Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
import os
|
3 |
+
from PIL import Image
|
4 |
+
import pytesseract
|
5 |
+
from pdf2image import convert_from_path
|
6 |
+
from langchain_community.embeddings import HuggingFaceEmbeddings
|
7 |
+
from langchain.prompts import PromptTemplate
|
8 |
+
from langchain.chains import RetrievalQA
|
9 |
+
from langchain.memory import ConversationBufferMemory
|
10 |
+
from langchain_groq import ChatGroq
|
11 |
+
from langchain_community.vectorstores import FAISS
|
12 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
13 |
+
|
14 |
+
def setup_environment():
|
15 |
+
os.environ["GROQ_API_KEY"] = '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'
|
16 |
+
|
17 |
+
# Define OCR functions for image and PDF files
|
18 |
+
def ocr_image(image_path, language='eng+guj'):
|
19 |
+
img = Image.open(image_path)
|
20 |
+
text = pytesseract.image_to_string(img, lang=language)
|
21 |
+
return text
|
22 |
+
|
23 |
+
def ocr_pdf(pdf_path, language='eng+guj'):
|
24 |
+
images = convert_from_path(pdf_path)
|
25 |
+
all_text = ""
|
26 |
+
for img in images:
|
27 |
+
text = pytesseract.image_to_string(img, lang=language)
|
28 |
+
all_text += text + "\n"
|
29 |
+
return all_text
|
30 |
+
|
31 |
+
def ocr_file(file_path):
|
32 |
+
file_extension = os.path.splitext(file_path)[1].lower()
|
33 |
+
if file_extension == ".pdf":
|
34 |
+
text_re = ocr_pdf(file_path, language='guj+eng')
|
35 |
+
elif file_extension in [".jpg", ".jpeg", ".png", ".bmp"]:
|
36 |
+
text_re = ocr_image(file_path, language='guj+eng')
|
37 |
+
else:
|
38 |
+
raise ValueError("Unsupported file format. Supported formats are PDF, JPG, JPEG, PNG, BMP.")
|
39 |
+
return text_re
|
40 |
+
|
41 |
+
def get_text_chunks(text):
|
42 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=500, chunk_overlap=100)
|
43 |
+
chunks = text_splitter.split_text(text)
|
44 |
+
return chunks
|
45 |
+
|
46 |
+
def get_vector_store(text_chunks):
|
47 |
+
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2", model_kwargs={'device': 'cpu'}, encode_kwargs={'normalize_embeddings': True})
|
48 |
+
vector_store = FAISS.from_texts(text_chunks, embedding=embeddings)
|
49 |
+
os.makedirs("faiss_index", exist_ok=True)
|
50 |
+
vector_store.save_local("faiss_index")
|
51 |
+
return vector_store
|
52 |
+
|
53 |
+
def process_ocr_and_pdf_files(file_paths):
|
54 |
+
raw_text = ""
|
55 |
+
for file_path in file_paths:
|
56 |
+
raw_text += ocr_file(file_path) + "\n"
|
57 |
+
text_chunks = get_text_chunks(raw_text)
|
58 |
+
return get_vector_store(text_chunks)
|
59 |
+
|
60 |
+
def get_conversational_chain():
|
61 |
+
template = """You are an intelligent educational assistant specialized in handling queries about documents. You have been provided with OCR-processed text from the uploaded files that contains important educational information.
|
62 |
+
|
63 |
+
Core Responsibilities:
|
64 |
+
1. Language Processing:
|
65 |
+
- Identify the language of the user's query (English or Gujarati)
|
66 |
+
- Respond in the same language as the query
|
67 |
+
- If the query is in Gujarati, ensure the response maintains proper Gujarati grammar and terminology
|
68 |
+
- For technical terms, provide both English and Gujarati versions when relevant
|
69 |
+
|
70 |
+
2. Document Understanding:
|
71 |
+
- Analyze the OCR-processed text from the uploaded files
|
72 |
+
- Account for potential OCR errors or misinterpretations
|
73 |
+
- Focus on extracting accurate information despite possible OCR imperfections
|
74 |
+
|
75 |
+
3. Response Guidelines:
|
76 |
+
- Provide direct, clear answers based solely on the document content
|
77 |
+
- If information is unclear due to OCR quality, mention this limitation
|
78 |
+
- For numerical data (dates, percentages, marks), double-check accuracy before responding
|
79 |
+
- If information is not found in the documents, clearly state: \"This information is not present in the uploaded documents\"
|
80 |
+
|
81 |
+
4. Educational Context:
|
82 |
+
- Maintain focus on educational queries related to the document content
|
83 |
+
- For admission-related queries, emphasize important deadlines and requirements
|
84 |
+
- For scholarship information, highlight eligibility criteria and application processes
|
85 |
+
- For course-related queries, provide detailed, accurate information from the documents
|
86 |
+
|
87 |
+
5. Response Format:
|
88 |
+
- Structure responses clearly with relevant subpoints when necessary
|
89 |
+
- For complex information, break down the answer into digestible parts
|
90 |
+
- Include relevant reference points from the documents when applicable
|
91 |
+
- Format numerical data and dates clearly
|
92 |
+
|
93 |
+
6. Quality Control:
|
94 |
+
- Verify that responses align with the document content
|
95 |
+
- Don't make assumptions beyond the provided information
|
96 |
+
- If multiple interpretations are possible due to OCR quality, mention all possibilities
|
97 |
+
- Maintain consistency in terminology throughout the conversation
|
98 |
+
|
99 |
+
Important Rules:
|
100 |
+
- Never make up information not present in the documents
|
101 |
+
- Don't combine information from previous conversations or external knowledge
|
102 |
+
- Always indicate if certain parts of the documents are unclear due to OCR quality
|
103 |
+
- Maintain professional tone while being accessible to students and parents
|
104 |
+
- If the query is out of scope of the uploaded documents, politely redirect to relevant official sources
|
105 |
+
|
106 |
+
Context from uploaded documents:
|
107 |
+
{context}
|
108 |
+
|
109 |
+
Chat History:
|
110 |
+
{history}
|
111 |
+
|
112 |
+
Current Question: {question}
|
113 |
+
Assistant: Let me provide a clear and accurate response based on the uploaded documents..."""
|
114 |
+
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/paraphrase-MiniLM-L6-v2", model_kwargs={'device': 'cpu'}, encode_kwargs={'normalize_embeddings': True})
|
115 |
+
new_vector_store = FAISS.load_local(
|
116 |
+
"faiss_index", embeddings, allow_dangerous_deserialization=True
|
117 |
+
)
|
118 |
+
QA_CHAIN_PROMPT = PromptTemplate(input_variables=["history", "context", "question"], template=template)
|
119 |
+
qa_chain = RetrievalQA.from_chain_type(llm, retriever=new_vector_store.as_retriever(), chain_type='stuff', verbose=True, chain_type_kwargs={"verbose": True,"prompt": QA_CHAIN_PROMPT,"memory": ConversationBufferMemory(memory_key="history",input_key="question"),})
|
120 |
+
return qa_chain
|
121 |
+
|
122 |
+
def user_input(user_question):
|
123 |
+
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2", model_kwargs={'device': 'cpu'}, encode_kwargs={'normalize_embeddings': True})
|
124 |
+
new_db = FAISS.load_local("faiss_index", embeddings, allow_dangerous_deserialization=True)
|
125 |
+
docs = new_db.similarity_search(user_question)
|
126 |
+
chain = get_conversational_chain()
|
127 |
+
response = chain({"input_documents": docs, "query": user_question}, return_only_outputs=True)
|
128 |
+
return response.get("result", "No result found")
|
129 |
+
|
130 |
+
def gradio_interface():
|
131 |
+
def process_files(files):
|
132 |
+
file_paths = []
|
133 |
+
for file in files:
|
134 |
+
file_path = os.path.join("temp", file.name)
|
135 |
+
os.makedirs(os.path.dirname(file_path), exist_ok=True)
|
136 |
+
with open(file_path, "wb") as f:
|
137 |
+
f.write(file.read())
|
138 |
+
file_paths.append(file_path)
|
139 |
+
process_ocr_and_pdf_files(file_paths)
|
140 |
+
return "Files processed and vector store updated!"
|
141 |
+
|
142 |
+
def ask_question(user_question):
|
143 |
+
return user_input(user_question)
|
144 |
+
|
145 |
+
file_upload = gr.inputs.File(label="Upload Files", type="file", multiple=True)
|
146 |
+
text_input = gr.inputs.Textbox(label="Ask a question related to the uploaded documents:")
|
147 |
+
|
148 |
+
outputs = [gr.outputs.Textbox(label="Output"), gr.outputs.Textbox(label="Conversation History")]
|
149 |
+
interface = gr.Interface(
|
150 |
+
fn=[process_files, ask_question],
|
151 |
+
inputs=[file_upload, text_input],
|
152 |
+
outputs=outputs,
|
153 |
+
live=True
|
154 |
+
)
|
155 |
+
interface.launch()
|
156 |
+
|
157 |
+
if __name__ == "__main__":
|
158 |
+
setup_environment()
|
159 |
+
gradio_interface()
|