|
646 |
+
color: #ef4444;
|
647 |
+
}
|
648 |
+
.no-sources {
|
649 |
+
padding: 1rem;
|
650 |
+
text-align: center;
|
651 |
+
color: #a8a9ab;
|
652 |
+
background: #2c2d30;
|
653 |
+
border-radius: 8px;
|
654 |
+
}
|
655 |
+
@keyframes pulse {
|
656 |
+
0% { opacity: 0.6; }
|
657 |
+
50% { opacity: 1; }
|
658 |
+
100% { opacity: 0.6; }
|
659 |
+
}
|
660 |
+
.searching {
|
661 |
+
animation: pulse 1.5s infinite;
|
662 |
}
|
663 |
"""
|
664 |
|
665 |
+
# --- Gradio Interface ---
|
666 |
with gr.Blocks(title="AI Search Assistant", css=css, theme="dark") as demo:
|
667 |
chat_history = gr.State([])
|
668 |
|
|
|
678 |
scale=5,
|
679 |
container=False
|
680 |
)
|
|
|
681 |
voice_select = gr.Dropdown(
|
682 |
+
choices=list(VOICE_CHOICES.keys()),
|
683 |
+
value=list(VOICE_CHOICES.keys())[0],
|
684 |
+
label="Voice",
|
685 |
+
elem_classes="voice-selector",
|
686 |
+
scale=1
|
687 |
)
|
688 |
+
search_btn = gr.Button("Search", variant="primary", scale=1)
|
689 |
|
690 |
with gr.Row(elem_classes="results-container"):
|
691 |
with gr.Column(scale=2):
|
|
|
703 |
with gr.Row(elem_classes="examples-container"):
|
704 |
gr.Examples(
|
705 |
examples=[
|
706 |
+
"Latest news about artificial intelligence advances",
|
707 |
+
"How does blockchain technology work?",
|
708 |
"What are the best practices for sustainable living?",
|
709 |
+
"Compare electric vehicles and traditional cars"
|
710 |
],
|
711 |
inputs=search_input,
|
712 |
label="Try these examples"
|
713 |
)
|
714 |
|
715 |
+
# Handle voice selection mapping
|
716 |
+
def get_voice_id(voice_name):
|
717 |
+
return VOICE_CHOICES.get(voice_name, 'af')
|
718 |
+
|
719 |
# Handle interactions
|
720 |
search_btn.click(
|
721 |
fn=process_query,
|
722 |
+
inputs=[search_input, chat_history, lambda x: get_voice_id(x), voice_select],
|
723 |
outputs=[answer_output, sources_output, search_btn, chat_history_display, audio_output]
|
724 |
)
|
725 |
|
726 |
# Also trigger search on Enter key
|
727 |
search_input.submit(
|
728 |
fn=process_query,
|
729 |
+
inputs=[search_input, chat_history, lambda x: get_voice_id(x), voice_select],
|
730 |
outputs=[answer_output, sources_output, search_btn, chat_history_display, audio_output]
|
731 |
)
|
732 |
|
733 |
if __name__ == "__main__":
|
734 |
+
# Start the app with optimized settings
|
735 |
+
demo.queue(concurrency_count=5, max_size=20).launch(share=True)
|