Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,6 +1,8 @@
|
|
1 |
import os
|
2 |
import base64
|
|
|
3 |
from io import BytesIO
|
|
|
4 |
|
5 |
import gradio as gr
|
6 |
from gradio_pdf import PDF
|
@@ -13,179 +15,272 @@ from tqdm import tqdm
|
|
13 |
|
14 |
from colpali_engine.models import ColQwen2, ColQwen2Processor
|
15 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
|
|
|
|
|
|
|
|
17 |
|
18 |
model = ColQwen2.from_pretrained(
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
processor = ColQwen2Processor.from_pretrained("vidore/colqwen2-v1.0")
|
25 |
|
26 |
|
27 |
-
|
|
|
|
|
|
|
28 |
"""Encodes a PIL image to a base64 string."""
|
29 |
buffered = BytesIO()
|
30 |
image.save(buffered, format="JPEG")
|
31 |
return base64.b64encode(buffered.getvalue()).decode("utf-8")
|
32 |
-
|
33 |
|
34 |
-
def query_gpt4o_mini(query, images, api_key):
|
35 |
-
"""Calls OpenAI's GPT-4o-mini with the query and image data."""
|
36 |
|
|
|
|
|
37 |
if api_key and api_key.startswith("sk"):
|
38 |
try:
|
39 |
from openai import OpenAI
|
40 |
-
|
41 |
-
base64_images = [encode_image_to_base64(
|
42 |
client = OpenAI(api_key=api_key.strip())
|
43 |
PROMPT = """
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
response = client.chat.completions.create(
|
56 |
-
|
57 |
-
|
58 |
-
{
|
59 |
-
"role": "user",
|
60 |
-
"content": [
|
61 |
{
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
}
|
70 |
-
|
71 |
-
|
72 |
-
max_tokens=500,
|
73 |
)
|
74 |
return response.choices[0].message.content
|
75 |
-
except Exception
|
76 |
-
return "OpenAI API connection failure. Verify
|
77 |
-
|
78 |
-
return "Enter your OpenAI API key to get a custom response"
|
79 |
-
|
80 |
-
|
81 |
-
def search(query: str, ds, images, k, api_key):
|
82 |
-
k = min(k, len(ds))
|
83 |
-
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
84 |
-
if device != model.device:
|
85 |
-
model.to(device)
|
86 |
-
|
87 |
-
qs = []
|
88 |
-
with torch.no_grad():
|
89 |
-
batch_query = processor.process_queries([query]).to(model.device)
|
90 |
-
embeddings_query = model(**batch_query)
|
91 |
-
qs.extend(list(torch.unbind(embeddings_query.to("cpu"))))
|
92 |
-
|
93 |
-
scores = processor.score(qs, ds, device=device)
|
94 |
|
95 |
-
top_k_indices = scores[0].topk(k).indices.tolist()
|
96 |
-
|
97 |
-
results = []
|
98 |
-
for idx in top_k_indices:
|
99 |
-
results.append((images[idx], f"Page {idx}"))
|
100 |
-
|
101 |
-
# Generate response from GPT-4o-mini
|
102 |
-
ai_response = query_gpt4o_mini(query, results, api_key)
|
103 |
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
images = convert_files(files)
|
110 |
-
print(f"Files converted with {len(images)} images.")
|
111 |
-
return index_gpu(images, ds)
|
112 |
-
|
113 |
|
114 |
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
if len(images) >= 500:
|
123 |
raise gr.Error("The number of images in the dataset should be less than 500.")
|
124 |
-
return
|
|
|
125 |
|
|
|
|
|
|
|
|
|
126 |
|
127 |
-
|
128 |
-
|
|
|
129 |
|
130 |
-
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
131 |
-
if device != model.device:
|
132 |
-
model.to(device)
|
133 |
-
|
134 |
-
# run inference - docs
|
135 |
dataloader = DataLoader(
|
136 |
images,
|
137 |
batch_size=4,
|
138 |
-
# num_workers=4,
|
139 |
shuffle=False,
|
140 |
collate_fn=lambda x: processor.process_images(x).to(model.device),
|
141 |
)
|
142 |
|
143 |
-
for batch_doc in tqdm(dataloader):
|
144 |
with torch.no_grad():
|
145 |
batch_doc = {k: v.to(device) for k, v in batch_doc.items()}
|
146 |
embeddings_doc = model(**batch_doc)
|
147 |
ds.extend(list(torch.unbind(embeddings_doc.to("cpu"))))
|
148 |
-
return f"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
|
|
|
|
|
|
|
150 |
|
|
|
|
|
|
|
|
|
|
|
151 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
153 |
gr.Markdown("# ColPali: Efficient Document Retrieval with Vision Language Models (ColQwen2) 📚")
|
154 |
-
gr.Markdown(
|
155 |
-
|
|
|
156 |
|
157 |
-
|
158 |
-
|
|
|
|
|
|
|
159 |
|
160 |
-
⚠️ This demo uses a model trained exclusively on A4 PDFs in portrait mode, containing english text. Performance is expected to drop for other page formats and languages.
|
161 |
-
Other models will be released with better robustness towards different languages and document formats !
|
162 |
-
""")
|
163 |
with gr.Row():
|
164 |
with gr.Column(scale=2):
|
165 |
-
gr.Markdown("## 1️⃣
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
|
|
175 |
|
176 |
with gr.Column(scale=3):
|
177 |
gr.Markdown("## 2️⃣ Search")
|
178 |
query = gr.Textbox(placeholder="Enter your query here", label="Query")
|
179 |
-
|
180 |
-
|
|
|
|
|
181 |
|
182 |
-
#
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
convert_button.click(index, inputs=[file, embeds], outputs=[message, embeds, imgs])
|
188 |
-
search_button.click(search, inputs=[query, embeds, imgs, k, api_key], outputs=[output_gallery, output_text])
|
189 |
|
190 |
if __name__ == "__main__":
|
191 |
-
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import os
|
2 |
import base64
|
3 |
+
import tempfile
|
4 |
from io import BytesIO
|
5 |
+
from urllib.request import urlretrieve
|
6 |
|
7 |
import gradio as gr
|
8 |
from gradio_pdf import PDF
|
|
|
15 |
|
16 |
from colpali_engine.models import ColQwen2, ColQwen2Processor
|
17 |
|
18 |
+
# -----------------------------
|
19 |
+
# Globals
|
20 |
+
# -----------------------------
|
21 |
+
api_key = os.getenv("OPENAI_API_KEY", "") # <- use env var
|
22 |
+
ds = [] # list of document embeddings (torch tensors)
|
23 |
+
images = [] # list of PIL images (page-order)
|
24 |
+
current_pdf_path = None # last (indexed) pdf path for preview
|
25 |
|
26 |
+
# -----------------------------
|
27 |
+
# Model & processor
|
28 |
+
# -----------------------------
|
29 |
+
device_map = "cuda:0" if torch.cuda.is_available() else ("mps" if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available() else "cpu")
|
30 |
|
31 |
model = ColQwen2.from_pretrained(
|
32 |
+
"vidore/colqwen2-v1.0",
|
33 |
+
torch_dtype=torch.bfloat16,
|
34 |
+
device_map=device_map,
|
35 |
+
attn_implementation="flash_attention_2"
|
36 |
+
).eval()
|
37 |
processor = ColQwen2Processor.from_pretrained("vidore/colqwen2-v1.0")
|
38 |
|
39 |
|
40 |
+
# -----------------------------
|
41 |
+
# Utilities
|
42 |
+
# -----------------------------
|
43 |
+
def encode_image_to_base64(image: Image.Image) -> str:
|
44 |
"""Encodes a PIL image to a base64 string."""
|
45 |
buffered = BytesIO()
|
46 |
image.save(buffered, format="JPEG")
|
47 |
return base64.b64encode(buffered.getvalue()).decode("utf-8")
|
|
|
48 |
|
|
|
|
|
49 |
|
50 |
+
def query_gpt(query: str, retrieved_images: list[tuple[Image.Image, str]]) -> str:
|
51 |
+
"""Calls OpenAI's GPT model with the query and image data."""
|
52 |
if api_key and api_key.startswith("sk"):
|
53 |
try:
|
54 |
from openai import OpenAI
|
55 |
+
|
56 |
+
base64_images = [encode_image_to_base64(im_caption[0]) for im_caption in retrieved_images]
|
57 |
client = OpenAI(api_key=api_key.strip())
|
58 |
PROMPT = """
|
59 |
+
You are a smart assistant designed to answer questions about a PDF document.
|
60 |
+
You are given relevant information in the form of PDF pages. Use them to construct a short response to the question, and cite your sources (page numbers, etc).
|
61 |
+
If it is not possible to answer using the provided pages, do not attempt to provide an answer and simply say the answer is not present within the documents.
|
62 |
+
Give detailed and extensive answers, only containing info in the pages you are given.
|
63 |
+
You can answer using information contained in plots and figures if necessary.
|
64 |
+
Answer in the same language as the query.
|
65 |
+
|
66 |
+
Query: {query}
|
67 |
+
PDF pages:
|
68 |
+
""".strip()
|
69 |
+
|
70 |
response = client.chat.completions.create(
|
71 |
+
model="gpt-5-mini",
|
72 |
+
messages=[
|
|
|
|
|
|
|
73 |
{
|
74 |
+
"role": "user",
|
75 |
+
"content": (
|
76 |
+
[{"type": "text", "text": PROMPT.format(query=query)}] +
|
77 |
+
[{"type": "image_url",
|
78 |
+
"image_url": {"url": f"data:image/jpeg;base64,{im}"}}
|
79 |
+
for im in base64_images]
|
80 |
+
)
|
81 |
+
}
|
82 |
+
],
|
83 |
+
max_tokens=500,
|
|
|
84 |
)
|
85 |
return response.choices[0].message.content
|
86 |
+
except Exception:
|
87 |
+
return "OpenAI API connection failure. Verify that OPENAI_API_KEY is set and valid (sk-***)."
|
88 |
+
return "Set OPENAI_API_KEY in your environment to get a custom response."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
90 |
|
91 |
+
def _ensure_model_device():
|
92 |
+
dev = "cuda:0" if torch.cuda.is_available() else ("mps" if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available() else "cpu")
|
93 |
+
if str(model.device) != dev:
|
94 |
+
model.to(dev)
|
95 |
+
return dev
|
|
|
|
|
|
|
|
|
96 |
|
97 |
|
98 |
+
# -----------------------------
|
99 |
+
# Indexing helpers
|
100 |
+
# -----------------------------
|
101 |
+
def convert_files(pdf_path: str) -> list[Image.Image]:
|
102 |
+
"""Convert a single PDF path into a list of PIL Images (pages)."""
|
103 |
+
imgs = convert_from_path(pdf_path, thread_count=4)
|
104 |
+
if len(imgs) >= 500:
|
|
|
105 |
raise gr.Error("The number of images in the dataset should be less than 500.")
|
106 |
+
return imgs
|
107 |
+
|
108 |
|
109 |
+
def index_gpu(imgs: list[Image.Image]) -> str:
|
110 |
+
"""Embed a list of images (pages) with ColPali and store in globals."""
|
111 |
+
global ds, images
|
112 |
+
device = _ensure_model_device()
|
113 |
|
114 |
+
# reset previous dataset
|
115 |
+
ds = []
|
116 |
+
images = imgs
|
117 |
|
|
|
|
|
|
|
|
|
|
|
118 |
dataloader = DataLoader(
|
119 |
images,
|
120 |
batch_size=4,
|
|
|
121 |
shuffle=False,
|
122 |
collate_fn=lambda x: processor.process_images(x).to(model.device),
|
123 |
)
|
124 |
|
125 |
+
for batch_doc in tqdm(dataloader, desc="Indexing pages"):
|
126 |
with torch.no_grad():
|
127 |
batch_doc = {k: v.to(device) for k, v in batch_doc.items()}
|
128 |
embeddings_doc = model(**batch_doc)
|
129 |
ds.extend(list(torch.unbind(embeddings_doc.to("cpu"))))
|
130 |
+
return f"Indexed {len(images)} pages successfully."
|
131 |
+
|
132 |
+
|
133 |
+
def index_from_path(pdf_path: str) -> str:
|
134 |
+
"""Public: index a local PDF file path."""
|
135 |
+
imgs = convert_files(pdf_path)
|
136 |
+
return index_gpu(imgs)
|
137 |
+
|
138 |
+
|
139 |
+
def index_from_url(url: str) -> tuple[str, str]:
|
140 |
+
"""
|
141 |
+
Download a PDF from URL and index it.
|
142 |
+
|
143 |
+
Returns:
|
144 |
+
status message, saved pdf path
|
145 |
+
"""
|
146 |
+
tmp_dir = tempfile.mkdtemp(prefix="colpali_")
|
147 |
+
local_path = os.path.join(tmp_dir, "document.pdf")
|
148 |
+
urlretrieve(url, local_path)
|
149 |
+
status = index_from_path(local_path)
|
150 |
+
return status, local_path
|
151 |
+
|
152 |
+
|
153 |
+
# -----------------------------
|
154 |
+
# Search (MCP tool-friendly)
|
155 |
+
# -----------------------------
|
156 |
+
def search(query: str, k: int):
|
157 |
+
"""
|
158 |
+
Search the currently indexed PDF pages for the most relevant content and
|
159 |
+
generate an answer grounded ONLY in those pages.
|
160 |
+
|
161 |
+
MCP tool description:
|
162 |
+
- name: search
|
163 |
+
- description: Retrieve top-k PDF pages relevant to a query and answer using only those pages.
|
164 |
+
- input_schema:
|
165 |
+
type: object
|
166 |
+
properties:
|
167 |
+
query: {type: string, description: "User query in natural language."}
|
168 |
+
k: {type: integer, minimum: 1, maximum: 10, description: "Number of top pages to retrieve."}
|
169 |
+
required: ["query"]
|
170 |
+
|
171 |
+
Args:
|
172 |
+
query (str): Natural-language question to search for.
|
173 |
+
k (int): Number of top results to return (1–10).
|
174 |
+
|
175 |
+
Returns:
|
176 |
+
tuple:
|
177 |
+
- results (list[tuple[PIL.Image.Image, str]]): List of (page_image, caption) pairs for a gallery.
|
178 |
+
- ai_response (str): Answer grounded only in retrieved pages, with citations (page numbers).
|
179 |
+
|
180 |
+
Notes:
|
181 |
+
• Requires that a PDF has been indexed first.
|
182 |
+
• Citations reference 1-based page numbers as shown in the gallery captions.
|
183 |
+
"""
|
184 |
+
global ds, images
|
185 |
+
|
186 |
+
if not images or not ds:
|
187 |
+
return [], "No document indexed yet. Upload a PDF or load the sample, then run Search."
|
188 |
+
|
189 |
+
k = max(1, min(int(k), len(images)))
|
190 |
+
device = _ensure_model_device()
|
191 |
+
|
192 |
+
# Encode query
|
193 |
+
qs = []
|
194 |
+
with torch.no_grad():
|
195 |
+
batch_query = processor.process_queries([query]).to(model.device)
|
196 |
+
embeddings_query = model(**batch_query)
|
197 |
+
qs.extend(list(torch.unbind(embeddings_query.to("cpu"))))
|
198 |
|
199 |
+
# Score and select top-k
|
200 |
+
scores = processor.score(qs, ds, device=device)
|
201 |
+
top_k_indices = scores[0].topk(k).indices.tolist()
|
202 |
|
203 |
+
# Build gallery results with 1-based page numbering
|
204 |
+
results = []
|
205 |
+
for idx in top_k_indices:
|
206 |
+
page_num = idx + 1
|
207 |
+
results.append((images[idx], f"Page {page_num}"))
|
208 |
|
209 |
+
# Generate grounded response
|
210 |
+
ai_response = query_gpt(query, results)
|
211 |
+
return results, ai_response
|
212 |
+
|
213 |
+
|
214 |
+
# -----------------------------
|
215 |
+
# Gradio UI callbacks
|
216 |
+
# -----------------------------
|
217 |
+
def handle_upload(file) -> tuple[str, str | None]:
|
218 |
+
"""Index a user-uploaded PDF file."""
|
219 |
+
global current_pdf_path
|
220 |
+
if file is None:
|
221 |
+
return "Please upload a PDF.", None
|
222 |
+
path = getattr(file, "name", file)
|
223 |
+
status = index_from_path(path)
|
224 |
+
current_pdf_path = path
|
225 |
+
return status, path
|
226 |
+
|
227 |
+
|
228 |
+
def handle_url(url: str) -> tuple[str, str | None]:
|
229 |
+
"""Index a PDF from URL (e.g., a sample)."""
|
230 |
+
global current_pdf_path
|
231 |
+
if not url or not url.lower().endswith(".pdf"):
|
232 |
+
return "Please provide a direct PDF URL.", None
|
233 |
+
status, path = index_from_url(url)
|
234 |
+
current_pdf_path = path
|
235 |
+
return status, path
|
236 |
+
|
237 |
+
|
238 |
+
# -----------------------------
|
239 |
+
# Gradio App
|
240 |
+
# -----------------------------
|
241 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
242 |
gr.Markdown("# ColPali: Efficient Document Retrieval with Vision Language Models (ColQwen2) 📚")
|
243 |
+
gr.Markdown(
|
244 |
+
"""Demo to test ColQwen2 (ColPali) on PDF documents.
|
245 |
+
ColPali is implemented from the [ColPali paper](https://arxiv.org/abs/2407.01449).
|
246 |
|
247 |
+
This demo lets you **upload a PDF or load a sample**, then **search** for the most relevant pages and get a grounded answer.
|
248 |
+
|
249 |
+
⚠️ The model was trained on A4 portrait English PDFs; performance may drop on other formats/languages.
|
250 |
+
"""
|
251 |
+
)
|
252 |
|
|
|
|
|
|
|
253 |
with gr.Row():
|
254 |
with gr.Column(scale=2):
|
255 |
+
gr.Markdown("## 1️⃣ Load a PDF")
|
256 |
+
pdf_input = gr.File(label="Upload PDF", file_types=[".pdf"])
|
257 |
+
index_btn = gr.Button("📥 Index Uploaded PDF", variant="secondary")
|
258 |
+
url_box = gr.Textbox(
|
259 |
+
label="Or index from URL",
|
260 |
+
placeholder="https://example.com/file.pdf",
|
261 |
+
value="https://sist.sathyabama.ac.in/sist_coursematerial/uploads/SAR1614.pdf",
|
262 |
+
)
|
263 |
+
index_url_btn = gr.Button("🌐 Load Sample / From URL", variant="secondary")
|
264 |
+
status_box = gr.Textbox(label="Status", interactive=False)
|
265 |
+
pdf_view = PDF(label="PDF Preview")
|
266 |
|
267 |
with gr.Column(scale=3):
|
268 |
gr.Markdown("## 2️⃣ Search")
|
269 |
query = gr.Textbox(placeholder="Enter your query here", label="Query")
|
270 |
+
k_slider = gr.Slider(minimum=1, maximum=10, step=1, label="Number of results", value=5)
|
271 |
+
search_button = gr.Button("🔍 Search", variant="primary")
|
272 |
+
output_gallery = gr.Gallery(label="Retrieved Pages", height=600, show_label=True)
|
273 |
+
output_text = gr.Textbox(label="AI Response", placeholder="Generated response based on retrieved documents")
|
274 |
|
275 |
+
# Wiring
|
276 |
+
index_btn.click(handle_upload, inputs=[pdf_input], outputs=[status_box, pdf_view])
|
277 |
+
index_url_btn.click(handle_url, inputs=[url_box], outputs=[status_box, pdf_view])
|
278 |
+
search_button.click(search, inputs=[query, k_slider], outputs=[output_gallery, output_text])
|
|
|
|
|
|
|
279 |
|
280 |
if __name__ == "__main__":
|
281 |
+
# Optional: pre-load the default sample at startup.
|
282 |
+
# Comment these two lines if you prefer a "cold" start.
|
283 |
+
# msg, path = index_from_url("https://sist.sathyabama.ac.in/sist_coursematerial/uploads/SAR1614.pdf")
|
284 |
+
# print(msg, "->", path)
|
285 |
+
|
286 |
+
demo.queue(max_size=5).launch(debug=True, mcp_server=True)
|