In this tutorial, we build a GPU‑capable local LLM stack that unifies Ollama and LangChain. We install the required libraries, launch the Ollama server, pull a model, and wrap it in a custom LangChain LLM, allowing us to control temperature, token limits, and context. We add a Retrieval-Augmented Generation layer that ingests PDFs or text, chunks them, embeds them with Sentence-Transformers, and serves grounded answers. We manage multi‑session chat memory, register tools (web search + RAG query), and spin up an agent that reasons about when to call them.
import os
import sys
import subprocess
import time
import threading
import queue
import json
from typing import List, Dict, Any, Optional, Tuple
from dataclasses import dataclass
from contextlib import contextmanager
import asyncio
from concurrent.futures import ThreadPoolExecutor
def install_packages():
"""Install required packages for Colab environment"""
packages = [
"langchain",
"langchain-community",
"langchain-core",
"chromadb",
"sentence-transformers",
"faiss-cpu",
"pypdf",
"python-docx",
"requests",
"psutil",
"pyngrok",
"gradio"
]
for package in packages:
subprocess.check_call([sys.executable, "-m", "pip", "install", package])
install_packages()
import requests
import psutil
import threading
from queue import Queue
from langchain.llms.base import LLM
from langchain.callbacks.manager import CallbackManagerForLLMRun
from langchain.schema import BaseMessage, HumanMessage, AIMessage, SystemMessage
from langchain.memory import ConversationBufferWindowMemory, ConversationSummaryBufferMemory
from langchain.chains import ConversationChain, RetrievalQA
from langchain.prompts import PromptTemplate, ChatPromptTemplate
from langchain.document_loaders import PyPDFLoader, TextLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.vectorstores import FAISS, Chroma
from langchain.agents import AgentType, initialize_agent, Tool
from langchain.tools import DuckDuckGoSearchRunWe import the necessary Python utilities in Colab for concurrency, system calls, and JSON handling. We define and run install_packages() to pull LangChain, embeddings, vector stores, document loaders, monitoring, and UI dependencies. We then import LangChain LLM, memory, retrieval, and agent tools (including DuckDuckGo search) to build an extensible RAG and agent workflow.
[Download the full codes with notebook here]
@dataclass
class OllamaConfig:
"""Configuration for Ollama setup"""
model_name: str = "llama2"
base_url: str = "http://localhost:11434"
max_tokens: int = 2048
temperature: float = 0.7
gpu_layers: int = -1
context_window: int = 4096
batch_size: int = 512
threads: int = 4We define an OllamaConfig dataclass so we keep all Ollama runtime settings in one clean place. We set the model name and local API endpoint, as well as the generation behavior (max_tokens, temperature, and context_window). We control performance with gpu_layers (‑1 = load all to GPU when possible), batch_size, and threads for parallelism.
@dataclass
class OllamaConfig:
"""Configuration for Ollama setup"""
model_name: str = "llama2"
base_url: str = "http://localhost:11434"
max_tokens: int = 2048
temperature: float = 0.7
gpu_layers: int = -1
context_window: int = 4096
batch_size: int = 512
threads: int = 4
We define an OllamaConfig dataclass so we keep all Ollama runtime settings in one clean place. We set the model name and local API endpoint, as well as the generation behavior (max_tokens, temperature, and context_window). We control performance with gpu_layers (‑1 = load all to GPU when possible), batch_size, and threads for parallelism.
class OllamaManager:
"""Advanced Ollama manager for Colab environment"""
def __init__(self, config: OllamaConfig):
self.config = config
self.process = None
self.is_running = False
self.models_cache = {}
self.performance_monitor = PerformanceMonitor()
def install_ollama(self):
"""Install Ollama in Colab environment"""
try:
subprocess.run([
"curl", "-fsSL", "https://ollama.com/install.sh", "-o", "/tmp/install.sh"
], check=True)
subprocess.run(["bash", "/tmp/install.sh"], check=True)
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> Ollama installed successfully")
except subprocess.CalledProcessError as e:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Failed to install Ollama: {e}")
raise
def start_server(self):
"""Start Ollama server with GPU support"""
if self.is_running:
print("Ollama server is already running")
return
try:
env = os.environ.copy()
env["OLLAMA_NUM_PARALLEL"] = str(self.config.threads)
env["OLLAMA_MAX_LOADED_MODELS"] = "3"
self.process = subprocess.Popen(
["ollama", "serve"],
env=env,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE
)
time.sleep(5)
if self.health_check():
self.is_running = True
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> Ollama server started successfully")
self.performance_monitor.start()
else:
raise Exception("Server failed to start properly")
except Exception as e:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Failed to start Ollama server: {e}")
raise
def health_check(self) -> bool:
"""Check if Ollama server is healthy"""
try:
response = requests.get(f"{self.config.base_url}/api/tags", timeout=10)
return response.status_code == 200
except:
return False
def pull_model(self, model_name: str) -> bool:
"""Pull a model from Ollama registry"""
try:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f504.png" alt="🔄" class="wp-smiley" /> Pulling model: {model_name}")
result = subprocess.run(
["ollama", "pull", model_name],
capture_output=True,
text=True,
timeout=1800
)
if result.returncode == 0:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> Model {model_name} pulled successfully")
self.models_cache[model_name] = True
return True
else:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Failed to pull model {model_name}: {result.stderr}")
return False
except subprocess.TimeoutExpired:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Timeout pulling model {model_name}")
return False
except Exception as e:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Error pulling model {model_name}: {e}")
return False
def list_models(self) -> List[str]:
"""List available local models"""
try:
result = subprocess.run(
["ollama", "list"],
capture_output=True,
text=True
)
models = []
for line in result.stdout.split('n')[1:]:
if line.strip():
model_name = line.split()[0]
models.append(model_name)
return models
except Exception as e:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Error listing models: {e}")
return []
def stop_server(self):
"""Stop Ollama server"""
if self.process:
self.process.terminate()
self.process.wait()
self.is_running = False
self.performance_monitor.stop()
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> Ollama server stopped")We create the OllamaManager class to install, start, monitor, and manage the Ollama server in the Colab environment. We set environment variables for GPU parallelism, run the server in the background, and verify it’s up with a health check. We pull models on demand, cache them, list available ones locally, and gracefully shut down the server when the task is complete, all while tracking performance.
[Download the full codes with notebook here]
class PerformanceMonitor:
"""Monitor system performance and resource usage"""
def __init__(self):
self.monitoring = False
self.stats = {
"cpu_usage": [],
"memory_usage": [],
"gpu_usage": [],
"inference_times": []
}
self.monitor_thread = None
def start(self):
"""Start performance monitoring"""
self.monitoring = True
self.monitor_thread = threading.Thread(target=self._monitor_loop)
self.monitor_thread.daemon = True
self.monitor_thread.start()
def stop(self):
"""Stop performance monitoring"""
self.monitoring = False
if self.monitor_thread:
self.monitor_thread.join()
def _monitor_loop(self):
"""Main monitoring loop"""
while self.monitoring:
try:
cpu_percent = psutil.cpu_percent(interval=1)
memory = psutil.virtual_memory()
self.stats["cpu_usage"].append(cpu_percent)
self.stats["memory_usage"].append(memory.percent)
for key in ["cpu_usage", "memory_usage"]:
if len(self.stats[key]) > 100:
self.stats[key] = self.stats[key][-100:]
time.sleep(5)
except Exception as e:
print(f"Monitoring error: {e}")
def get_stats(self) -> Dict[str, Any]:
"""Get current performance statistics"""
return {
"avg_cpu": sum(self.stats["cpu_usage"][-10:]) / max(len(self.stats["cpu_usage"][-10:]), 1),
"avg_memory": sum(self.stats["memory_usage"][-10:]) / max(len(self.stats["memory_usage"][-10:]), 1),
"total_inferences": len(self.stats["inference_times"]),
"avg_inference_time": sum(self.stats["inference_times"]) / max(len(self.stats["inference_times"]), 1)
}
We define a PerformanceMonitor class to track CPU, memory, and inference times in real-time while the Ollama server runs. We launch a background thread to collect stats every few seconds, store recent metrics, and provide average usage summaries. This helps us monitor system load and optimize performance during model inference.
[Download the full codes with notebook here]
class OllamaLLM(LLM):
"""Custom LangChain LLM for Ollama"""
model_name: str = "llama2"
base_url: str = "http://localhost:11434"
temperature: float = 0.7
max_tokens: int = 2048
performance_monitor: Optional[PerformanceMonitor] = None
@property
def _llm_type(self) -> str:
return "ollama"
def _call(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> str:
"""Make API call to Ollama"""
start_time = time.time()
try:
payload = {
"model": self.model_name,
"prompt": prompt,
"stream": False,
"options": {
"temperature": self.temperature,
"num_predict": self.max_tokens,
"stop": stop or []
}
}
response = requests.post(
f"{self.base_url}/api/generate",
json=payload,
timeout=120
)
response.raise_for_status()
result = response.json()
inference_time = time.time() - start_time
if self.performance_monitor:
self.performance_monitor.stats["inference_times"].append(inference_time)
return result.get("response", "")
except Exception as e:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Ollama API error: {e}")
return f"Error: {str(e)}"We wrap the Ollama API inside a custom OllamaLLM class compatible with LangChain’s LLM interface. We define how prompts are sent to the Ollama server and record each inference time for performance tracking. This lets us plug Ollama directly into LangChain chains, agents, and memory components while monitoring efficiency.
class RAGSystem:
"""Retrieval-Augmented Generation system"""
def __init__(self, llm: OllamaLLM, embedding_model: str = "sentence-transformers/all-MiniLM-L6-v2"):
self.llm = llm
self.embeddings = HuggingFaceEmbeddings(model_name=embedding_model)
self.vector_store = None
self.qa_chain = None
self.text_splitter = RecursiveCharacterTextSplitter(
chunk_size=1000,
chunk_overlap=200,
length_function=len
)
def add_documents(self, file_paths: List[str]):
"""Add documents to the vector store"""
documents = []
for file_path in file_paths:
try:
if file_path.endswith('.pdf'):
loader = PyPDFLoader(file_path)
else:
loader = TextLoader(file_path)
docs = loader.load()
documents.extend(docs)
except Exception as e:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Error loading {file_path}: {e}")
if documents:
splits = self.text_splitter.split_documents(documents)
if self.vector_store is None:
self.vector_store = FAISS.from_documents(splits, self.embeddings)
else:
self.vector_store.add_documents(splits)
self.qa_chain = RetrievalQA.from_chain_type(
llm=self.llm,
chain_type="stuff",
retriever=self.vector_store.as_retriever(search_kwargs={"k": 3}),
return_source_documents=True
)
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> Added {len(splits)} document chunks to vector store")
def query(self, question: str) -> Dict[str, Any]:
"""Query the RAG system"""
if not self.qa_chain:
return {"answer": "No documents loaded. Please add documents first."}
try:
result = self.qa_chain({"query": question})
return {
"answer": result["result"],
"sources": [doc.metadata for doc in result.get("source_documents", [])]
}
except Exception as e:
return {"answer": f"Error: {str(e)}"}We use ConversationManager to manage multi-session memory, enabling both buffer-based and summary-based chat histories for each session. Then, in OllamaLangChainSystem, we bring all components together, server, LLM, RAG, memory, tools, and agents, into one unified interface. We configure the system to install Ollama, pull models, build agents with tools like web search and RAG, and expose chat, document upload, and model-switching capabilities for seamless interaction.
class ConversationManager:
"""Manage conversation history and memory"""
def __init__(self, llm: OllamaLLM, memory_type: str = "buffer"):
self.llm = llm
self.conversations = {}
self.memory_type = memory_type
def get_conversation(self, session_id: str) -> ConversationChain:
"""Get or create conversation for session"""
if session_id not in self.conversations:
if self.memory_type == "buffer":
memory = ConversationBufferWindowMemory(k=10)
elif self.memory_type == "summary":
memory = ConversationSummaryBufferMemory(
llm=self.llm,
max_token_limit=1000
)
else:
memory = ConversationBufferWindowMemory(k=10)
self.conversations[session_id] = ConversationChain(
llm=self.llm,
memory=memory,
verbose=True
)
return self.conversations[session_id]
def chat(self, session_id: str, message: str) -> str:
"""Chat with specific session"""
conversation = self.get_conversation(session_id)
return conversation.predict(input=message)
def clear_session(self, session_id: str):
"""Clear conversation history for session"""
if session_id in self.conversations:
del self.conversations[session_id]
class OllamaLangChainSystem:
"""Main system integrating all components"""
def __init__(self, config: OllamaConfig):
self.config = config
self.manager = OllamaManager(config)
self.llm = None
self.rag_system = None
self.conversation_manager = None
self.tools = []
self.agent = None
def setup(self):
"""Complete system setup"""
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f680.png" alt="🚀" class="wp-smiley" /> Setting up Ollama + LangChain system...")
self.manager.install_ollama()
self.manager.start_server()
if not self.manager.pull_model(self.config.model_name):
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Failed to pull default model")
return False
self.llm = OllamaLLM(
model_name=self.config.model_name,
base_url=self.config.base_url,
temperature=self.config.temperature,
max_tokens=self.config.max_tokens,
performance_monitor=self.manager.performance_monitor
)
self.rag_system = RAGSystem(self.llm)
self.conversation_manager = ConversationManager(self.llm)
self._setup_tools()
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> System setup complete!")
return True
def _setup_tools(self):
"""Setup tools for the agent"""
search = DuckDuckGoSearchRun()
self.tools = [
Tool(
name="Search",
func=search.run,
description="Search the internet for current information"
),
Tool(
name="RAG_Query",
func=lambda q: self.rag_system.query(q)["answer"],
description="Query loaded documents using RAG"
)
]
self.agent = initialize_agent(
tools=self.tools,
llm=self.llm,
agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
verbose=True
)
def chat(self, message: str, session_id: str = "default") -> str:
"""Simple chat interface"""
return self.conversation_manager.chat(session_id, message)
def rag_chat(self, question: str) -> Dict[str, Any]:
"""RAG-based chat"""
return self.rag_system.query(question)
def agent_chat(self, message: str) -> str:
"""Agent-based chat with tools"""
return self.agent.run(message)
def switch_model(self, model_name: str) -> bool:
"""Switch to different model"""
if self.manager.pull_model(model_name):
self.llm.model_name = model_name
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> Switched to model: {model_name}")
return True
return False
def load_documents(self, file_paths: List[str]):
"""Load documents into RAG system"""
self.rag_system.add_documents(file_paths)
def get_performance_stats(self) -> Dict[str, Any]:
"""Get system performance statistics"""
return self.manager.performance_monitor.get_stats()
def cleanup(self):
"""Clean up resources"""
self.manager.stop_server()
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/2705.png" alt="✅" class="wp-smiley" /> System cleanup complete")We use the ConversationManager to maintain separate chat sessions, each with its memory type, either buffer-based or summary-based, allowing us to preserve or summarize context as needed. In the OllamaLangChainSystem, we integrate everything: we install and launch Ollama, pull the desired model, wrap it in a LangChain-compatible LLM, connect a RAG system, initialize chat memory, and register external tools like web search.
def main():
"""Main function demonstrating the system"""
config = OllamaConfig(
model_name="llama2",
temperature=0.7,
max_tokens=2048
)
system = OllamaLangChainSystem(config)
try:
if not system.setup():
return
print("n<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f5e3.png" alt="🗣" class="wp-smiley" /> Testing basic chat:")
response = system.chat("Hello! How are you?")
print(f"Response: {response}")
print("n<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f504.png" alt="🔄" class="wp-smiley" /> Testing model switching:")
models = system.manager.list_models()
print(f"Available models: {models}")
print("n<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f916.png" alt="🤖" class="wp-smiley" /> Testing agent:")
agent_response = system.agent_chat("What's the current weather like?")
print(f"Agent Response: {agent_response}")
print("n<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f4ca.png" alt="📊" class="wp-smiley" /> Performance Statistics:")
stats = system.get_performance_stats()
print(json.dumps(stats, indent=2))
except KeyboardInterrupt:
print("n<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/23f9.png" alt="⏹" class="wp-smiley" /> Interrupted by user")
except Exception as e:
print(f"<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/274c.png" alt="❌" class="wp-smiley" /> Error: {e}")
finally:
system.cleanup()
def create_gradio_interface(system: OllamaLangChainSystem):
"""Create a Gradio interface for easy interaction"""
try:
import gradio as gr
def chat_interface(message, history, mode):
if mode == "Basic Chat":
response = system.chat(message)
elif mode == "RAG Chat":
result = system.rag_chat(message)
response = result["answer"]
elif mode == "Agent Chat":
response = system.agent_chat(message)
else:
response = "Unknown mode"
history.append((message, response))
return "", history
def upload_docs(files):
if files:
file_paths = [f.name for f in files]
system.load_documents(file_paths)
return f"Loaded {len(file_paths)} documents into RAG system"
return "No files uploaded"
def get_stats():
stats = system.get_performance_stats()
return json.dumps(stats, indent=2)
with gr.Blocks(title="Ollama + LangChain System") as demo:
gr.Markdown("# <img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f999.png" alt="🦙" class="wp-smiley" /> Ollama + LangChain Advanced System")
with gr.Tab("Chat"):
chatbot = gr.Chatbot()
mode = gr.Dropdown(
["Basic Chat", "RAG Chat", "Agent Chat"],
value="Basic Chat",
label="Chat Mode"
)
msg = gr.Textbox(label="Message")
clear = gr.Button("Clear")
msg.submit(chat_interface, [msg, chatbot, mode], [msg, chatbot])
clear.click(lambda: ([], ""), outputs=[chatbot, msg])
with gr.Tab("Document Upload"):
file_upload = gr.File(file_count="multiple", label="Upload Documents")
upload_btn = gr.Button("Upload to RAG System")
upload_status = gr.Textbox(label="Status")
upload_btn.click(upload_docs, file_upload, upload_status)
with gr.Tab("Performance"):
stats_btn = gr.Button("Get Performance Stats")
stats_output = gr.Textbox(label="Performance Statistics")
stats_btn.click(get_stats, outputs=stats_output)
return demo
except ImportError:
print("Gradio not installed. Skipping interface creation.")
return None
if __name__ == "__main__":
print("<img src="https://s.w.org/images/core/emoji/16.0.1/72x72/1f680.png" alt="🚀" class="wp-smiley" /> Ollama + LangChain System for Google Colab")
print("=" * 50)
main()
# Or create a system instance for interactive use
# config = OllamaConfig(model_name="llama2")
# system = OllamaLangChainSystem(config)
# system.setup()
# # Create Gradio interface
# demo = create_gradio_interface(system)
# if demo:
# demo.launch(share=True) # share=True for public linkWe wrap everything up in the main function to run a full demo, setting up the system, testing chat, agent tools, model listing, and performance statistics. Then, in create_gradio_interface(), we build a user-friendly Gradio app with tabs for chatting, uploading documents to the RAG system, and monitoring performance. Finally, we call main() in the __main__ block for direct Colab execution, or optionally launch the Gradio UI for interactive exploration and public sharing.
In conclusion, we have a flexible playground: we switch Ollama models, converse with buffered or summary memory, question our own documents, reach out to search when context is missing, and monitor basic resource stats to stay within Colab limits. The code is modular, allowing us to extend the tool list, tune inference options (temperature, maximum tokens, concurrency) in OllamaConfig, or adapt the RAG pipeline to larger corpora or different embedding models. We launch the Gradio app with share=True to collaborate or embed these components in our projects. We now own an extensible template for fast local LLM experimentation.
Check out the Codes. All credit for this research goes to the researchers of this project. SUBSCRIBE NOW to our AI Newsletter
The post Building a GPU-Accelerated Ollama LangChain Workflow with RAG Agents, Multi-Session Chat Performance Monitoring appeared first on MarkTechPost.
Source: Read MoreÂ
