AK1239's picture
feat: updated prompt with better instructions
8a1b729
import os
from transformers import pipeline
import torch
import nltk
from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
import fitz
from sentence_transformers import SentenceTransformer
import faiss
import numpy as np
import pickle
import re
import logging
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
import uvicorn
import asyncio
from config import (
ALL_FILES,
MATH_FILES,
SCIENCE_FILES,
DATA_DIR,
DOCUMENTS_PATH,
FAISS_INDEX_PATH,
HUGGINGFACE_TOKEN,
MODEL_ID
)
app = FastAPI(title="Swahili Content Generation API")
# Configure CORS
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
class PromptRequest(BaseModel):
prompt: str
class ContentRequest(BaseModel):
grade: int
subject: str
topic: str
style: str = "normal"
TOPIC_KEYWORDS = {
# Grade 3 Science
'mazingira g3.pdf': ['mazingira'],
'nishati g3.pdf': ['nishati'],
'maada g3.pdf': ['maada'],
'mawasiliano g3.pdf': ['mawasiliano'],
'usafi g3.pdf': ['usafi'],
'vipimo g3.pdf': ['vipimo-s'],
'mlo g3.pdf': ['mlo'],
'mfumo g3.pdf': ['mfumo'],
'maambukizi g3.pdf': ['maambukizi'],
'huduma g3.pdf': ['huduma'],
'vifaa g3.pdf': ['vifaa'],
# Grade 4 Science
'kinga ya mwili g4.txt': ['kinga'],
'magonjwa g4.txt': ['magonjwa'],
'majaribio ya kisayansi g4.txt': ['majaribio'],
'maji g4.txt': ['maji'],
'ukimwi g4.txt': ['ukimwi'],
'huduma g4.txt': ['huduma-g4'],
'mazingira g4.txt': ['mazingira-g4'],
'matumizi ya nishati g4.txt': ['matumizi-nishati-g4'],
'nishati g4.txt': ['nishati-g4'],
'mfumo g4.txt': ['mfumo-g4'],
'mawasiliano g4.txt': ['mawasiliano-g4'],
# MATH TOPICS Grade 3
'namba g3.txt': ['namba'],
'mpangilio g3.txt': ['mpangilio'],
'matendo katika namba g3.txt': ['matendo'],
'kutambua sehemu g3.txt': ['sehemu'],
'kutambua maumbo g3.txt': ['maumbo'],
'vipimo g3.txt': ['vipimo'],
'fedha g3.txt': ['fedha'],
'takwimu kwa picha g3.txt': ['takwimu'],
# MATH TOPICS Grade 4
'kugawanya namba g4.txt': ['kugawanya'],
'kujumlisha namba g4.txt': ['kujumlisha'],
'kuzidisha namba g4.txt': ['kuzidisha'],
'namba nzima g4.txt': ['namba-g4'],
'namba za kirumi g4.txt': ['kirumi'],
'wakati g4.txt': ['wakati'],
'mpangilio g4.txt': ['mpangilio-g4'],
'vipimo g4.txt': ['vipimo-g4'],
'takwimu g4.txt': ['takwimu-g4'],
'kutoa namba g4.txt': ['kutoa'],
'fedha g4.txt': ['fedha-g4'],
'sehemu g4.txt': ['sehemu-g4'],
'maumbo g4.txt': ['maumbo-g4']
}
def preprocess_pdf_text(text):
words_to_remove = ['FOR', 'ONLINE', 'USE', 'ONLY', 'DO', 'NOT', 'DUPLICATE', 'SAYANSI', 'STD', 'PM']
pattern = r'\b(?:' + '|'.join(map(re.escape, words_to_remove)) + r')\b'
text = re.sub(pattern, '', text, flags=re.IGNORECASE)
text = ' '.join(text.split())
text = re.sub(r'[^\w\s\.\,\?\!\'\"àèìòùÀÈÌÒÙáéíóúÁÉÍÓÚâêîôûÂÊÎÔÛãẽĩõũÃẼĨÕŨ]', ' ', text)
text = ' '.join(text.split())
return text
def extract_text_from_file(file_path):
if file_path.lower().endswith('.pdf'):
return extract_text_from_pdf(file_path)
elif file_path.lower().endswith('.txt'):
try:
with open(file_path, 'r', encoding='utf-8') as file:
text = file.read()
return text.strip()
except Exception as e:
logging.error(f"Error reading text file {file_path}: {str(e)}")
return ""
else:
logging.error(f"Unsupported file type for {file_path}")
return ""
def extract_text_from_pdf(pdf_path):
text = ""
with fitz.open(pdf_path) as doc:
for page_num, page in enumerate(doc):
try:
blocks = page.get_text("blocks")
page_text = "\n".join(block[4] for block in blocks)
cleaned_text = preprocess_pdf_text(page_text)
text += cleaned_text + "\n"
except Exception as e:
logging.error(f"Error processing page {page_num + 1}: {str(e)}")
continue
return text.strip()
def split_text_into_chunks(text, source_file, chunk_size=500, overlap=50):
# Clean the text
text = text.strip().replace('\n', ' ').replace(' ', ' ')
# Get filename and keywords
filename = os.path.basename(source_file)
keywords = TOPIC_KEYWORDS.get(filename, [])
# Use NLTK for better sentence tokenization
sentences = nltk.sent_tokenize(text)
chunks = []
current_chunk = []
current_size = 0
for sentence in sentences:
sentence_words = len(sentence.split())
if current_size + sentence_words > chunk_size:
if current_chunk:
# Create chunk with metadata
chunk_text = ' '.join(current_chunk)
chunk_info = {
'text': chunk_text,
'source': filename,
'keywords': keywords
}
chunks.append(chunk_info)
# Calculate overlap
overlap_size = 0
overlap_chunk = []
for s in reversed(current_chunk):
if overlap_size + len(s.split()) <= overlap:
overlap_chunk.insert(0, s)
overlap_size += len(s.split())
else:
break
current_chunk = overlap_chunk
current_size = overlap_size
current_chunk.append(sentence)
current_size += sentence_words
if current_chunk:
chunk_text = ' '.join(current_chunk)
chunks.append({
'text': chunk_text,
'source': filename,
'keywords': keywords
})
return chunks
def create_faiss_index(texts, embedding_model):
doc_embeddings = embedding_model.encode(texts)
index = faiss.IndexFlatL2(doc_embeddings.shape[1])
index.add(np.array(doc_embeddings))
return index
def retrieve_documents(query, index, embedding_model, documents, top_k=5):
query_lower = query.lower()
target_topic = None
# Simple direct keyword matching since we only have one keyword per topic
for filename, keywords in TOPIC_KEYWORDS.items():
if keywords[0] == query_lower:
target_topic = filename
break
# Get embeddings and search
query_embedding = embedding_model.encode([query])
distances, indices = index.search(query_embedding, top_k * 3)
# Filter and organize retrieved documents
topic_docs = []
for idx in indices[0]:
doc = documents[idx]
if doc['source'] == target_topic:
# Check if content is not too repetitive
if not any(existing.get('text', '') == doc['text'] for existing in topic_docs):
topic_docs.append(doc)
if len(topic_docs) >= top_k:
break
final_content = "\n\n".join(doc['text'] for doc in topic_docs[:top_k])
logger.info(f"Retrieved content from: {target_topic}")
return final_content
def calculate_bleu(reference, candidate):
"""
Calculate BLEU score between reference and candidate texts.
"""
if isinstance(reference, list):
reference = " ".join(reference)
if isinstance(candidate, list):
candidate = " ".join(candidate)
reference_tokens = [reference.split()]
candidate_tokens = candidate.split()
smoothing = SmoothingFunction().method1
return sentence_bleu(reference_tokens, candidate_tokens, smoothing_function=smoothing)
def get_topic_files(grade: int, subject: str, topic: str):
# Convert topic to lowercase for case-insensitive matching
topic_lower = topic.lower()
# Get the appropriate file list
file_list = MATH_FILES if subject.lower() == "math" else SCIENCE_FILES
# Filter files by grade and topic
matching_files = []
for file in file_list:
if f"g{grade}" in file.lower(): # Check grade
filename = os.path.basename(file)
if filename in TOPIC_KEYWORDS: # Check if file is in our topics
keywords = TOPIC_KEYWORDS[filename]
if topic_lower == keywords[0]:
matching_files.append(file)
return matching_files
def generate_response_with_rag(prompt, index, embedding_model, documents, settings):
# Retrieve relevant documents
retrieved_context = retrieve_documents(prompt, index, embedding_model, documents)
# Log the retrieved context
logger.info("Context sent to model:")
logger.info("-" * 50)
logger.info(retrieved_context)
logger.info("-" * 50)
style_instructions = {
"simple": "Provide clear and easy-to-understand answers using common words and short sentences. Explain concepts as if talking to a young student.",
"creative": "Give creative and engaging answers, using real-life examples and illustrations to make the content interesting and memorable.",
"normal": ""
}
instruction = style_instructions.get(settings.get("style", "normal"), "")
# Create system prompt
system_prompt = f"""
Explain the topic of "{settings['topic']}" in detail following this structure:
1. Summary: Briefly explain what the student will learn in this topic (5-6 sentences).
2. Introduction to the topic: Provide background information about the topic before breaking it down into subtopics.
3. Subtopics: Explain each subtopic in detail, providing real-life examples where necessary. For each subtopic, Describe images that could help explain the topic in detail using text instead of actual images.
Use this format: [Picture: Image description]. Dont provide more than 3 [Picture: Image description].
4. Activities: After each subtopic, provide small exercises or activities that the student can do to enhance understanding (Activities).
5. Practice questions: Provide 6-8 questions related to the topic to reinforce the student's understanding.
**Respond to all questions and instructions in Swahili.**
{instruction}
Context:
{retrieved_context}
"""
# Generate response from the model
messages = [{"role": "system", "content": system_prompt}]
outputs = app.state.pipe(messages, max_new_tokens=2000)
try:
# Extract the generated text from pipeline output
if not outputs or len(outputs) == 0:
logger.error("No output generated")
return {
"content": "Failed to generate response",
"context": retrieved_context
}
generated_messages = outputs[0]['generated_text']
if isinstance(generated_messages, list):
# Find the assistant's message
for message in generated_messages:
if message.get('role') == 'assistant':
response_content = message.get('content', '')
break
else:
logger.error("No assistant response found in messages")
return {
"content": "Failed to generate response",
"context": retrieved_context
}
else:
response_content = generated_messages
if not response_content:
logger.error("Empty response content")
return {
"content": "Failed to generate response",
"context": retrieved_context
}
# Clean up the response
response_content = response_content.strip()
# Split text into paragraphs and ensure proper spacing
paragraphs = [p.strip() for p in response_content.split('\n\n') if p.strip()]
# Handle single-line paragraphs that should be split
formatted_paragraphs = []
for paragraph in paragraphs:
# If a paragraph is too long (more than 100 chars) and doesn't have proper line breaks,
# split it into sentences and add line breaks
if len(paragraph) > 100 and '\n' not in paragraph:
sentences = [s.strip() for s in nltk.sent_tokenize(paragraph)]
formatted_paragraphs.append('\n'.join(sentences))
else:
formatted_paragraphs.append(paragraph)
# Join paragraphs with double line breaks and convert to HTML breaks
response_content = '\n\n'.join(formatted_paragraphs)
response_content = response_content.replace('\n', '<br>')
return {
"content": response_content,
"context": retrieved_context
}
except Exception as e:
logger.error(f"Error processing response: {e}")
logger.error(f"Raw output: {outputs}")
return {
"content": "Error processing response",
"context": retrieved_context
}
async def load_or_create_index():
embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
os.makedirs(DATA_DIR, exist_ok=True)
os.makedirs(os.path.dirname(FAISS_INDEX_PATH), exist_ok=True)
try:
with open(DOCUMENTS_PATH, 'rb') as f:
documents = pickle.load(f)
index = faiss.read_index(FAISS_INDEX_PATH)
print("FAISS index and documents loaded successfully.")
return index, documents, embedding_model
except FileNotFoundError:
print("Index and documents not found. Proceeding to create them.")
documents = []
# Process all files (both PDFs and TXTs)
files_found = False
for file_path in ALL_FILES:
if not os.path.exists(file_path):
logger.warning(f"File not found: {file_path}")
continue
filename = os.path.basename(file_path)
logging.info(f"Processing {filename}")
text = extract_text_from_file(file_path)
if text:
files_found = True
chunks = split_text_into_chunks(text, filename)
documents.extend(chunks)
await asyncio.sleep(0)
if not files_found:
raise Exception(f"No valid text or PDF files found in the specified paths")
texts = [doc['text'] for doc in documents]
index = create_faiss_index(texts, embedding_model)
os.makedirs(os.path.dirname(DOCUMENTS_PATH), exist_ok=True)
# Save the index and documents
with open(DOCUMENTS_PATH, 'wb') as f:
pickle.dump(documents, f)
faiss.write_index(index, FAISS_INDEX_PATH)
print("FAISS index and documents saved successfully.")
return index, documents, embedding_model
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)
# Initialize global variables in app state
@app.on_event("startup")
async def startup_event():
"""Initialize the application on startup."""
logger = logging.getLogger(__name__)
logger.info("Starting application initialization...")
# Check if CUDA is available
device = "cuda" if torch.cuda.is_available() else "cpu"
logger.info(f"Using device: {device}")
if device == "cpu":
logger.warning("GPU not detected. Model will run slower on CPU.")
# Set NLTK data path
nltk_data_dir = os.environ.get('NLTK_DATA', os.path.join(os.path.expanduser('~'), 'nltk_data'))
os.makedirs(nltk_data_dir, exist_ok=True)
# Download NLTK data
logger.info("Downloading NLTK data...")
try:
# Check if punkt is already downloaded
import nltk.data
try:
nltk.data.find('tokenizers/punkt', paths=[nltk_data_dir])
logger.info("NLTK punkt already downloaded")
except LookupError:
await asyncio.to_thread(nltk.download, 'punkt', download_dir=nltk_data_dir, quiet=True)
try:
nltk.data.find('tokenizers/punkt_tab', paths=[nltk_data_dir])
logger.info("NLTK punkt_tab already downloaded")
except LookupError:
await asyncio.to_thread(nltk.download, 'punkt_tab', download_dir=nltk_data_dir, quiet=True)
except Exception as e:
logger.error(f"Error handling NLTK data: {str(e)}")
raise Exception(f"Failed to initialize application: {str(e)}")
# Initialize the model and index
try:
app.state.pipe = pipeline(
"text-generation",
model=MODEL_ID,
trust_remote_code=True,
token=HUGGINGFACE_TOKEN,
device_map="auto",
torch_dtype=torch.float16 if device == "cuda" else torch.float32
)
faiss_index, documents, embedding_model = await load_or_create_index()
# Store these in app.state for access across the application
app.state.faiss_index = faiss_index
app.state.documents = documents
app.state.embedding_model = embedding_model
logger.info("Application initialization completed successfully")
except Exception as e:
logger.error(f"Error initializing application: {str(e)}")
raise Exception(f"Failed to initialize application: {str(e)}")
@app.post("/generate")
async def generate_content(request: ContentRequest):
try:
logger.info(f"Generating content for grade {request.grade}, subject {request.subject}, topic {request.topic}")
# Validate inputs
if request.grade not in [3, 4]:
raise HTTPException(status_code=400, detail="Invalid grade level. Must be 3 or 4")
if request.subject.lower() not in ["math", "science"]:
raise HTTPException(status_code=400, detail="Invalid subject. Must be 'math' or 'science'")
if request.style not in ["normal", "simple", "creative"]:
raise HTTPException(status_code=400, detail="Invalid style. Must be 'normal', 'simple', or 'creative'")
# Get relevant topic files
topic_files = get_topic_files(request.grade, request.subject, request.topic)
if not topic_files:
raise HTTPException(status_code=404, detail="Topic not found for specified grade and subject")
# Create settings dictionary
settings = {
"style": request.style,
"topic": request.topic,
"grade": request.grade,
"subject": request.subject
}
response = generate_response_with_rag(
request.topic,
app.state.faiss_index,
app.state.embedding_model,
app.state.documents,
settings
)
logger.info("Content generated successfully")
return {"response": response['content']}
except Exception as e:
logger.error(f"Error generating response: {str(e)}")
raise HTTPException(status_code=500, detail=str(e))
@app.get("/health")
async def health_check():
try:
# Check if model is loaded
if not hasattr(app.state, "pipe"):
return {"status": "starting", "message": "Model is still loading"}
return {"status": "healthy"}
except Exception as e:
logger.error(f"Health check failed: {str(e)}")
raise HTTPException(status_code=500, detail="Internal server error")
if __name__ == "__main__":
try:
logger.info("Starting FastAPI server...")
uvicorn.run(app, host="0.0.0.0", port=8080, log_level="info")
except Exception as e:
logger.error(f"Application failed to start: {str(e)}")
raise