Spaces:
Sleeping
Sleeping
Samuel Thomas
commited on
Commit
·
38df4e4
1
Parent(s):
6f21ce8
reading from api
Browse files
app.py
CHANGED
@@ -3,6 +3,7 @@ import gradio as gr
|
|
3 |
import requests
|
4 |
import inspect
|
5 |
import pandas as pd
|
|
|
6 |
|
7 |
# (Keep Constants as is)
|
8 |
# --- Constants ---
|
@@ -53,11 +54,11 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
|
|
53 |
try:
|
54 |
response = requests.get(questions_url, timeout=15)
|
55 |
response.raise_for_status()
|
56 |
-
|
57 |
-
if not
|
58 |
print("Fetched questions list is empty.")
|
59 |
return "Fetched questions list is empty or invalid format.", None
|
60 |
-
print(f"Fetched {len(
|
61 |
except requests.exceptions.RequestException as e:
|
62 |
print(f"Error fetching questions: {e}")
|
63 |
return f"Error fetching questions: {e}", None
|
@@ -68,7 +69,26 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
|
|
68 |
except Exception as e:
|
69 |
print(f"An unexpected error occurred fetching questions: {e}")
|
70 |
return f"An unexpected error occurred fetching questions: {e}", None
|
71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
# 3. Run your Agent
|
73 |
results_log = []
|
74 |
answers_payload = []
|
@@ -138,7 +158,7 @@ def run_and_submit_all( profile: gr.OAuthProfile | None):
|
|
138 |
print(status_message)
|
139 |
results_df = pd.DataFrame(results_log)
|
140 |
return status_message, results_df
|
141 |
-
|
142 |
|
143 |
# --- Build Gradio Interface using Blocks ---
|
144 |
with gr.Blocks() as demo:
|
|
|
3 |
import requests
|
4 |
import inspect
|
5 |
import pandas as pd
|
6 |
+
from tools import intelligent_agent, get_file_type, write_bytes_to_temp_dir
|
7 |
|
8 |
# (Keep Constants as is)
|
9 |
# --- Constants ---
|
|
|
54 |
try:
|
55 |
response = requests.get(questions_url, timeout=15)
|
56 |
response.raise_for_status()
|
57 |
+
hf_questions = response.json()
|
58 |
+
if not hf_questions:
|
59 |
print("Fetched questions list is empty.")
|
60 |
return "Fetched questions list is empty or invalid format.", None
|
61 |
+
print(f"Fetched {len(hf_questions)} questions.")
|
62 |
except requests.exceptions.RequestException as e:
|
63 |
print(f"Error fetching questions: {e}")
|
64 |
return f"Error fetching questions: {e}", None
|
|
|
69 |
except Exception as e:
|
70 |
print(f"An unexpected error occurred fetching questions: {e}")
|
71 |
return f"An unexpected error occurred fetching questions: {e}", None
|
72 |
+
|
73 |
+
# 3. Create states
|
74 |
+
for item in hf_questions:
|
75 |
+
file_name = item.get('file_name', '')
|
76 |
+
if file_name == '':
|
77 |
+
item['input_file'] = None
|
78 |
+
item['file_type'] = None
|
79 |
+
item['file_path'] = None
|
80 |
+
else:
|
81 |
+
# Call the API to retrieve the file; adjust params as needed
|
82 |
+
task_id = item['task_id']
|
83 |
+
api_response = requests.get(f"{api_url}/{task_id}")
|
84 |
+
if api_response.status_code == 200:
|
85 |
+
item['input_file'] = api_response.content # Store file as bytes
|
86 |
+
item['file_type'] = get_file_type(file_name)
|
87 |
+
item['file_path'] = write_bytes_to_temp_dir(item['input_file'], file_name)
|
88 |
+
else:
|
89 |
+
item['input_file'] = None # Or handle error as needed
|
90 |
+
|
91 |
+
"""
|
92 |
# 3. Run your Agent
|
93 |
results_log = []
|
94 |
answers_payload = []
|
|
|
158 |
print(status_message)
|
159 |
results_df = pd.DataFrame(results_log)
|
160 |
return status_message, results_df
|
161 |
+
"""
|
162 |
|
163 |
# --- Build Gradio Interface using Blocks ---
|
164 |
with gr.Blocks() as demo:
|
tools.py
CHANGED
@@ -0,0 +1,601 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import spacy
|
3 |
+
import tempfile
|
4 |
+
import glob
|
5 |
+
import yt_dlp
|
6 |
+
import shutil
|
7 |
+
import cv2
|
8 |
+
import librosa
|
9 |
+
import wikipedia
|
10 |
+
|
11 |
+
from typing import TypedDict, List, Optional, Dict, Any
|
12 |
+
from langchain.docstore.document import Document
|
13 |
+
from langchain.prompts import PromptTemplate
|
14 |
+
from langchain_community.document_loaders import WikipediaLoader
|
15 |
+
from langgraph.graph import START, END, StateGraph
|
16 |
+
from langchain_core.messages import AnyMessage, HumanMessage, AIMessage # If you are using it
|
17 |
+
from langchain_community.retrievers import BM25Retriever # If you are using it
|
18 |
+
from langgraph.prebuilt import ToolNode, tools_condition # If you are using it
|
19 |
+
from langchain.vectorstores import FAISS
|
20 |
+
from langchain.embeddings import HuggingFaceEmbeddings
|
21 |
+
from langchain.schema import Document
|
22 |
+
from transformers import BlipProcessor, BlipForQuestionAnswering, pipeline
|
23 |
+
from io import BytesIO
|
24 |
+
from sentence_transformers import SentenceTransformer
|
25 |
+
|
26 |
+
nlp = spacy.load("en_core_web_sm")
|
27 |
+
|
28 |
+
# Define file extension sets for each category
|
29 |
+
PICTURE_EXTENSIONS = {'.jpg', '.jpeg', '.png', '.gif', '.bmp', '.tiff', '.webp'}
|
30 |
+
AUDIO_EXTENSIONS = {'.mp3', '.wav', '.aac', '.flac', '.ogg', '.m4a', '.wma'}
|
31 |
+
CODE_EXTENSIONS = {'.py', '.js', '.java', '.cpp', '.c', '.cs', '.rb', '.go', '.php', '.html', '.css', '.ts'}
|
32 |
+
SPREADSHEET_EXTENSIONS = {
|
33 |
+
'.xls', '.xlsx', '.xlsm', '.xlsb', '.xlt', '.xltx', '.xltm',
|
34 |
+
'.ods', '.ots', '.csv', '.tsv', '.sxc', '.stc', '.dif', '.gsheet',
|
35 |
+
'.numbers', '.numbers-tef', '.nmbtemplate', '.fods', '.123', '.wk1', '.wk2',
|
36 |
+
'.wks', '.wku', '.wr1', '.gnumeric', '.gnm', '.xml', '.pmvx', '.pmdx',
|
37 |
+
'.pmv', '.uos', '.txt'
|
38 |
+
}
|
39 |
+
|
40 |
+
def get_file_type(filename: str) -> str:
|
41 |
+
if not filename or '.' not in filename or filename == '':
|
42 |
+
return ''
|
43 |
+
ext = filename.lower().rsplit('.', 1)[-1]
|
44 |
+
dot_ext = f'.{ext}'
|
45 |
+
if dot_ext in PICTURE_EXTENSIONS:
|
46 |
+
return 'picture'
|
47 |
+
elif dot_ext in AUDIO_EXTENSIONS:
|
48 |
+
return 'audio'
|
49 |
+
elif dot_ext in CODE_EXTENSIONS:
|
50 |
+
return 'code'
|
51 |
+
elif dot_ext in SPREADSHEET_EXTENSIONS:
|
52 |
+
return 'spreadsheet'
|
53 |
+
else:
|
54 |
+
return 'unknown'
|
55 |
+
|
56 |
+
def write_bytes_to_temp_dir(file_bytes: bytes, file_name: str) -> str:
|
57 |
+
"""
|
58 |
+
Writes bytes to a file in the system temporary directory using the provided file_name.
|
59 |
+
Returns the full path to the saved file.
|
60 |
+
The file will persist until manually deleted or the OS cleans the temp directory.
|
61 |
+
"""
|
62 |
+
temp_dir = tempfile.gettempdir()
|
63 |
+
file_path = os.path.join(temp_dir, file_name)
|
64 |
+
with open(file_path, 'wb') as f:
|
65 |
+
f.write(file_bytes)
|
66 |
+
print(f"File written to: {file_path}")
|
67 |
+
return file_path
|
68 |
+
|
69 |
+
import os
|
70 |
+
import re
|
71 |
+
from PIL import Image # This is correctly imported, but was being used incorrectly
|
72 |
+
import numpy as np
|
73 |
+
from collections import Counter
|
74 |
+
import torch
|
75 |
+
from transformers import BlipProcessor, BlipForQuestionAnswering, pipeline
|
76 |
+
from typing import TypedDict, List, Optional, Dict, Any, Literal, Tuple
|
77 |
+
from langgraph.graph import StateGraph, START, END
|
78 |
+
from langchain.docstore.document import Document
|
79 |
+
|
80 |
+
# 1. Define the State type
|
81 |
+
class State(TypedDict, total=False):
|
82 |
+
question: str
|
83 |
+
task_id: str
|
84 |
+
input_file: bytes
|
85 |
+
file_type: str
|
86 |
+
context: List[Document] # Using LangChain's Document class
|
87 |
+
file_path: Optional[str]
|
88 |
+
youtube_url: Optional[str]
|
89 |
+
answer: Optional[str]
|
90 |
+
frame_answers: Optional[list]
|
91 |
+
next: Optional[str] # Added to track the next node
|
92 |
+
|
93 |
+
# --- LLM pipeline for general questions ---
|
94 |
+
llm_pipe = pipeline("text-generation",
|
95 |
+
#model="meta-llama/Llama-3.3-70B-Instruct",
|
96 |
+
#model="meta-llama/Meta-Llama-3-8B-Instruct",
|
97 |
+
#model="Qwen/Qwen2-7B-Instruct",
|
98 |
+
#model="microsoft/Phi-4-reasoning",
|
99 |
+
model="microsoft/Phi-3-mini-4k-instruct",
|
100 |
+
device_map="auto",
|
101 |
+
#device_map={ "": 0 }, # "" means the whole model
|
102 |
+
#max_memory={0: "10GiB"},
|
103 |
+
torch_dtype="auto",
|
104 |
+
max_new_tokens=256)
|
105 |
+
|
106 |
+
# Speech-to-text pipeline
|
107 |
+
asr_pipe = pipeline(
|
108 |
+
"automatic-speech-recognition",
|
109 |
+
model="openai/whisper-small",
|
110 |
+
device=-1
|
111 |
+
#device_map={"", 0},
|
112 |
+
#max_memory = {0: "4.5GiB"},
|
113 |
+
#device_map="auto"
|
114 |
+
)
|
115 |
+
|
116 |
+
# --- Your BLIP VQA setup ---
|
117 |
+
#device = "cuda" if torch.cuda.is_available() else "cpu"
|
118 |
+
device = "cpu"
|
119 |
+
vqa_model_name = "Salesforce/blip-vqa-base"
|
120 |
+
processor_vqa = BlipProcessor.from_pretrained(vqa_model_name)
|
121 |
+
|
122 |
+
# Attempt to load model to GPU; fall back to CPU if OOM
|
123 |
+
try:
|
124 |
+
model_vqa = BlipForQuestionAnswering.from_pretrained(vqa_model_name).to(device)
|
125 |
+
except torch.cuda.OutOfMemoryError:
|
126 |
+
print("WARNING: Loading model to CPU due to insufficient GPU memory.")
|
127 |
+
device = "cpu" # Switch device to CPU
|
128 |
+
model_vqa = BlipForQuestionAnswering.from_pretrained(vqa_model_name).to(device)
|
129 |
+
|
130 |
+
|
131 |
+
# --- Helper: Answer question on a single frame ---
|
132 |
+
def answer_question_on_frame(image_path, question):
|
133 |
+
# Fixed: Properly use the PIL Image module
|
134 |
+
image = Image.open(image_path).convert('RGB')
|
135 |
+
inputs = processor_vqa(image, question, return_tensors="pt").to(device)
|
136 |
+
out = model_vqa.generate(**inputs)
|
137 |
+
answer = processor_vqa.decode(out[0], skip_special_tokens=True)
|
138 |
+
return answer
|
139 |
+
|
140 |
+
# --- Helper: Answer question about the whole video ---
|
141 |
+
def answer_video_question(frames_dir, question):
|
142 |
+
valid_exts = ('.jpg', '.jpeg', '.png')
|
143 |
+
|
144 |
+
# Check if directory exists
|
145 |
+
if not os.path.exists(frames_dir):
|
146 |
+
return {
|
147 |
+
"most_common_answer": "No frames found to analyze.",
|
148 |
+
"all_answers": [],
|
149 |
+
"answer_counts": Counter()
|
150 |
+
}
|
151 |
+
|
152 |
+
frame_files = [os.path.join(frames_dir, f) for f in os.listdir(frames_dir)
|
153 |
+
if f.lower().endswith(valid_exts)]
|
154 |
+
|
155 |
+
# Sort frames properly by number
|
156 |
+
def get_frame_number(filename):
|
157 |
+
match = re.search(r'(\d+)', os.path.basename(filename))
|
158 |
+
return int(match.group(1)) if match else 0
|
159 |
+
|
160 |
+
frame_files = sorted(frame_files, key=get_frame_number)
|
161 |
+
|
162 |
+
if not frame_files:
|
163 |
+
return {
|
164 |
+
"most_common_answer": "No valid image frames found.",
|
165 |
+
"all_answers": [],
|
166 |
+
"answer_counts": Counter()
|
167 |
+
}
|
168 |
+
|
169 |
+
answers = []
|
170 |
+
for frame_path in frame_files:
|
171 |
+
try:
|
172 |
+
ans = answer_question_on_frame(frame_path, question)
|
173 |
+
answers.append(ans)
|
174 |
+
print(f"Processed frame: {os.path.basename(frame_path)}, Answer: {ans}")
|
175 |
+
except Exception as e:
|
176 |
+
print(f"Error processing frame {frame_path}: {str(e)}")
|
177 |
+
|
178 |
+
if not answers:
|
179 |
+
return {
|
180 |
+
"most_common_answer": "Could not analyze any frames successfully.",
|
181 |
+
"all_answers": [],
|
182 |
+
"answer_counts": Counter()
|
183 |
+
}
|
184 |
+
|
185 |
+
counted = Counter(answers)
|
186 |
+
most_common_answer, freq = counted.most_common(1)[0]
|
187 |
+
return {
|
188 |
+
"most_common_answer": most_common_answer,
|
189 |
+
"all_answers": answers,
|
190 |
+
"answer_counts": counted
|
191 |
+
}
|
192 |
+
|
193 |
+
|
194 |
+
def download_youtube_video(url, output_dir='/content/video/', output_filename='downloaded_video.mp4'):
|
195 |
+
# Ensure the output directory exists
|
196 |
+
os.makedirs(output_dir, exist_ok=True)
|
197 |
+
|
198 |
+
# Delete all files in the output directory
|
199 |
+
files = glob.glob(os.path.join(output_dir, '*'))
|
200 |
+
for f in files:
|
201 |
+
try:
|
202 |
+
os.remove(f)
|
203 |
+
except Exception as e:
|
204 |
+
print(f"Error deleting {f}: {str(e)}")
|
205 |
+
|
206 |
+
# Set output path for yt-dlp
|
207 |
+
output_path = os.path.join(output_dir, output_filename)
|
208 |
+
|
209 |
+
ydl_opts = {
|
210 |
+
'format': 'bestvideo[ext=mp4]+bestaudio[ext=m4a]/best[ext=mp4]/best',
|
211 |
+
'outtmpl': output_path,
|
212 |
+
'quiet': True,
|
213 |
+
'merge_output_format': 'mp4', # Ensures merged output is mp4
|
214 |
+
'postprocessors': [{
|
215 |
+
'key': 'FFmpegVideoConvertor',
|
216 |
+
'preferedformat': 'mp4', # Recode if needed
|
217 |
+
}]
|
218 |
+
}
|
219 |
+
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
|
220 |
+
ydl.download([url])
|
221 |
+
return output_path
|
222 |
+
|
223 |
+
|
224 |
+
|
225 |
+
# --- Helper: Extract frames from video ---
|
226 |
+
def extract_frames(video_path, output_dir, frame_interval_seconds=10):
|
227 |
+
# --- Clean output directory before extracting new frames ---
|
228 |
+
if os.path.exists(output_dir):
|
229 |
+
for filename in os.listdir(output_dir):
|
230 |
+
file_path = os.path.join(output_dir, filename)
|
231 |
+
try:
|
232 |
+
if os.path.isfile(file_path) or os.path.islink(file_path):
|
233 |
+
os.unlink(file_path)
|
234 |
+
elif os.path.isdir(file_path):
|
235 |
+
shutil.rmtree(file_path)
|
236 |
+
except Exception as e:
|
237 |
+
print(f'Failed to delete {file_path}. Reason: {e}')
|
238 |
+
else:
|
239 |
+
os.makedirs(output_dir, exist_ok=True)
|
240 |
+
|
241 |
+
try:
|
242 |
+
cap = cv2.VideoCapture(video_path)
|
243 |
+
if not cap.isOpened():
|
244 |
+
print("Error: Could not open video.")
|
245 |
+
return False
|
246 |
+
fps = cap.get(cv2.CAP_PROP_FPS)
|
247 |
+
frame_interval = int(fps * frame_interval_seconds)
|
248 |
+
count = 0
|
249 |
+
saved = 0
|
250 |
+
while True:
|
251 |
+
ret, frame = cap.read()
|
252 |
+
if not ret:
|
253 |
+
break
|
254 |
+
if count % frame_interval == 0:
|
255 |
+
frame_filename = os.path.join(output_dir, f"frame_{count:06d}.jpg")
|
256 |
+
cv2.imwrite(frame_filename, frame)
|
257 |
+
saved += 1
|
258 |
+
count += 1
|
259 |
+
cap.release()
|
260 |
+
print(f"Extracted {saved} frames.")
|
261 |
+
return saved > 0
|
262 |
+
except Exception as e:
|
263 |
+
print(f"Exception during frame extraction: {e}")
|
264 |
+
return False
|
265 |
+
|
266 |
+
def image_qa(image_path: str, question: str, model_name: str = vqa_model_name) -> str:
|
267 |
+
"""
|
268 |
+
Answers questions about images using Hugging Face's VQA pipeline.
|
269 |
+
|
270 |
+
Args:
|
271 |
+
image_path: Path to local image file or URL
|
272 |
+
question: Natural language question about the image
|
273 |
+
model_name: Pretrained VQA model (default: good general-purpose model)
|
274 |
+
|
275 |
+
Returns:
|
276 |
+
str: The model's best answer
|
277 |
+
"""
|
278 |
+
# Create VQA pipeline with specified model
|
279 |
+
vqa_pipeline = pipeline("visual-question-answering", model=model_name)
|
280 |
+
|
281 |
+
# Get predictions (automatically handles local files/URLs)
|
282 |
+
results = vqa_pipeline(image=image_path, question=question, top_k=1)
|
283 |
+
|
284 |
+
# Return top answer
|
285 |
+
return results[0]['answer']
|
286 |
+
|
287 |
+
|
288 |
+
def router(state: Dict[str, Any]) -> str:
|
289 |
+
"""Determine the next node based on whether the question contains a YouTube URL or references Wikipedia."""
|
290 |
+
question = state.get('question', '')
|
291 |
+
|
292 |
+
|
293 |
+
# Pattern for Wikipedia and similar sources
|
294 |
+
wiki_pattern = r"(wikipedia\.org|wiki|encyclopedia|britannica\.com|encyclop[a|æ]dia)"
|
295 |
+
has_wiki = re.search(wiki_pattern, question, re.IGNORECASE) is not None
|
296 |
+
|
297 |
+
# Pattern for YouTube
|
298 |
+
yt_pattern = r"(https?://)?(www\.)?(youtube\.com|youtu\.be)/[^\s]+"
|
299 |
+
has_youtube = re.search(yt_pattern, question) is not None
|
300 |
+
|
301 |
+
# Check for image
|
302 |
+
has_image = state.get('file_type') == 'picture'
|
303 |
+
|
304 |
+
# Check for audio
|
305 |
+
has_audio = state.get('file_type') == 'audio'
|
306 |
+
|
307 |
+
print(f"Has Wikipedia reference: {has_wiki}")
|
308 |
+
print(f"Has YouTube link: {has_youtube}")
|
309 |
+
print(f"Has picture file: {has_image}")
|
310 |
+
print(f"Has audio file: {has_audio}")
|
311 |
+
|
312 |
+
if has_wiki:
|
313 |
+
return "retrieve"
|
314 |
+
elif has_youtube:
|
315 |
+
# Store the extracted YouTube URL in the state
|
316 |
+
url_match = re.search(r"(https?://[^\s]+)", question)
|
317 |
+
if url_match:
|
318 |
+
state['youtube_url'] = url_match.group(0)
|
319 |
+
return "video"
|
320 |
+
elif has_image:
|
321 |
+
return "image"
|
322 |
+
elif has_audio:
|
323 |
+
return "audio"
|
324 |
+
else:
|
325 |
+
return "llm"
|
326 |
+
|
327 |
+
|
328 |
+
# --- Node Implementation ---
|
329 |
+
def node_image(state: Dict[str, Any]) -> Dict[str, Any]:
|
330 |
+
"""Router node that decides which node to go to next."""
|
331 |
+
print("Running node_image")
|
332 |
+
# Add the next state to the state dict
|
333 |
+
img = Image.open(state['file_path'])
|
334 |
+
state['answer'] = image_qa(state['file_path'], state['question'])
|
335 |
+
return state
|
336 |
+
|
337 |
+
|
338 |
+
def node_decide(state: Dict[str, Any]) -> Dict[str, Any]:
|
339 |
+
"""Router node that decides which node to go to next."""
|
340 |
+
print("Running node_decide")
|
341 |
+
# Add the next state to the state dict
|
342 |
+
state["next"] = router(state)
|
343 |
+
print(f"Routing to: {state['next']}")
|
344 |
+
return state
|
345 |
+
|
346 |
+
def node_video(state: Dict[str, Any]) -> Dict[str, Any]:
|
347 |
+
print("Running node_video")
|
348 |
+
youtube_url = state.get('youtube_url')
|
349 |
+
if not youtube_url:
|
350 |
+
state['answer'] = "No YouTube URL found in the question."
|
351 |
+
return state
|
352 |
+
|
353 |
+
question = state['question']
|
354 |
+
# Extract the actual question part (remove the URL)
|
355 |
+
question_text = re.sub(r'https?://[^\s]+', '', question).strip()
|
356 |
+
if not question_text.endswith('?'):
|
357 |
+
question_text += '?'
|
358 |
+
|
359 |
+
video_file = download_youtube_video(youtube_url)
|
360 |
+
if not video_file or not os.path.exists(video_file):
|
361 |
+
state['answer'] = "Failed to download the video."
|
362 |
+
return state
|
363 |
+
|
364 |
+
frames_dir = "/tmp/frames"
|
365 |
+
os.makedirs(frames_dir, exist_ok=True)
|
366 |
+
|
367 |
+
success = extract_frames(video_path=video_file, output_dir=frames_dir, frame_interval_seconds=10)
|
368 |
+
if not success:
|
369 |
+
state['answer'] = "Failed to extract frames from the video."
|
370 |
+
return state
|
371 |
+
|
372 |
+
result = answer_video_question(frames_dir, question_text)
|
373 |
+
state['answer'] = result['most_common_answer']
|
374 |
+
state['frame_answers'] = result['all_answers']
|
375 |
+
|
376 |
+
# Create Document objects for each frame analysis
|
377 |
+
frame_documents = []
|
378 |
+
for i, ans in enumerate(result['all_answers']):
|
379 |
+
doc = Document(
|
380 |
+
page_content=f"Frame {i}: {ans}",
|
381 |
+
metadata={"frame_number": i, "source": "video_analysis"}
|
382 |
+
)
|
383 |
+
frame_documents.append(doc)
|
384 |
+
|
385 |
+
# Add documents to state if not already present
|
386 |
+
if 'context' not in state:
|
387 |
+
state['context'] = []
|
388 |
+
state['context'].extend(frame_documents)
|
389 |
+
|
390 |
+
print(f"Video answer: {state['answer']}")
|
391 |
+
return state
|
392 |
+
|
393 |
+
def node_audio_rag(state: Dict[str, Any]) -> Dict[str, Any]:
|
394 |
+
print(f"Processing audio file: {state['file_path']}")
|
395 |
+
|
396 |
+
try:
|
397 |
+
# Step 1: Transcribe audio
|
398 |
+
audio, sr = librosa.load(state['file_path'], sr=16000)
|
399 |
+
asr_result = asr_pipe({"raw": audio, "sampling_rate": sr})
|
400 |
+
audio_transcript = asr_result['text']
|
401 |
+
print(f"Audio transcript: {audio_transcript}")
|
402 |
+
|
403 |
+
# Step 2: Store ONLY the transcript in the vector store
|
404 |
+
transcript_doc = [Document(page_content=audio_transcript)]
|
405 |
+
embeddings = HuggingFaceEmbeddings(model_name='BAAI/bge-large-en-v1.5')
|
406 |
+
vector_db = FAISS.from_documents(transcript_doc, embedding=embeddings)
|
407 |
+
|
408 |
+
# Step 3: Retrieve relevant docs for the user's question
|
409 |
+
question = state['question']
|
410 |
+
similar_docs = vector_db.similarity_search(question, k=1) # Only one doc in store
|
411 |
+
retrieved_context = "\n".join([doc.page_content for doc in similar_docs])
|
412 |
+
|
413 |
+
# Step 4: Augment prompt and generate answer
|
414 |
+
prompt = (
|
415 |
+
f"Use the following context to answer the question.\n"
|
416 |
+
f"Context:\n{retrieved_context}\n\n"
|
417 |
+
f"Question: {question}\nAnswer:"
|
418 |
+
)
|
419 |
+
llm_response = llm_pipe(prompt)
|
420 |
+
state['answer'] = llm_response[0]['generated_text']
|
421 |
+
|
422 |
+
except Exception as e:
|
423 |
+
error_msg = f"Audio processing error: {str(e)}"
|
424 |
+
print(error_msg)
|
425 |
+
state['answer'] = error_msg
|
426 |
+
|
427 |
+
return state
|
428 |
+
|
429 |
+
def node_llm(state: Dict[str, Any]) -> Dict[str, Any]:
|
430 |
+
print("Running node_llm")
|
431 |
+
question = state['question']
|
432 |
+
|
433 |
+
# Optionally add context from state (e.g., Wikipedia/Wikidata content)
|
434 |
+
context_text = ""
|
435 |
+
if 'article_content' in state and state['article_content']:
|
436 |
+
context_text = f"\n\nBackground Information:\n{state['article_content']}\n"
|
437 |
+
elif 'context' in state and state['context']:
|
438 |
+
context_text = "\n\n".join([doc.page_content for doc in state['context']])
|
439 |
+
|
440 |
+
# Compose a detailed prompt
|
441 |
+
prompt = (
|
442 |
+
"You are an expert researcher. Answer the user's question as accurately as possible. "
|
443 |
+
"If the text appears to be scrambled, try to unscramble the text for the user"
|
444 |
+
"If the information is incomplete or ambiguous, provide your best estimate based on the available evidence, and clearly explain any assumptions or reasoning you use. "
|
445 |
+
"If the answer requires multiple steps or deeper analysis, break down the question into sub-questions and answer them step by step, citing the relevant context for each step.\n\n"
|
446 |
+
f"Question: {question}"
|
447 |
+
f"{context_text}\n"
|
448 |
+
"Answer:"
|
449 |
+
)
|
450 |
+
|
451 |
+
# Add document to state for traceability
|
452 |
+
query_doc = Document(
|
453 |
+
page_content=prompt,
|
454 |
+
metadata={"source": "llm_prompt"}
|
455 |
+
)
|
456 |
+
if 'context' not in state:
|
457 |
+
state['context'] = []
|
458 |
+
state['context'].append(query_doc)
|
459 |
+
|
460 |
+
try:
|
461 |
+
result = llm_pipe(prompt)
|
462 |
+
state['answer'] = result[0]['generated_text']
|
463 |
+
except Exception as e:
|
464 |
+
print(f"Error in LLM processing: {str(e)}")
|
465 |
+
state['answer'] = f"An error occurred while processing your question: {str(e)}"
|
466 |
+
|
467 |
+
print(f"LLM answer: {state['answer']}")
|
468 |
+
return state
|
469 |
+
|
470 |
+
|
471 |
+
# --- Define the edge condition function ---
|
472 |
+
def get_next_node(state: Dict[str, Any]) -> str:
|
473 |
+
"""Get the next node from the state."""
|
474 |
+
return state["next"]
|
475 |
+
|
476 |
+
|
477 |
+
# 2. Improved Wikipedia Retrieval Node
|
478 |
+
def extract_keywords(question: str) -> List[str]:
|
479 |
+
doc = nlp(question)
|
480 |
+
keywords = [token.text for token in doc if token.pos_ in ("PROPN", "NOUN")] # Extract proper nouns and nouns
|
481 |
+
return keywords
|
482 |
+
|
483 |
+
def extract_entities(question: str) -> List[str]:
|
484 |
+
doc = nlp(question)
|
485 |
+
entities = [ent.text for ent in doc.ents]
|
486 |
+
return entities if entities else [token.text for token in doc if token.pos_ in ("PROPN", "NOUN")]
|
487 |
+
|
488 |
+
|
489 |
+
def retrieve(state: State) -> dict:
|
490 |
+
keywords = extract_entities(state["question"])
|
491 |
+
query = " ".join(keywords)
|
492 |
+
search_results = wikipedia.search(query)
|
493 |
+
selected_page = search_results[0] if search_results else None
|
494 |
+
|
495 |
+
if selected_page:
|
496 |
+
loader = WikipediaLoader(
|
497 |
+
query=selected_page,
|
498 |
+
lang="en",
|
499 |
+
load_max_docs=1,
|
500 |
+
doc_content_chars_max=100000,
|
501 |
+
load_all_available_meta=True
|
502 |
+
)
|
503 |
+
docs = loader.load()
|
504 |
+
# Chunk the article for finer retrieval
|
505 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
506 |
+
splitter = RecursiveCharacterTextSplitter(chunk_size=1500, chunk_overlap=200)
|
507 |
+
all_chunks = []
|
508 |
+
for doc in docs:
|
509 |
+
chunks = splitter.split_text(doc.page_content)
|
510 |
+
all_chunks.extend([Document(page_content=chunk) for chunk in chunks])
|
511 |
+
# Optionally: re-rank or filter chunks here
|
512 |
+
return {"context": all_chunks}
|
513 |
+
else:
|
514 |
+
return {"context": []}
|
515 |
+
|
516 |
+
# 3. Prompt Template for General QA
|
517 |
+
prompt = PromptTemplate(
|
518 |
+
input_variables=["question", "context"],
|
519 |
+
template=(
|
520 |
+
"You are an expert researcher. Given the following context from Wikipedia, answer the user's question as accurately as possible. "
|
521 |
+
"If the text appears to be scrambled, try to unscramble the text for the user"
|
522 |
+
"If the information is incomplete or ambiguous, provide your best estimate based on the available evidence, and clearly explain any assumptions or reasoning you use. "
|
523 |
+
"If the answer requires multiple steps or deeper analysis, break down the question into sub-questions and answer them step by step, citing the relevant context for each step."
|
524 |
+
"Context:\n{context}\n\n"
|
525 |
+
"Question: {question}\n\n"
|
526 |
+
"Best Estimate Answer:"
|
527 |
+
)
|
528 |
+
)
|
529 |
+
|
530 |
+
"""
|
531 |
+
def generate(state: State) -> dict:
|
532 |
+
# Concatenate all context documents into a single string
|
533 |
+
docs_content = "\n\n".join(doc.page_content for doc in state["context"])
|
534 |
+
# Format the prompt for the LLM
|
535 |
+
prompt_str = prompt.format(question=state["question"], context=docs_content)
|
536 |
+
# Generate answer
|
537 |
+
response = llm.invoke(prompt_str)
|
538 |
+
return {"answer": response}
|
539 |
+
"""
|
540 |
+
|
541 |
+
def generate(state: dict) -> dict:
|
542 |
+
# Concatenate all context documents into a single string
|
543 |
+
docs_content = "\n\n".join(doc.page_content for doc in state["context"])
|
544 |
+
# Format the prompt for the LLM
|
545 |
+
prompt_str = prompt.format(question=state["question"], context=docs_content)
|
546 |
+
# Generate answer using Hugging Face pipeline
|
547 |
+
response = llm_pipe(prompt_str)
|
548 |
+
# Extract generated text
|
549 |
+
answer = response[0]["generated_text"]
|
550 |
+
return {"answer": answer}
|
551 |
+
|
552 |
+
# Create the StateGraph
|
553 |
+
graph = StateGraph(State)
|
554 |
+
|
555 |
+
# Add nodes
|
556 |
+
graph.add_node("decide", node_decide)
|
557 |
+
graph.add_node("video", node_video)
|
558 |
+
graph.add_node("llm", node_llm)
|
559 |
+
graph.add_node("retrieve", retrieve)
|
560 |
+
graph.add_node("generate", generate)
|
561 |
+
graph.add_node("image", node_image)
|
562 |
+
graph.add_node("audio", node_audio_rag)
|
563 |
+
|
564 |
+
# Add edge from START to decide
|
565 |
+
graph.add_edge(START, "decide")
|
566 |
+
graph.add_edge("retrieve", "generate")
|
567 |
+
|
568 |
+
# Add conditional edges from decide to video or llm based on question
|
569 |
+
graph.add_conditional_edges(
|
570 |
+
"decide",
|
571 |
+
get_next_node,
|
572 |
+
{
|
573 |
+
"video": "video",
|
574 |
+
"llm": "llm",
|
575 |
+
"retrieve": "retrieve",
|
576 |
+
"image": "image",
|
577 |
+
"audio": "audio"
|
578 |
+
}
|
579 |
+
)
|
580 |
+
|
581 |
+
# Add edges from video and llm to END to terminate the graph
|
582 |
+
graph.add_edge("video", END)
|
583 |
+
graph.add_edge("llm", END)
|
584 |
+
graph.add_edge("generate", END)
|
585 |
+
graph.add_edge("image", END)
|
586 |
+
graph.add_edge("audio", END)
|
587 |
+
|
588 |
+
# Compile the graph
|
589 |
+
agent = graph.compile()
|
590 |
+
|
591 |
+
# --- Usage Example ---
|
592 |
+
def intelligent_agent(state: State) -> str:
|
593 |
+
"""Process a question using the appropriate pipeline based on content."""
|
594 |
+
#state = State(question= question)
|
595 |
+
try:
|
596 |
+
final_state = agent.invoke(state)
|
597 |
+
return final_state.get('answer', "No answer found.")
|
598 |
+
except Exception as e:
|
599 |
+
print(f"Error in agent execution: {str(e)}")
|
600 |
+
return f"An error occurred: {str(e)}"
|
601 |
+
|