SmartManuals-AI / app.py
damoojeje's picture
Update app.py
18069c2 verified
raw
history blame
25.1 kB
import os
import json
import fitz # PyMuPDF
import nltk
import chromadb
from tqdm import tqdm
from nltk.tokenize import sent_tokenize
from sentence_transformers import SentenceTransformer, util
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import pytesseract
from PIL import Image
import io
import gradio as gr
# ---------------------------
# βš™οΈ Configuration
# ---------------------------
pdf_folder = r"./Manuals" # Path relative to the app.py file in the Space
output_jsonl_pages = "manual_pages_with_ocr.jsonl"
output_jsonl_chunks = "manual_chunks_with_ocr.jsonl"
chroma_path = "./chroma_store"
collection_name = "manual_chunks"
chunk_size = 750
chunk_overlap = 100
MAX_CONTEXT_CHUNKS = 3 # Max chunks to send to the LLM
# Hugging Face Model Configuration
HF_MODEL_ID = "meta-llama/Llama-3.1-8B-Instruct"
# Read HF Token from environment variable for security
HF_TOKEN = os.environ.get("HF_TOKEN") # Hugging Face Space secret name
# ---------------------------
# Ensure NLTK resources are available
# ---------------------------
try:
nltk.data.find('tokenizers/punkt')
except nltk.downloader.DownloadError:
nltk.download('punkt')
except LookupError:
nltk.download('punkt')
# ---------------------------
# πŸ“„ Utility: Read PDF to text (with OCR fallback)
# ---------------------------
# This combines logic from extract_text_from_pdf and extract_text_from_page
def extract_text_from_page_with_ocr(page):
text = page.get_text().strip()
if text:
return text, False # native text found, no OCR needed
# If native text is missing, try OCR
try:
pix = page.get_pixmap(dpi=300)
img_data = pix.tobytes("png")
img = Image.open(io.BytesIO(img_data))
ocr_text = pytesseract.image_to_string(img).strip()
return ocr_text, True
except Exception as e:
print(f"OCR failed for a page: {e}")
return "", False # Return empty and indicate OCR was not used if it fails
# ---------------------------
# 🧹 Clean up lines (from original notebook)
# ---------------------------
def clean_text(text):
lines = text.splitlines()
lines = [line.strip() for line in lines if line.strip()]
return "\n".join(lines)
# ---------------------------
# βœ‚οΈ Sentence Tokenizer (from original notebook)
# ---------------------------
def tokenize_sentences(text):
return sent_tokenize(text)
# ---------------------------
# πŸ“¦ Chunk into fixed size blocks (from original notebook)
# ---------------------------
def split_into_chunks(sentences, max_tokens=750, overlap=100):
chunks = []
current_chunk = []
current_len = 0
for sentence in sentences:
token_count = len(sentence.split())
# Check if adding the next sentence exceeds max_tokens
# If it does, and the current chunk is not empty, save the current chunk
if current_len + token_count > max_tokens and current_chunk:
chunks.append(" ".join(current_chunk))
# Start the next chunk with the overlap
current_chunk = current_chunk[-overlap:]
# Recalculate current_len based on the overlap
current_len = sum(len(s.split()) for s in current_chunk)
# Add the current sentence and update length
current_chunk.append(sentence)
current_len += token_count
# Add the last chunk if it's not empty
if current_chunk:
chunks.append(" ".join(current_chunk))
return chunks
# ---------------------------
# 🧠 Extract Metadata from Filename (from original notebook)
# ---------------------------
def extract_metadata_from_filename(filename):
name = filename.lower().replace("_", " ").replace("-", " ")
metadata = {
"model": "unknown",
"doc_type": "unknown",
"brand": "life fitness" # Assuming 'life fitness' is constant based on your notebook
}
if "om" in name or "owner" in name:
metadata["doc_type"] = "owner manual"
elif "sm" in name or "service" in name:
metadata["doc_type"] = "service manual"
elif "assembly" in name:
metadata["doc_type"] = "assembly instructions"
elif "alert" in name:
metadata["doc_type"] = "installer alert"
elif "parts" in name:
metadata["doc_type"] = "parts manual"
elif "bulletin" in name:
metadata["doc_type"] = "service bulletin"
known_models = [
"se3hd", "se3", "se4", "symbio", "explore", "integrity x", "integrity sl",
"everest", "engage", "inspire", "discover", "95t", "95x", "95c", "95r", "97c"
]
for model in known_models:
# Use regex for more robust matching if needed, but simple 'in' check from notebook
if model.replace(" ", "") in name.replace(" ", ""):
metadata["model"] = model
break
return metadata
# ---------------------------
# πŸš€ Step 1: Process PDFs, Extract Pages with OCR
# ---------------------------
def process_pdfs_for_pages(pdf_folder, output_jsonl):
print("Starting PDF processing and OCR...")
all_pages = []
if not os.path.exists(pdf_folder):
print(f"Error: PDF folder not found at {pdf_folder}")
return [] # Return empty list if folder doesn't exist
pdf_files = [f for f in os.listdir(pdf_folder) if f.lower().endswith(".pdf")]
if not pdf_files:
print(f"No PDF files found in {pdf_folder}")
return []
for pdf_file in tqdm(pdf_files, desc="Scanning PDFs"):
path = os.path.join(pdf_folder, pdf_file)
try:
doc = fitz.open(path)
for page_num, page in enumerate(doc, start=1):
text, used_ocr = extract_text_from_page_with_ocr(page)
if text: # Only save pages with extracted text
all_pages.append({
"source_file": pdf_file,
"page": page_num,
"text": text,
"ocr_used": used_ocr
})
doc.close() # Close the document
except Exception as e:
print(f"Error processing {pdf_file}: {e}")
continue # Skip to the next file
with open(output_jsonl, "w", encoding="utf-8") as f:
for page in all_pages:
json.dump(page, f)
f.write("\n")
print(f"βœ… Saved {len(all_pages)} pages to {output_jsonl} (with OCR fallback)")
return all_pages # Return the list of pages
# ---------------------------
# πŸš€ Step 2: Chunk the Pages
# ---------------------------
def chunk_pages(input_jsonl, output_jsonl, chunk_size, chunk_overlap):
print("Starting page chunking...")
all_chunks = []
if not os.path.exists(input_jsonl):
print(f"Error: Input JSONL file not found at {input_jsonl}. Run PDF processing first.")
return []
try:
with open(input_jsonl, "r", encoding="utf-8") as f:
# Count lines for tqdm progress bar
total_lines = sum(1 for _ in f)
f.seek(0) # Reset file pointer to the beginning
for line in tqdm(f, total=total_lines, desc="Chunking pages"):
try:
page = json.loads(line)
source_file = page["source_file"]
page_number = page["page"]
text = page["text"]
metadata = extract_metadata_from_filename(source_file)
sentences = tokenize_sentences(clean_text(text)) # Clean and tokenize the page text
chunks = split_into_chunks(sentences, max_tokens=chunk_size, overlap=chunk_overlap)
for i, chunk in enumerate(chunks):
# Ensure chunk text is not empty
if chunk.strip():
all_chunks.append({
"source_file": source_file,
"chunk_id": f"{source_file}::page_{page_number}::chunk_{i+1}",
"page": page_number,
"ocr_used": page.get("ocr_used", False), # Use .get for safety
"model": metadata.get("model", "unknown"),
"doc_type": metadata.get("doc_type", "unknown"),
"brand": metadata.get("brand", "life fitness"),
"text": chunk.strip() # Ensure no leading/trailing whitespace
})
except json.JSONDecodeError:
print(f"Skipping invalid JSON line: {line}")
except Exception as e:
print(f"Error processing page from {line}: {e}")
continue # Continue with the next line
except Exception as e:
print(f"Error opening or reading input JSONL file: {e}")
return []
if not all_chunks:
print("No chunks were created.")
with open(output_jsonl, "w", encoding="utf-8") as f:
for chunk in all_chunks:
json.dump(chunk, f)
f.write("\n")
print(f"βœ… Done! {len(all_chunks)} chunks saved to {output_jsonl}")
return all_chunks # Return the list of chunks
# ---------------------------
# πŸš€ Step 3: Embed Chunks into Chroma
# ---------------------------
def embed_chunks_into_chroma(jsonl_path, chroma_path, collection_name):
print("Starting ChromaDB embedding...")
try:
embedder = SentenceTransformer("all-MiniLM-L6-v2")
embedder.eval()
print("βœ… SentenceTransformer model loaded.")
except Exception as e:
print(f"❌ Error loading SentenceTransformer model: {e}")
return None, "Error loading SentenceTransformer model."
try:
# Use a persistent client
client = chromadb.PersistentClient(path=chroma_path)
# Check if collection exists and delete if it does to rebuild
try:
client.get_collection(name=collection_name)
client.delete_collection(collection_name)
print(f"Deleted existing collection: {collection_name}")
except Exception: # Collection does not exist, which is fine
pass
collection = client.create_collection(name=collection_name)
print(f"βœ… ChromaDB collection '{collection_name}' created.")
except Exception as e:
print(f"❌ Error initializing ChromaDB: {e}")
return None, "Error initializing ChromaDB."
texts, metadatas, ids = [], [], []
batch_size = 16 # Define batch size for embedding
if not os.path.exists(jsonl_path):
print(f"Error: Input JSONL file not found at {jsonl_path}. Run chunking first.")
return None, "Input chunk file not found."
try:
with open(jsonl_path, "r", encoding="utf-8") as f:
# Count lines for tqdm progress bar
total_lines = sum(1 for _ in f)
f.seek(0) # Reset file pointer to the beginning
for line in tqdm(f, total=total_lines, desc="Embedding chunks"):
try:
item = json.loads(line)
texts.append(item.get("text", "")) # Use .get for safety
ids.append(item.get("chunk_id", f"unknown_{len(ids)}")) # Ensure chunk_id exists
# Prepare metadata, ensuring all keys are strings and handling potential missing keys
metadata = {str(k): str(v) for k, v in item.items() if k != "text"}
metadatas.append(metadata)
if len(texts) >= batch_size:
embeddings = embedder.encode(texts).tolist()
collection.add(documents=texts, metadatas=metadatas, ids=ids, embeddings=embeddings)
texts, metadatas, ids = [], [], [] # Reset batches
except json.JSONDecodeError:
print(f"Skipping invalid JSON line during embedding: {line}")
except Exception as e:
print(f"Error processing chunk line {line} during embedding: {e}")
continue # Continue with the next line
# Add any remaining items in the last batch
if texts:
embeddings = embedder.encode(texts).tolist()
collection.add(documents=texts, metadatas=metadatas, ids=ids, embeddings=embeddings)
print("βœ… All OCR-enhanced chunks embedded in Chroma!")
return collection, None # Return collection and no error
except Exception as e:
print(f"❌ Error reading input JSONL file for embedding: {e}")
return None, "Error reading input file for embedding."
# ---------------------------
# 🧠 Load Hugging Face Model and Tokenizer
# ---------------------------
# This needs to happen after imports but before the Gradio interface
tokenizer = None
model = None
pipe = None
print(f"Attempting to load Hugging Face model: {HF_MODEL_ID}")
print(f"Using HF_TOKEN (present: {HF_TOKEN is not None})")
if not HF_TOKEN:
print("❌ HF_TOKEN environment variable not set. Cannot load Hugging Face model.")
else:
try:
# Check if CUDA is available
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")
# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(HF_MODEL_ID, token=HF_TOKEN)
model = AutoModelForCausalLM.from_pretrained(
HF_MODEL_ID,
token=HF_TOKEN,
torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32, # Use bfloat16 on GPU
device_map="auto" if torch.cuda.is_available() else None # Auto device mapping on GPU
).to(device) # Move model to selected device
# Create a pipeline for easy inference
pipe = pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
max_new_tokens=512,
temperature=0.1,
top_p=0.9,
do_sample=True,
device=0 if torch.cuda.is_available() else -1 # Specify device for pipeline
)
print(f"βœ… Successfully loaded Hugging Face model: {HF_MODEL_ID} on {device}")
except Exception as e:
print(f"❌ Error loading Hugging Face model: {e}")
print("Please ensure:")
print("- The HF_TOKEN secret is set in your Hugging Face Space settings.")
print("- Your Space has sufficient resources (GPU, RAM) for the model.")
print("- You have accepted the model's terms on Hugging Face (if required).")
tokenizer, model, pipe = None, None, None # Set to None if loading fails
# ---------------------------
# πŸ”Ž Query Function (Uses Embedder and Chroma)
# ---------------------------
# Embedder is loaded during the embedding step, need to ensure it's accessible
embedder = None # Initialize embedder as None
def query_manuals(question, model_filter=None, doc_type_filter=None, top_k=5, rerank_keywords=None):
global embedder # Access the global embedder variable
if collection is None or embedder is None:
print("⚠️ ChromaDB or Embedder not loaded. Cannot perform vector search.")
return [] # Return empty if Chroma or Embedder is not loaded
where_filter = {}
if model_filter:
where_filter["model"] = model_filter.lower()
if doc_type_filter:
where_filter["doc_type"] = doc_type_filter.lower()
# ChromaDB query expects a dictionary for 'where'
results = collection.query(
query_texts=[question],
n_results=top_k * 5, # fetch more for reranking
where={} if not where_filter else where_filter # Pass empty dict if no filter
)
if not results or not results.get("documents") or not results["documents"][0]:
return [] # No matches
try:
question_embedding = embedder.encode(question, convert_to_tensor=True)
except Exception as e:
print(f"Error encoding question: {e}")
return [] # Return empty if embedding fails
# Step 3: Compute semantic + keyword score
reranked = []
# Ensure results["documents"] and results["metadatas"] are not empty before iterating
if results.get("documents") and results["documents"][0]:
for i, text in enumerate(results["documents"][0]):
meta = results["metadatas"][0][i]
# Handle potential encoding errors during text embedding
try:
embedding = embedder.encode(text, convert_to_tensor=True)
# Semantic similarity
similarity_score = float(util.cos_sim(question_embedding, embedding))
except Exception as e:
print(f"Error encoding chunk text for reranking: {e}. Skipping chunk.")
continue # Skip this chunk if encoding fails
# Keyword score
keyword_score = 0
if rerank_keywords and text: # Ensure text is not None or empty
for kw in rerank_keywords:
if kw.lower() in text.lower():
keyword_score += 1
# Combine with tunable weights
# Weights should sum to 1 for a simple weighted average
final_score = (0.8 * similarity_score) + (0.2 * keyword_score)
reranked.append({
"score": final_score,
"text": text,
"metadata": meta
})
# Sort by combined score
reranked.sort(key=lambda x: x["score"], reverse=True)
return reranked[:top_k]
# ---------------------------
# πŸ’¬ Ask Hugging Face Model
# ---------------------------
def ask_hf_model(prompt):
if pipe is None:
return "Hugging Face model not loaded. Cannot generate response."
try:
# Use the Llama 3.1 chat template
messages = [
{"role": "system", "content": "You are a technical assistant trained to answer questions using equipment manuals. Use only the provided context to answer the question. If the answer is not clearly in the context, reply: 'I don't know.'"},
{"role": "user", "content": prompt}
]
# Apply chat template and generate text
prompt_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipe(
prompt_text,
do_sample=True,
temperature=0.1, # Keep temperature low for more factual answers
top_p=0.9,
max_new_tokens=512,
pad_token_id=tokenizer.eos_token_id # Set pad_token_id for generation
)
# The output includes the prompt, we need to extract just the generated part
# Find the end of the prompt text in the generated output
generated_text = outputs[0]["generated_text"]
# The chat template adds the assistant's turn token, look for that to find the response start
response_start_token = tokenizer.apply_chat_template([{"role": "assistant", "content": ""}], tokenize=False, add_generation_prompt=False)
response_start_index = generated_text.find(response_start_token)
if response_start_index != -1:
response = generated_text[response_start_index + len(response_start_token):].strip()
else:
# Fallback if the assistant token isn't found
response = generated_text.strip()
# Remove any trailing EOS tokens or similar artifacts
if response.endswith(tokenizer.eos_token):
response = response[:-len(tokenizer.eos_token)].strip()
return response
except Exception as e:
return f"❌ Error generating response from Hugging Face model: {str(e)}"
# ---------------------------
# 🎯 Full RAG Pipeline
# ---------------------------
def run_rag_qa(user_question, model_filter=None, doc_type_filter=None): # Added filters as optional inputs
# Ensure ChromaDB and the HF model pipeline are loaded before proceeding
if collection is None:
return "ChromaDB is not loaded. Ensure PDFs are in ./Manuals and the app started correctly."
if pipe is None:
return "Hugging Face model pipeline is not loaded. Ensure HF_TOKEN is set and the model loaded successfully."
results = query_manuals(
question=user_question,
model_filter=model_filter, # Use the optional filter inputs
doc_type_filter=doc_type_filter,
top_k=MAX_CONTEXT_CHUNKS,
rerank_keywords=["diagnostic", "immobilize", "system", "screen", "service", "error"] # Example keywords
)
if not results:
# Attempt a broader search if initial filter yields no results
if model_filter or doc_type_filter:
print("No results with specified filters, trying broader search...")
results = query_manuals(
question=user_question,
model_filter=None, # Remove filters for broader search
doc_type_filter=None,
top_k=MAX_CONTEXT_CHUNKS,
rerank_keywords=["diagnostic", "immobilize", "system", "screen", "service", "error"]
)
if not results:
return "No relevant documents found for the query, even with broader search."
else:
return "No relevant documents found for the query."
context = "\n\n".join([f"Source File: {r['metadata'].get('source_file', 'N/A')}, Page: {r['metadata'].get('page', 'N/A')}\nText: {r['text'].strip()}" for r in results])
prompt = f"""
Context:
{context}
Question: {user_question}
"""
return ask_hf_model(prompt)
# ---------------------------
# --- Initial Setup ---
# This code runs when the app starts on Hugging Face Spaces
# It processes PDFs, chunks, and builds the ChromaDB
# ---------------------------
print("Starting initial setup...")
# Ensure Tesseract is available on the system (Hugging Face Spaces usually has it, but this command is good practice)
# Using ! in app.py is generally discouraged, better to ensure the environment has it
# For HF Spaces, you might need to use a Dockerfile or rely on the default environment.
# If Tesseract isn't found, the OCR part might fail.
# Process PDFs and extract pages
all_pages = process_pdfs_for_pages(pdf_folder, output_jsonl_pages)
# Chunk the pages
all_chunks = []
if all_pages: # Only chunk if pages were processed
all_chunks = chunk_pages(output_jsonl_pages, output_jsonl_chunks, chunk_size, chunk_overlap)
# Embed chunks into ChromaDB
collection = None # Initialize collection
if all_chunks: # Only embed if chunks were created
collection, embed_error = embed_chunks_into_chroma(output_jsonl_chunks, chroma_path, collection_name)
if embed_error:
print(f"Error during embedding: {embed_error}")
print("Initial setup complete.")
# ---------------------------
# πŸ–₯️ Gradio Interface
# ---------------------------
# Only define and launch the interface if the necessary components loaded
if collection is not None and pipe is not None:
with gr.Blocks() as demo:
gr.Markdown("""# 🧠 Manual QA via Hugging Face Llama 3.1
Ask a technical question and get answers using your own PDF manual database and a Hugging Face model.
**Note:** Initial startup might take time to process manuals and build the search index. Ensure your `Manuals` folder is uploaded and the `HF_TOKEN` secret is set in Space settings.
""")
with gr.Row():
question = gr.Textbox(label="Your Question", placeholder="e.g. How do I access diagnostics on the SE3 console?")
with gr.Row():
model_filter_input = gr.Textbox(label="Filter by Model (Optional)", placeholder="e.g. se3hd")
doc_type_filter_input = gr.Dropdown(label="Filter by Document Type (Optional)", choices=["owner manual", "service manual", "assembly instructions", "installer alert", "parts manual", "service bulletin", "unknown", None], value=None, allow_custom_value=True)
submit = gr.Button("πŸ” Ask")
answer = gr.Textbox(label="Answer", lines=10) # Increased lines for better readability
# Call the run_rag_qa function when the button is clicked
submit.click(
fn=run_rag_qa,
inputs=[question, model_filter_input, doc_type_filter_input],
outputs=[answer]
)
# In Hugging Face Spaces, the app is launched automatically.
# The demo.launch() call is removed.
# demo.launch()
else:
print("Gradio demo will not launch because RAG components (ChromaDB or HF Model) failed to load during setup.")
# You could add a simple Gradio interface here to show an error message
# if you wanted to provide user feedback in the Space UI even on failure.
# Example:
# with gr.Blocks() as error_demo:
# gr.Markdown("## Application Failed to Load")
# gr.Textbox(label="Error Details", value="RAG components (ChromaDB or HF Model) failed to initialize. Check logs and Space settings (HF_TOKEN, resources).", interactive=False)
# error_demo.launch()