import gradio as gr
import gc
import os
import time
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"
os.environ["CUDA_VISIBLE_DEVICES"] = "" # Force CPU only
import uuid
import threading
import pandas as pd
import torch
from langchain.document_loaders import CSVLoader
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.vectorstores import FAISS
from langchain.llms import HuggingFacePipeline
from langchain.chains import LLMChain
from transformers import AutoTokenizer, AutoModelForCausalLM, T5Tokenizer, T5ForConditionalGeneration, BitsAndBytesConfig, pipeline
from langchain.prompts import PromptTemplate
from llama_cpp import Llama
import re
import datetime
import warnings
warnings.filterwarnings('ignore')
# Global model cache
MODEL_CACHE = {
"model": None,
"tokenizer": None,
"init_lock": threading.Lock(),
"model_name": None
}
# Create directories for user data
os.makedirs("user_data", exist_ok=True)
os.makedirs("performance_metrics", exist_ok=True)
# Model configuration dictionary
MODEL_CONFIG = {
"Llama 2 Chat": {
"name": "TheBloke/Llama-2-7B-Chat-GGUF",
"description": "Llama 2 7B Chat model with good general performance",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"TinyLlama Chat": {
"name": "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF",
"description": "Lightweight model with 1.1B parameters, fast and efficient",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"Mistral Instruct": {
"name": "TheBloke/Mistral-7B-Instruct-v0.2-GGUF",
"description": "7B instruction-tuned model with excellent reasoning",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"Phi-4 Mini Instruct": {
"name": "microsoft/Phi-4-mini-instruct",
"description": "Lightweight model from Microsoft suitable for instructional tasks",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"DeepSeek Coder Instruct": {
"name": "deepseek-ai/deepseek-coder-1.3b-instruct",
"description": "1.3B model for code and data analysis",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"DeepSeek Lite Chat": {
"name": "deepseek-ai/DeepSeek-V2-Lite-Chat",
"description": "Light but powerful chat model from DeepSeek",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"Qwen2.5 Coder Instruct": {
"name": "Qwen/Qwen2.5-Coder-3B-Instruct-GGUF",
"description": "3B model specialized for code and technical applications",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"DeepSeek Distill Qwen": {
"name": "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
"description": "1.5B distilled model with good balance of speed and quality",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32
},
"Flan T5 Small": {
"name": "google/flan-t5-small",
"description": "Lightweight T5 model optimized for instruction following",
"dtype": torch.float16 if torch.cuda.is_available() else torch.float32,
"is_t5": True
}
}
# Performance metrics tracking
class PerformanceTracker:
def __init__(self):
self.metrics_file = "performance_metrics/model_performance.csv"
# Create metrics file if it doesn't exist
if not os.path.exists(self.metrics_file):
with open(self.metrics_file, "w") as f:
f.write("timestamp,model,question,processing_time,response_length\n")
def log_performance(self, model_name, question, processing_time, response):
timestamp = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
response_length = len(response)
with open(self.metrics_file, "a") as f:
f.write(f'"{timestamp}","{model_name}","{question}",{processing_time},{response_length}\n')
print(f"Logged performance for {model_name}: {processing_time:.2f}s")
# Initialize performance tracker
performance_tracker = PerformanceTracker()
def initialize_model_once(model_key):
with MODEL_CACHE["init_lock"]:
current_model = MODEL_CACHE["model_name"]
if MODEL_CACHE["model"] is None or current_model != model_key:
# Clear previous model
if MODEL_CACHE["model"] is not None:
del MODEL_CACHE["model"]
if MODEL_CACHE["tokenizer"] is not None:
del MODEL_CACHE["tokenizer"]
torch.cuda.empty_cache() if torch.cuda.is_available() else None
model_info = MODEL_CONFIG[model_key]
model_name = model_info["name"]
MODEL_CACHE["model_name"] = model_key
try:
print(f"Loading model: {model_name}")
# Check if this is a GGUF model
if "GGUF" in model_name:
# Download the model file first if it doesn't exist
from huggingface_hub import hf_hub_download
try:
# Try to find the GGUF file in the repo
repo_id = model_name
model_path = hf_hub_download(
repo_id=repo_id,
filename="model.gguf" # File name may differ
)
except Exception as e:
print(f"Couldn't find model.gguf, trying other filenames: {str(e)}")
# Try to find GGUF file with other names
import requests
from huggingface_hub import list_repo_files
files = list_repo_files(repo_id)
gguf_files = [f for f in files if f.endswith('.gguf')]
if not gguf_files:
raise ValueError(f"No GGUF files found in {repo_id}")
# Use first GGUF file found
model_path = hf_hub_download(repo_id=repo_id, filename=gguf_files[0])
# Load GGUF model with llama-cpp-python
MODEL_CACHE["model"] = Llama(
model_path=model_path,
n_ctx=2048, # Smaller context for memory savings
n_batch=512,
n_threads=2 # Adjust for 2 vCPU
)
MODEL_CACHE["tokenizer"] = None # GGUF doesn't need separate tokenizer
MODEL_CACHE["is_gguf"] = True
# Handle T5 models
elif model_info.get("is_t5", False):
MODEL_CACHE["tokenizer"] = T5Tokenizer.from_pretrained(model_name)
MODEL_CACHE["model"] = T5ForConditionalGeneration.from_pretrained(
model_name,
torch_dtype=model_info["dtype"],
device_map="auto" if torch.cuda.is_available() else None,
low_cpu_mem_usage=True
)
MODEL_CACHE["is_gguf"] = False
# Handle standard HF models
else:
# Only use quantization if CUDA is available
if torch.cuda.is_available():
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True
)
MODEL_CACHE["tokenizer"] = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
MODEL_CACHE["model"] = AutoModelForCausalLM.from_pretrained(
model_name,
quantization_config=quantization_config,
torch_dtype=model_info["dtype"],
device_map="auto",
low_cpu_mem_usage=True,
trust_remote_code=True
)
else:
# For CPU-only environments, load without quantization
MODEL_CACHE["tokenizer"] = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
MODEL_CACHE["model"] = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float32, # Use float32 for CPU
device_map=None,
low_cpu_mem_usage=True,
trust_remote_code=True
)
MODEL_CACHE["is_gguf"] = False
print(f"Model {model_name} loaded successfully")
except Exception as e:
import traceback
print(f"Error loading model {model_name}: {str(e)}")
print(traceback.format_exc())
raise RuntimeError(f"Failed to load model {model_name}: {str(e)}")
return MODEL_CACHE["tokenizer"], MODEL_CACHE["model"], MODEL_CACHE.get("is_gguf", False)
def create_llm_pipeline(model_key):
"""Create a new pipeline using the specified model"""
try:
print(f"Creating pipeline for model: {model_key}")
tokenizer, model, is_gguf = initialize_model_once(model_key)
# Get the model info for reference
model_info = MODEL_CONFIG[model_key]
if model is None:
raise ValueError(f"Model is None for {model_key}")
# For GGUF models from llama-cpp-python
if is_gguf:
# Create adapter to use GGUF model like HF pipeline
from langchain.llms import LlamaCpp
llm = LlamaCpp(
model_path=model.model_path,
temperature=0.3,
max_tokens=256, # Increased for more comprehensive answers
top_p=0.9,
n_ctx=2048,
streaming=False
)
return llm
# Create appropriate pipeline for HF models
elif model_info.get("is_t5", False):
print("Creating T5 pipeline")
pipe = pipeline(
"text2text-generation",
model=model,
tokenizer=tokenizer,
max_new_tokens=256, # Increased for more comprehensive answers
temperature=0.3,
top_p=0.9,
return_full_text=False,
)
else:
print("Creating causal LM pipeline")
pipe = pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
max_new_tokens=256, # Increased for more comprehensive answers
temperature=0.3,
top_p=0.9,
top_k=30,
repetition_penalty=1.2,
return_full_text=False,
)
print("Pipeline created successfully")
return HuggingFacePipeline(pipeline=pipe)
except Exception as e:
import traceback
print(f"Error creating pipeline: {str(e)}")
print(traceback.format_exc())
raise RuntimeError(f"Failed to create pipeline: {str(e)}")
def handle_model_loading_error(model_key, session_id):
"""Handle model loading errors by providing alternative model suggestions"""
suggested_models = [
"DeepSeek Coder Instruct", # 1.3B model
"Phi-4 Mini Instruct", # Light model
"TinyLlama Chat", # 1.1B model
"Flan T5 Small" # Lightweight T5
]
# Remove the current model from suggestions if it's in the list
if model_key in suggested_models:
suggested_models.remove(model_key)
suggestions = ", ".join(suggested_models[:3]) # Only show top 3 suggestions
return None, f"Unable to load model {model_key}. Please try another model such as: {suggestions}"
def create_conversational_chain(db, file_path, model_key):
llm = create_llm_pipeline(model_key)
# Load the file into pandas to get metadata about the CSV
df = pd.read_csv(file_path)
# Create improved prompt template that focuses on pure LLM analysis
template = """
You are an expert data analyst tasked with answering questions about a CSV file. The file has been analyzed, and its structure is provided below.
CSV File Structure:
- Total rows: {row_count}
- Total columns: {column_count}
- Columns: {columns_list}
Sample data (first few rows):
{sample_data}
Additional context from the document:
{context}
User Question: {question}
IMPORTANT INSTRUCTIONS:
1. Answer the question directly about the CSV data with accurate information.
2. If asked for basic statistics (mean, sum, max, min, count, etc.), perform the calculation mentally and provide the result. Include up to 2 decimal places for non-integer values.
3. If asked about patterns or trends, analyze the data thoughtfully.
4. Keep answers concise but informative. Respond in the same language as the question.
5. If you are not certain of a precise answer, explain what you can determine from the available data.
6. You can perform simple calculations including: counts, sums, averages, minimums, maximums, and basic filtering.
7. For questions about specific values in the data, reference the sample data and available context.
8. Do not mention any programming language or how you would code the solution.
Your analysis:
"""
PROMPT = PromptTemplate(
template=template,
input_variables=["row_count", "column_count", "columns_list", "sample_data", "context", "question"]
)
# Create retriever
retriever = db.as_retriever(search_kwargs={"k": 5}) # Increase k for better context
# Process query with better error handling
def process_query(query, chat_history):
try:
start_time = time.time()
# Get information from dataframe for context
columns_list = ", ".join(df.columns.tolist())
sample_data = df.head(5).to_string() # Show 5 rows for better context
row_count = len(df)
column_count = len(df.columns)
# Get context from vector database
docs = retriever.get_relevant_documents(query)
context = "\n\n".join([doc.page_content for doc in docs])
# Run the chain
chain = LLMChain(llm=llm, prompt=PROMPT)
raw_result = chain.run(
row_count=row_count,
column_count=column_count,
columns_list=columns_list,
sample_data=sample_data,
context=context,
question=query
)
# Clean the result
cleaned_result = raw_result.strip()
# If result is empty after cleaning, use a fallback
if not cleaned_result:
cleaned_result = "I couldn't process a complete answer to your question. Please try asking in a different way or provide more specific details about what you'd like to know about the data."
processing_time = time.time() - start_time
# Log performance metrics
performance_tracker.log_performance(
model_key,
query,
processing_time,
cleaned_result
)
# Add processing time to the response for comparison purposes
result_with_metrics = f"{cleaned_result}\n\n[Processing time: {processing_time:.2f} seconds]"
return {"answer": result_with_metrics}
except Exception as e:
import traceback
print(f"Error in process_query: {str(e)}")
print(traceback.format_exc())
return {"answer": f"An error occurred while processing your question: {str(e)}"}
return process_query
class ChatBot:
def __init__(self, session_id, model_key="DeepSeek Coder Instruct"):
self.session_id = session_id
self.chat_history = []
self.chain = None
self.user_dir = f"user_data/{session_id}"
self.csv_file_path = None
self.model_key = model_key
os.makedirs(self.user_dir, exist_ok=True)
def process_file(self, file, model_key=None):
if model_key:
self.model_key = model_key
if file is None:
return "Please upload a CSV file first."
try:
start_time = time.time()
print(f"Processing file using model: {self.model_key}")
# Handle file from Gradio
file_path = file.name if hasattr(file, 'name') else str(file)
self.csv_file_path = file_path
print(f"CSV file path: {file_path}")
# Copy to user directory
user_file_path = f"{self.user_dir}/uploaded.csv"
# Verify the CSV can be loaded
try:
df = pd.read_csv(file_path)
print(f"CSV verified: {df.shape[0]} rows, {len(df.columns)} columns")
# Save a copy in user directory
df.to_csv(user_file_path, index=False)
self.csv_file_path = user_file_path
print(f"CSV saved to {user_file_path}")
except Exception as e:
print(f"Error reading CSV: {str(e)}")
return f"Error reading CSV: {str(e)}"
# Load document with reduced chunk size for better memory usage
try:
loader = CSVLoader(file_path=user_file_path, encoding="utf-8", csv_args={
'delimiter': ','})
data = loader.load()
print(f"Documents loaded: {len(data)}")
except Exception as e:
print(f"Error loading documents: {str(e)}")
return f"Error loading documents: {str(e)}"
# Create vector database with optimized settings
try:
db_path = f"{self.user_dir}/db_faiss"
# Use CPU-friendly embeddings with smaller dimensions
embeddings = HuggingFaceEmbeddings(
model_name='sentence-transformers/all-MiniLM-L6-v2',
model_kwargs={'device': 'cpu'}
)
db = FAISS.from_documents(data, embeddings)
db.save_local(db_path)
print(f"Vector database created at {db_path}")
except Exception as e:
print(f"Error creating vector database: {str(e)}")
return f"Error creating vector database: {str(e)}"
# Create custom chain
try:
print(f"Creating conversation chain with model: {self.model_key}")
self.chain = create_conversational_chain(db, self.csv_file_path, self.model_key)
print("Chain created successfully")
except Exception as e:
import traceback
print(f"Error creating chain: {str(e)}")
print(traceback.format_exc())
return f"Error creating chain: {str(e)}"
# Add basic file info to chat history for context
file_processing_time = time.time() - start_time
file_info = f"CSV successfully loaded with {df.shape[0]} rows and {len(df.columns)} columns using model {self.model_key}. Columns: {', '.join(df.columns.tolist())}"
self.chat_history.append(("System", file_info))
return f"CSV file successfully processed with model {self.model_key}! You can now chat with the model to analyze the data.\n\n[Processing time: {file_processing_time:.2f} seconds]"
except Exception as e:
import traceback
print(traceback.format_exc())
return f"File processing error: {str(e)}"
def change_model(self, model_key):
"""Change the model being used and recreate the chain if necessary"""
try:
if model_key == self.model_key:
return f"Model {model_key} is already in use."
print(f"Changing model from {self.model_key} to {model_key}")
self.model_key = model_key
# If we have an active session with a file already loaded, recreate the chain
if self.csv_file_path and os.path.exists(self.csv_file_path):
try:
# Load existing database
db_path = f"{self.user_dir}/db_faiss"
if not os.path.exists(db_path):
return f"Error: Database not found. Please upload the CSV file again."
print(f"Loading embeddings from {db_path}")
embeddings = HuggingFaceEmbeddings(
model_name='sentence-transformers/all-MiniLM-L6-v2',
model_kwargs={'device': 'cpu'}
)
# Add allow_dangerous_deserialization=True flag
db = FAISS.load_local(db_path, embeddings, allow_dangerous_deserialization=True)
print(f"FAISS database loaded successfully")
# Create new chain with the selected model
print(f"Creating new conversation chain with {model_key}")
self.chain = create_conversational_chain(db, self.csv_file_path, self.model_key)
print(f"Chain created successfully")
# Add notification to chat history
self.chat_history.append(("System", f"Model successfully changed to {model_key}."))
return f"Model successfully changed to {model_key}."
except Exception as e:
import traceback
error_trace = traceback.format_exc()
print(f"Detailed error in change_model: {error_trace}")
return f"Error changing model: {str(e)}"
else:
# Just update the model key if no file is loaded yet
print(f"No CSV file loaded yet, just updating model preference to {model_key}")
return f"Model changed to {model_key}. Please upload a CSV file to begin."
except Exception as e:
import traceback
error_trace = traceback.format_exc()
print(f"Unexpected error in change_model: {error_trace}")
return f"Unexpected error while changing model: {str(e)}"
def chat(self, message, history):
if self.chain is None:
return "Please upload a CSV file first."
try:
# Process the question with the chain
result = self.chain(message, self.chat_history)
# Get the answer with fallback
answer = result.get("answer", "Sorry, I couldn't generate an answer. Please try asking a different question.")
# Ensure we never return empty
if not answer or answer.strip() == "":
answer = "Sorry, I couldn't generate an appropriate answer. Please try asking the question differently."
# Update internal chat history
self.chat_history.append((message, answer))
# Return just the answer for Gradio
return answer
except Exception as e:
import traceback
print(traceback.format_exc())
return f"Error: {str(e)}"
# UI Code
def create_gradio_interface():
with gr.Blocks(title="Chat with CSV using AI Models") as interface:
session_id = gr.State(lambda: str(uuid.uuid4()))
chatbot_state = gr.State(lambda: None)
model_selected = gr.State(lambda: False) # Track if model is already in use
# Get model choices
model_choices = list(MODEL_CONFIG.keys())
default_model = "DeepSeek Coder Instruct" # Default model
gr.HTML("Chat with CSV using AI Models
")
gr.HTML("Asisten analisis CSV untuk berbagai kebutuhan
")
with gr.Row():
with gr.Column(scale=1):
with gr.Group():
gr.Markdown("### Step 1: Choose AI Model")
model_dropdown = gr.Dropdown(
label="Model",
choices=model_choices,
value=default_model,
interactive=True
)
model_info = gr.Markdown(
value=f"**{default_model}**: {MODEL_CONFIG[default_model]['description']}"
)
with gr.Group():
gr.Markdown("### Step 2: Upload and Process CSV")
file_input = gr.File(
label="Upload CSV Anda",
file_types=[".csv"]
)
process_button = gr.Button("Process CSV")
reset_button = gr.Button("Reset Session (To Change Model)")
with gr.Column(scale=2):
chatbot_interface = gr.Chatbot(
label="Chat History",
# type="messages",
height=400
)
message_input = gr.Textbox(
label="Type your message",
placeholder="Ask questions about your CSV data...",
lines=2
)
submit_button = gr.Button("Send")
clear_button = gr.Button("Clear Chat")
# Update model info when selection changes
def update_model_info(model_key):
return f"**{model_key}**: {MODEL_CONFIG[model_key]['description']}"
model_dropdown.change(
fn=update_model_info,
inputs=[model_dropdown],
outputs=[model_info]
)
# Process file handler - disables model selection after file is processed
def handle_process_file(file, model_key, sess_id):
if file is None:
return None, None, False, "Please upload a CSV file first."
try:
chatbot = ChatBot(sess_id, model_key)
result = chatbot.process_file(file)
return chatbot, True, [(None, result)]
except Exception as e:
import traceback
print(f"Error processing file with {model_key}: {str(e)}")
print(traceback.format_exc())
error_msg = f"Error with model {model_key}: {str(e)}\n\nPlease try another model."
return None, False, [(None, error_msg)]
process_button.click(
fn=handle_process_file,
inputs=[file_input, model_dropdown, session_id],
outputs=[chatbot_state, model_selected, chatbot_interface]
).then(
# Disable model dropdown after processing file
fn=lambda selected: gr.update(interactive=not selected),
inputs=[model_selected],
outputs=[model_dropdown]
)
# Reset handler - enables model selection again
def reset_session():
return None, False, [], gr.update(interactive=True)
reset_button.click(
fn=reset_session,
inputs=[],
outputs=[chatbot_state, model_selected, chatbot_interface, model_dropdown]
)
# Chat handlers
def user_message_submitted(message, history, chatbot, sess_id):
history = history + [(message, None)]
return history, "", chatbot, sess_id
def bot_response(history, chatbot, sess_id):
if chatbot is None:
chatbot = ChatBot(sess_id)
history[-1] = (history[-1][0], "Please upload a CSV file first.")
return chatbot, history
user_message = history[-1][0]
response = chatbot.chat(user_message, history[:-1])
history[-1] = (user_message, response)
return chatbot, history
submit_button.click(
fn=user_message_submitted,
inputs=[message_input, chatbot_interface, chatbot_state, session_id],
outputs=[chatbot_interface, message_input, chatbot_state, session_id]
).then(
fn=bot_response,
inputs=[chatbot_interface, chatbot_state, session_id],
outputs=[chatbot_state, chatbot_interface]
)
message_input.submit(
fn=user_message_submitted,
inputs=[message_input, chatbot_interface, chatbot_state, session_id],
outputs=[chatbot_interface, message_input, chatbot_state, session_id]
).then(
fn=bot_response,
inputs=[chatbot_interface, chatbot_state, session_id],
outputs=[chatbot_state, chatbot_interface]
)
# Clear chat handler
def handle_clear_chat(chatbot):
if chatbot is not None:
chatbot.chat_history = []
return chatbot, []
clear_button.click(
fn=handle_clear_chat,
inputs=[chatbot_state],
outputs=[chatbot_state, chatbot_interface]
)
return interface
# Launch the interface
demo = create_gradio_interface()
demo.launch(share=True)