Introduction to AI Product Development

Picking the stack - Navigating JavaScript and Python

Designing a Hybrid Web Application Architecture with JavaScript and Python

Streaming events with Server-Sent Events and WebSockets

System Design for AI applications

Discovering the OpenAI Completion API

Handling Server-Sent Events with JavaScript

Building the useCompletion hook

Rendering Completion Output

Mocking Streams

Testing the useCompletion hook

In this module, we're diving into the creation of an AI product with a frontend focus.

Building the Front-End

Creating a FastAPI server

Exploring asynchronous programming in Python

Integrating Langchain with FastAPI for Asynchronous Streaming

Testing with PyTest and LangChain

In this module, we're diving into the creation of an AI product with a Backend focus.

Building the Backend

Building the useChat hook

Building the User Interface

In this module, we learn how to build a chat.

Building a Chat

Discovering Retrieval Augmented Generation

Building a Semantic Search Engine with Chroma

Adding Retrieval-Augmented Generation to the chat

In this module, we implement Retrieval Augmented Generation"

Implementing Retrieval Augmented Generation

Final words

Conclusion

Exploring AI Agents - Toward Autonomous Systems

Building an Asynchronous Event-Driven Service Using Context Variables

Adding Tools To The Chat

Building a Generative UI

In this lessons we will explore how AI agents function with a focus on autonomy, tool usage, and self-correction. They will also be introduced to the potential future developments in AI agent technology.

Building an Agent

Join us as we dive into the world of AI-powered streaming text and chat UIs using TypeScript, React, and Python. 

Large Language Models (LLMs) are revolutionizing various industries, but creating simple demos in notebooks or with tools like Gradio is not enough. To truly harness the potential of these models, we need to build sophisticated, real-world applications.

This comprehensive course is structured into six detailed modules where you will:
- Develop a dynamic front-end with React.
- Integrate FastAPI with LangChain to create a robust backend.
- Utilize Server-Sent Events (SSE) for seamless real-time data streaming.

This comprehensive course comes packed with valuable extras:
- The complete source code under MIT licence
- Exclusive access to our community forum
- Detailed exploration of Retrieval-Augmented Generation (RAG) for enhancing model responses
- Hands-on experience with developing intelligent agents capable of maintaining context in conversations

Taught by [Louis Sanna](https://www.linkedin.com/in/louis-sanna-0168363a/), a seasoned Tech Lead with expertise in GenAI applications and digital transformation projects for major organizations like UNESCO and Renault. His deep industry insights and practical experience will guide you from basics to advanced implementations.

Taking this course will spare you countless days of troubleshooting and challenges, offering a shortcut through the learning curve. Moreover, it presents a unique opportunity to engage with one of the most significant technological trends of our era. This is your chance to become a part of the groundbreaking wave of LLM advancements, positioning you at the forefront of innovation.

### Problem this Course Solves

This course is designed to tackle specific problems that AI engineers face when integrating LLM capabilities into web applications. Streaming text presents a unique challenge for engineers, requiring the integration of multiple technologies and concepts. This course addresses this by:

- Exploring the use of Langchain for seamless LLM provider switching. Given Langchain's rapid development, understanding its streaming, asynchronous, and callback functionalities can be daunting.
- Demystifying asynchronous programming in Python, which can be more complex than in JavaScript. We'll cover the differences between asynchronous and synchronous programming, the benefits of asynchronous operations, and how to implement and integrate asynchronous function calls within a streaming context.
- Selecting and implementing a protocol for streaming text from the backend to the frontend. We will explain why Server-Sent Events (SSE) is an ideal standard, despite being relatively new and unfamiliar to many.
- Creating an SSE endpoint using a Python server framework, such as FastAPI, to facilitate real-time data streaming.
- Reading the stream in the front-end using the ReadableStream interface of the Fetch API, ensuring efficient data handling.
- Updating the React state of prompts with consideration for React's batching behavior to prevent unexpected issues during rapid updates.
- Integrating custom components within text responses to enable dynamic, on-the-fly generative UIs.

By the end of this course, you will have a comprehensive understanding of how to effectively integrate and manage LLM capabilities within web applications, overcoming the technical hurdles associated with real-time text streaming.

### Preview the Application

Experience the power of what you'll build firsthand. Try out the demo app we will develop during the course:
[The Demo App](https://build-ai-driven-web-applications.vercel.app/)

Enroll now and start building transformative AI-powered applications that streamline communication and enhance user engagement!

### Course Content Overview

##  Basic Package
- Total Modules: **6**
- Total Lessons: **20**
- Total Video Runtime: **1 hour, 10 minutes**

##  Advanced Package
- Total Modules: **7**
- Total Lessons: **24**
- Total Lines of Code: **3706**
- Total Video Runtime: **1 hour, 34 minutes**

Dive into Retrieval Augmented Generation and Autonomous Agents with LangChain, Chroma and FastAPI

Large Language Models are reshaping industries, yet integrating them into real-time streaming UIs presents unique challenges. In this course we will learn how to seamlessly integrate LLM APIs into applications and build AI-powered streaming text and chat UIs with TypeScript, React, and Python. Step-by-step, we will build a full-stack AI application with quality code and very flexible implementation.

The LLM application in this course includes:
- Completion Use-Case (english to emojis)
- Chat
- Retrieval Augmented Generation use-case
- AI Agent Use-Cases (code execution, data-Analyste agent)

This app can be used as a starting point in most projects, saving a huge amount of time, and its flexibilty allows new tools to be added as needed.

At the end of this course, you will have mastered end-to-end implementation of a flexible and high-quality LLM application. This course will also equip you with the knowledge and skills necessary to create sophisticated LLM solutions of your own.

Responsive LLM Applications with Server-Sent Events

How to design systems for AI applications

How to stream the answer of a Large Language Model

Differences between Server-Sent Events and WebSockets

How asynchronous programming in Python works

What problems Retrieval Augmented Generation can solve

In this lesson, we'll focus on integrating and managing tools using large language models (LLMs). Let's dive into how these tools can be applied to enhance functionality and user interaction.

First, let's examine our demo application which includes an array of tools such as the addition function. This simple example will help us understand how tools are structured and executed within our application.

def add(a: int, b: int) -> int:
    """Return the sum of two integers."""
    return a + b

This function is straightforward—it takes two integers, computes their sum, and returns the result. The inclusion of descriptive docstrings and type annotations aids the model in utilizing the function correctly.

Code sandboxes can be powerful tools, they also come with potential risks such as security vulnerabilities, race conditions, and the possibility of infinite loops.

A preferred approach is to use an external container solution like e2b. This provides a more isolated and manageable environment for code execution.

Here's a basic example of how you might set up and use e2b (

) for code execution (requires an API key):

import asyncio
import multiprocessing
from e2b_code_interpreter import CodeInterpreter

TIMEOUT = 3


def _execute(code: str) -> dict:
    results = ""
    stdout = ""
    stderr = ""
    pngs = []
    with CodeInterpreter() as sandbox:
        try:
            execution = sandbox.notebook.exec_cell(code, timeout=TIMEOUT)
            pngs = [r.png for r in execution.results if r.png]
            results = "\n".join([r.text for r in execution.results])
            stdout = "\n".join([out for out in execution.logs.stdout])
            stderr = "\n".join([err for err in execution.logs.stderr])
            if execution.error:
                stderr = f"{execution.error.name}: {execution.error.value}"
        except TimeoutError:
            stderr = "Execution timed out."
        except Exception as e:
            stderr = f"Unexpected error: {e}"

    return {"results": results, "stdout": stdout, "stderr": stderr, "pngs": pngs}


async def exec_python(code: str) -> dict:
    loop = asyncio.get_running_loop()
    with multiprocessing.Pool(1) as pool:
        result = await loop.run_in_executor(None, pool.apply, _execute, (code,))
    return result

---
title: Adding Tools To The Chat
isPublicLesson: true
privateVideoUrl: https://stream.mux.com/BOX93pMZHpTRsBshYjPefmwKGAjddtAxnalp5DPf2lg.m3u8
code: https://gitlab.com/fullstackio/books/newline-course-apps/louis-sanna-real-time-streaming-for-generative-ui-app/-/tree/main/module-7
---

# Adding Tools To the Chat

In this lesson, we'll focus on integrating and managing tools using large language models (LLMs). Let's dive into how these tools can be applied to enhance functionality and user interaction.

### Add function example

First, let's examine our demo application which includes an array of tools such as the addition function. This simple example will help us understand how tools are structured and executed within our application.

**Add Function Example:**

```python
def add(a: int, b: int) -> int:
    """Return the sum of two integers."""
    return a + b
```

This function is straightforward—it takes two integers, computes their sum, and returns the result. The inclusion of descriptive docstrings and type annotations aids the model in utilizing the function correctly.

### Adding a code sandbox

Code sandboxes can be powerful tools, they also come with potential risks such as security vulnerabilities, race conditions, and the possibility of infinite loops.

A preferred approach is to use an external container solution like e2b. This provides a more isolated and manageable environment for code execution.

Here's a basic example of how you might set up and use e2b (<https://github.com/e2b-dev/E2B>) for code execution (requires an API key):

```python
import asyncio
import multiprocessing
from e2b_code_interpreter import CodeInterpreter

TIMEOUT = 3


def _execute(code: str) -> dict:
    results = ""
    stdout = ""
    stderr = ""
    pngs = []
    with CodeInterpreter() as sandbox:
        try:
            execution = sandbox.notebook.exec_cell(code, timeout=TIMEOUT)
            pngs = [r.png for r in execution.results if r.png]
            results = "\n".join([r.text for r in execution.results])
            stdout = "\n".join([out for out in execution.logs.stdout])
            stderr = "\n".join([err for err in execution.logs.stderr])
            if execution.error:
                stderr = f"{execution.error.name}: {execution.error.value}"
        except TimeoutError:
            stderr = "Execution timed out."
        except Exception as e:
            stderr = f"Unexpected error: {e}"

    return {"results": results, "stdout": stdout, "stderr": stderr, "pngs": pngs}


async def exec_python(code: str) -> dict:
    loop = asyncio.get_running_loop()
    with multiprocessing.Pool(1) as pool:
        result = await loop.run_in_executor(None, pool.apply, _execute, (code,))
    return result
```

### Registering the tools

To make various tools available within our chat application, we first define a set of functions and then specifically register the ones we need for a session. Here's how this is done using LangChain:

```python
all_functions = {
    "add": add,
    "divide": divide,
    "exec_python": exec_python,
    "get_weather": get_weather,
    "multiply": multiply,
    "query_titanic": query_titanic,
    "search_IPCC": search_IPCC,
}

available_functions = {
        key: value for key, value in all_functions.items() if key in tools
    }
# https://python.langchain.com/docs/modules/model_io/chat/function_calling
chat = chat.bind_tools(available_functions.values())
```

The bind_tools method connects the selected tools to the chat instance, making it possible for the chat model to utilize these tools based on user inputs or internal logic. This approach ensures that only necessary tools are loaded, optimizing resource use and performance.

### Managing Tool Calls with a Loop

Next, we dive into the loop that manages tool calls, designed to handle errors, and facilitate retries which allows the model to correct its own mistakes or complete tasks that were only partially solved on the first attempt.

**Loop Mechanism for Tool Calls:**

```python
while completion_count < MAX_ITERATIONS:
    message_id = str(uuid.uuid4())
    new_assistant_message = {
        "role": "assistant",
        "content": "",
        "events": [],
        "tool_calls": [],
        "id": message_id,
    }
    await emit_event(NewMessageEvent(message={**new_assistant_message}))
    truncated_messages = truncate_message(messages)
    response_content, tool_calls = await complete(chat, truncated_messages)
    messages.append(new_assistant_message)
    new_assistant_message["content"] = response_content
    new_assistant_message["tool_calls"] = tool_calls
    await emit_event(
        MessagePatchEvent(
            message_id=message_id,
            patch=[{"op": "add", "path": "/tool_calls", "value": tool_calls}],
        )
    )
    if not tool_calls:
        break

     for tool_call in tool_calls:
            if tool_call["type"] != "function":
                continue
            new_tool_message = await process_tool_call(tool_call, message_id)
            messages.append(new_tool_message)
```

In this segment, the loop runs until no more tool calls are necessary or a maximum number of iterations is reached. This structure is critical to prevent infinite loops and ensure efficient handling of tasks.

### Calling the tool

After tool calls are processed, we generate a of message of type `role` that includes the result of the tool's use. This step is vital for maintaining coherent communication and ensuring the continuity of the tool's interaction within the chat.

Note that the `tool_call_id` should match the id of a tool_calls of the previous assistant message.

```python
async def process_tool_call(tool_call, message_id):
    function_response = await call_tool(tool_call, message_id=message_id)
    new_content = truncate_content(
        function_response if isinstance(function_response, str)
        else yaml.dump(function_response), tokens_max=4000
    )
    new_tool_message = {
        "id": str(uuid.uuid4()),
        "role": "tool",
        "tool_call_id": tool_call["id"],
        "content": new_content,
    }
    await emit_event(NewMessageEvent(message={**new_tool_message}))
    return new_tool_message
```