gradio/guides/05_chatbots/02_creating-a-custom-chatbot-with-blocks.md

How to Create a Custom Chatbot with Gradio Blocks

Tags: NLP, TEXT, CHAT Related spaces: https://huggingface.co/spaces/gradio/chatbot_streaming, https://huggingface.co/spaces/project-baize/Baize-7B,

Introduction

Important Note: if you are getting started, we recommend using the gr.ChatInterface to create chatbots -- its a high-level abstraction that makes it possible to create beautiful chatbot applications fast, often with a single line of code. Read more about it here.

This tutorial will show how to make chatbot UIs from scratch with Gradio's low-level Blocks API. This will give you full control over your Chatbot UI. You'll start by first creating a a simple chatbot to display text, a second one to stream text responses, and finally a chatbot that can handle media files as well. The chatbot interface that we create will look something like this:

$demo_chatbot_streaming

Prerequisite: We'll be using the gradio.Blocks class to build our Chatbot demo. You can read the Guide to Blocks first if you are not already familiar with it. Also please make sure you are using the latest version version of Gradio: pip install --upgrade gradio.

A Simple Chatbot Demo

Let's start with recreating the simple demo above. As you may have noticed, our bot simply randomly responds "How are you?", "I love you", or "I'm very hungry" to any input. Here's the code to create this with Gradio:

$code_chatbot_simple

There are three Gradio components here:

• A Chatbot, whose value stores the entire history of the conversation, as a list of response pairs between the user and bot.
• A Textbox where the user can type their message, and then hit enter/submit to trigger the chatbot response
• A ClearButton button to clear the Textbox and entire Chatbot history

We have a single function, respond(), which takes in the entire history of the chatbot, appends a random message, waits 1 second, and then returns the updated chat history. The respond() function also clears the textbox when it returns.

Of course, in practice, you would replace respond() with your own more complex function, which might call a pretrained model or an API, to generate a response.

$demo_chatbot_simple

Add Streaming to your Chatbot

There are several ways we can improve the user experience of the chatbot above. First, we can stream responses so the user doesn't have to wait as long for a message to be generated. Second, we can have the user message appear immediately in the chat history, while the chatbot's response is being generated. Here's the code to achieve that:

$code_chatbot_streaming

You'll notice that when a user submits their message, we now chain three event events with .then():

1. The first method user() updates the chatbot with the user message and clears the input field. This method also makes the input field non interactive so that the user can't send another message while the chatbot is responding. Because we want this to happen instantly, we set queue=False, which would skip any queue had it been enabled. The chatbot's history is appended with (user_message, None), the None signifying that the bot has not responded.

2. The second method, bot() updates the chatbot history with the bot's response. Instead of creating a new message, we just replace the previously-created None message with the bot's response. Finally, we construct the message character by character and yield the intermediate outputs as they are being constructed. Gradio automatically turns any function with the yield keyword into a streaming output interface.

3. The third method makes the input field interactive again so that users can send another message to the bot.

Of course, in practice, you would replace bot() with your own more complex function, which might call a pretrained model or an API, to generate a response.

Finally, we enable queuing by running demo.queue(), which is required for streaming intermediate outputs. You can try the improved chatbot by scrolling to the demo at the top of this page.

Liking / Disliking Chat Messages

Once you've created your gr.Chatbot, you can add the ability for users to like or dislike messages. This can be useful if you would like users to vote on a bot's responses or flag inappropriate results.

To add this functionality to your Chatbot, simply attach a .like() event to your Chatbot. A chatbot that has the .like() event will automatically feature a thumbs-up icon and a thumbs-down icon next to every bot message.

The .like() method requires you to pass in a function that is called when a user clicks on these icons. In your function, you should have an argument whose type is gr.LikeData. Gradio will automatically supply the parameter to this argument with an object that contains information about the liked or disliked message. Here's a simplistic example of how you can have users like or dislike chat messages:

import gradio as gr

def greet(history, input):
    return history + [(input, "Hello, " + input)]

def vote(data: gr.LikeData):
    if data.liked:
        print("You upvoted this response: " + data.value["value"])
    else:
        print("You downvoted this response: " + data.value["value"])
    

with gr.Blocks() as demo:
    chatbot = gr.Chatbot()
    textbox = gr.Textbox()
    textbox.submit(greet, [chatbot, textbox], [chatbot])
    chatbot.like(vote, None, None)  # Adding this line causes the like/dislike icons to appear in your chatbot
    
demo.launch()

Adding Markdown, Images, Audio, or Videos

The gr.Chatbot component supports a subset of markdown including bold, italics, and code. For example, we could write a function that responds to a user's message, with a bold That's cool!, like this:

def bot(history):
    response = "**That's cool!**"
    history[-1][1] = response
    return history

In addition, it can handle media files, such as images, audio, and video. You can use the MultimodalTextbox component to easily upload all types of media files to your chatbot. To pass in a media file, we must pass in the file as a tuple of two strings, like this: (filepath, alt_text). The alt_text is optional, so you can also just pass in a tuple with a single element (filepath,), like this:

def add_message(history, message):
    for x in message["files"]:
        history.append(((x["path"],), None))  
    if message["text"] is not None:
        history.append((message["text"], None))
    return history, gr.MultimodalTextbox(value=None, interactive=False, file_types=["image"])

Putting this together, we can create a multimodal chatbot with a multimodal textbox for a user to submit text and media files. The rest of the code looks pretty much the same as before:

$code_chatbot_multimodal $demo_chatbot_multimodal

And you're done! That's all the code you need to build an interface for your chatbot model. Finally, we'll end our Guide with some links to Chatbots that are running on Spaces so that you can get an idea of what else is possible:

• project-baize/Baize-7B: A stylized chatbot that allows you to stop generation as well as regenerate responses.
• MAGAer13/mPLUG-Owl: A multimodal chatbot that allows you to upvote and downvote responses.