project / app.py
anomaly05's picture
Update app.py
45024a1 verified
Raw History Blame Contribute Delete
6.38 kB
import os
import numpy as np
from PyPDF2 import PdfReader
import docx
import faiss
import torch
from sentence_transformers import SentenceTransformer
from transformers import pipeline, AutoTokenizer
import gradio as gr
###########################
# CONFIGURATION
###########################
DATA_DIR = "." # Looks in the Space/project root
MAX_CHUNKS = 10 # Conservative for Spaces, raises as needed
MAX_RESPONSE_CHARS = 4000 # Prevents protocol errors by truncating huge responses
###########################
# DATA EXTRACTION & CHUNKING
###########################
def extract_text(file_path):
ext = os.path.splitext(file_path)[1].lower()
if ext == '.pdf':
text = ''
with open(file_path, 'rb') as f:
reader = PdfReader(f)
for page in reader.pages:
t = page.extract_text()
if t: text += t + ' '
return text
elif ext == '.docx':
docf = docx.Document(file_path)
return ' '.join([p.text for p in docf.paragraphs])
elif ext == '.txt':
with open(file_path, 'r', encoding='utf-8') as f:
return f.read()
return ''
def chunk_text(text, max_len=300):
# Smaller chunks are safer for Spaces
sents = text.split('. ')
chunks, chunk = [], ''
for s in sents:
if len(chunk) + len(s) < max_len:
chunk += s + '. '
else:
if chunk.strip():
chunks.append(chunk.strip())
chunk = s + '. '
if chunk.strip():
chunks.append(chunk.strip())
return chunks
all_chunks, all_meta = [], []
filelist = [f for f in os.listdir(DATA_DIR) if f.lower().endswith(('.pdf', '.docx', '.txt'))]
if filelist:
for fname in filelist:
file_path = os.path.join(DATA_DIR, fname)
subject = 'general'
try:
text = extract_text(file_path)
c = chunk_text(text, max_len=300)
for chunk in c:
# Only index nonempty, reasonbly sized chunks
if 40 < len(chunk) <= 1000:
all_chunks.append(chunk)
all_meta.append({'subject': subject, 'file': fname})
if len(all_chunks) >= MAX_CHUNKS:
break
if len(all_chunks) >= MAX_CHUNKS:
break
except Exception as e:
print(f"Failed {file_path}: {e}")
if len(all_chunks) >= MAX_CHUNKS:
break
print('Chunks loaded:', len(all_chunks))
else:
print("No textbook files found in current directory.")
###########################
# EMBEDDINGS & INDEX
###########################
if all_chunks:
encoder = SentenceTransformer('all-MiniLM-L6-v2')
embeddings = encoder.encode(all_chunks, show_progress_bar=False, batch_size=8)
faiss.normalize_L2(embeddings)
index = faiss.IndexFlatIP(embeddings.shape[1])
index.add(np.array(embeddings).astype('float32'))
else:
encoder, embeddings, index = None, None, None
###########################
# ANSWER FUNCTION
###########################
def answer_query(
question,
subject=None,
top_k=1, # One chunk per answer is safest!
model_name="Qwen/Qwen2-0.5B"
):
if encoder is None or index is None or not all_chunks:
return "⚠️ No textbooks found or processed. Please check your file uploads."
# Retrieval, only the most relevant chunk
q_emb = encoder.encode([question])
faiss.normalize_L2(q_emb)
D, I = index.search(np.asarray(q_emb, dtype='float32'), top_k*2)
found = []
for idx in I[0]:
if subject:
if subject.lower() in all_meta[idx]['subject'].lower():
found.append(all_chunks[idx])
else:
found.append(all_chunks[idx])
if len(found) >= top_k: break
context = "\n".join(found)
if not context.strip():
return "Sorry, I could not find an answer from the loaded textbooks."
prompt = (
"You are an ICSE board AI tutor. Answer using ONLY the following textbook excerpt. "
"Answer concisely and do not repeat the prompt or content.\n\n"
f"Textbook excerpt:\n{context}\n\n"
f"Student question: {question}\n"
"Answer:"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
pipe = pipeline(
"text-generation",
model=model_name,
tokenizer=tokenizer,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
device=0 if torch.cuda.is_available() else -1,
max_new_tokens=128,
do_sample=False,
pad_token_id=tokenizer.eos_token_id
)
output = pipe(prompt)[0]['generated_text']
# Clean output: only what's after the last "Answer:" line
answer_part = output.split("Answer:")[-1].strip()
# Truncate at first re-ask or echo marker
for stop_word in ["Textbook", "Student question:", "Context:"]:
if stop_word in answer_part:
answer_part = answer_part.split(stop_word)[0].strip()
# Remove duplicate lines and enforce length limit
lines = answer_part.splitlines()
seen, unique = set(), []
for l in lines:
l = l.strip()
if l and l not in seen:
unique.append(l)
seen.add(l)
final_answer = "\n".join(unique)
if not final_answer.strip():
return "Sorry, I could not generate an answer from the content."
if len(final_answer) > MAX_RESPONSE_CHARS:
final_answer = final_answer[:MAX_RESPONSE_CHARS] + "\n\n[Truncated]"
return final_answer
###########################
# GRADIO APP
###########################
def gradio_chatbot(user_input, subject=""):
subject_input = subject.strip() if subject.strip() else None
return answer_query(user_input, subject=subject_input)
TITLE = "ICSE Textbook Chatbot (Qwen2-0.5B)"
DESC = "Ask any question using ONLY your uploaded ICSE textbooks. The AI answers strictly from your syllabus content. For demo use only."
iface = gr.Interface(
fn=gradio_chatbot,
inputs=[
gr.Textbox(label="Your question", placeholder="Type your question here...", lines=2),
gr.Textbox(label="Subject (optional)", placeholder="e.g., Physics, Geography", lines=1)
],
outputs="text",
title=TITLE,
description=DESC,
allow_flagging="never"
)
if __name__ == "__main__":
iface.launch(share=True)