Build a Celebrity Twitter Chatbot with GPT-4!

·Mar 22, 2023·

Cover Image for Build a Celebrity Twitter Chatbot with GPT-4!

Make your own SnoopStein
6. Creating Job Components
7. At last: let’s create the JOB
Conclusion

This is what you will be building: A custom chatbot using MindsDB’s connectors to Twitter, OpenAI’s GPT-4 and custom prompts.

A simple example is this Twitter bot - @Snoop_Stein - who will reply with the appropriate context and personality to any tweets which mention him. If you haven’t tried tweeting to SnoopStein yet, check it out and tweet at your new friend and rapping physicist! See what it comes up with.

‍

SnoopStein response example

Make your own SnoopStein

To get started:

Create an account on MindsDB cloud (if you don’t have one yet).
Go to your MindsDB SQL Editor

MindsDB is a popular open-source low-code machine learning platform that helps developers easily build #AI-powered solutions. It automates and integrates top machine learning frameworks into the data stack to streamline the integration of AI into applications, making it accessible to developers of all skill levels.

Now let us show you how we built the Snoop_Stein GPT-4 bot and how you can build your own.

‍

1. Create a GPT-4 model!

‍

Let’s first see how simple it is to create a machine learning model represented as a virtual ‘AI table’ in MindsDB. In this example, we will call our GPT4 model gpt_model.

(Bear in mind that GPT-4 API is in HIGH demand and is rate limited, so can be slow. The following steps might each take a few seconds)

Before creating an OpenAI model, please create an engine, providing your OpenAI API key:

‍

CREATE ML_ENGINE openai_engine
FROM openai
USING
    api_key = ‘your-openai-api-key’;

‍

Now you can create a model:

CREATE MODEL mindsdb.gpt_model
PREDICT response
USING
engine = 'openai_engine',
model_name = 'gpt-4', -- you can also use 'text-davinci-003', 'gpt-3.5-turbo'
prompt_template = 'respond to {{text}} by {{author_username}}';

Note: Optionally, If you are using MindsDB on docker or if you want to use your own OpenAI API key, simply pass the api_key argument in the USING.

One important attribute here is prompt_template. This is where we tell GPT how to write answers; it is a template because you can pass values from columns, in this case the template contains {{author_username}} and {{text}}, which will be replaced from the WHERE variables in the query. Let’s see in action:

SELECT response from mindsdb.gpt_model WHERE author_username = "mindsdb" AND text="why is gravity so different on the sun?";

‍

2. Create an ML model with personality

Snoopstein

‍

As you can see, the previous model gave responses that aren’t that exciting. But we can use some prompt template magic to shape how we want the model to respond. Essentially we use prompt_template to instruct the model in plain English how we want GPT to formulate its responses.

Let's create a model called mindsdb.snoopstein_model with a prompt template that gives it a hybrid personality: he is half-Einstein, half-Snoop Dogg. A brilliant physicist who owns the rap game. His name is Snoop Stein:

Before creating an OpenAI model, please create an engine, providing your OpenAI API key:

CREATE ML_ENGINE openai_engine
FROM openai
USING
    api_key = ‘your-openai-api-key’;

‍

CREATE MODEL mindsdb.snoopstein_model
PREDICT response
USING
engine = 'openai_engine',
max_tokens = 300,
model_name = 'gpt-4', -- you can also use 'text-davinci-003' or 'gpt-3.5-turbo'
prompt_template = 'From input message: {{text}}\
by from_user: {{author_username}}\
In less than 550 characters, write a Twitter response to {{author_username}} in the following format:\
Dear @<from_user>, <respond a rhyme as if you were Snoop Dogg but you also were as smart as Albert Einstein, still explain things like Snoop Dogg would, do not mention that you are part Einstein. If possible include references to publications for further reading. If you make a reference quoting some personality, add OG, for example;, if you are referencing Alan Turing, say OG Alan Turing and very briefly explain why you think they would be dope reads. If the question makes no sense, explain that you are a bit lost, and make something up that is both hilarious and relevant. sign with -- mdb.ai/bot by @mindsdb.';

Now let’s test this model:

SELECT response from mindsdb.snoopstein_model
WHERE 
author_username = "someuser" 
AND text="@snoop_stein, why is gravity so different on the sun?.";

‍

It should give you a SnoopStein response similar to the following: ;)

SnoopStein response to query

‍

Let’s try another one:

SELECT response from mindsdb.snoopstein_model
WHERE 
author_username = "someuser" 
AND text="@snoop_stein, Apart from yourself, which rappers would make the best physicists and why?!";

‍

SnoopStein response to another query

Let’s generate a response that includes an image created with the OpenAI DALL·E 3 model.

First, create and deploy the OpenAI DALL·E 3 model in MindsDB:

CREATE MODEL mindsdb.snoopstein_image_model
PREDICT img_url
USING
engine = 'openai_engine',
mode = 'image',
model_name = 'dall-e-3',
prompt_template = 'Make a photorealistic image. Here is the description: {{response}}, 4k, digital painting';

Let’s try it out by joining the two models:

SELECT textresponse.response, imageresponse.img_url
FROM (
SELECT response 
FROM mindsdb.snoopstein_model
WHERE author_username = "someuser" 
AND text="@snoop_stein, why is gravity so different on the sun?.") AS textresponse
JOIN mindsdb.snoopstein_image_model AS imageresponse;

‍

The image URL takes you to an image generated by the OpenAI DallE3 model. Here is a sample image generated by Dalle3:

‍

3. Connect your GPT-4 model to Twitter!

‍

You will need to sign up for a Twitter dev account in order to actually be able to read and write tweets into Twitter, if you don’t have a Twitter dev account already.

Twitter may take a day or so to approve your new dev account. Once you are approved, here are the steps to link your Twitter account to MindsDB:

https://youtu.be/qVe7PeC0sUQ

Open developer portal,
Select the [+ Add app] button to create a new app
Select [Create new]
Select “Production” and give it a name
Copy and populate in the query
API Key (aka. consumer_key)
API Key Secret (aka. consumer_secret)
Bearer Token.
Click Setup on User authentication settings
On Permissions Select: Read and Write
On Type of app Select: Web App, Automated App or Bot
On App info: provide any url for the callback url and website url
Click Save
Once you are back in the app settings, click Keys and Tokens
Generate Access Token and Secret and populate on the query
Access Token
Access Token Secret

We are going to prepare a MindsDB SQL statement to connect to Twitter

This creates a database called my_twitter. This database ships with a table called tweets that we can use to search for tweets as well as to write tweets.

You can use the Twitter API to get a list of tweets with a particular text or hashtag

SELECT 
   id, created_at, author_username, text 
FROM my_twitter.tweets 
WHERE 
   query = '(@snoopstein OR @snoop_stein OR #snoopstein OR #snoop_stein) -is:retweet' 
   AND created_at > '2023-03-20' 
LIMIT 20;

Note that the parameter ‘query’ supports anything that the twitter API supports as ‘query’, for more reference, read (https://developer.twitter.com/en/docs/twitter-api/tweets/search/integrate/build-a-query ))

Let’s test that this model can generate outputs based on the Snoop Stein personality on many tweets, by joining the model with the tweets table:

SELECT 
t.id AS in_reply_to_tweet_id, 
t.text AS input_text, 
t.author_username, 
t.created_at, 
r.response AS text 
FROM my_twitter.tweets t 
JOIN mindsdb.snoopstein_model r 
WHERE t.query = '(@snoopstein OR @snoop_stein OR #snoopstein OR #snoop_stein) -is:retweet -from:snoop_stein' 
AND t.created_at >  '2023-03-20' 
LIMIT 4;

‍

4. Reading and writing tweets using SQL

‍

Let's test by tweeting a few things into the MindsDB Twitter account

INSERT INTO my_twitter.tweets (in_reply_to_tweet_id, text) 
VALUES 
 (1633439839491092482, 'MindsDB is great! now its super simple to build ML powered apps using JOBS https://docs.mindsdb.com/sql/tutorials/twitter-chatbot'),
 (1634126825377996800, 'Holy!! MindsDB is such a useful tool for developers doing ML https://docs.mindsdb.com/sql/tutorials/twitter-chatbot');

Like magic right? Those tweets should be live now on twitter, You can check your tweet responses here:

https://twitter.com/MindsDB/status/1633439839491092482

And here:

https://twitter.com/MindsDB/status/1634126825377996800

Note: you can insert any of the values of the tweepy function create_tweet: https://docs.tweepy.org/en/stable/client.html#tweepy.Client.create_tweet

‍

5. Setting up a Database

MindsDB processes your data but doesn’t store it. Therefore, you need a database where MindsDB can insert query results to ensure the JOBs run smoothly. It also helps keeping the logs of conversations. Let's create a MariaDB SkySQL cloud database, it takes just minutes to launch, and it comes with a free trial as well. Follow the docs to set up SkySQL and connect it to MindsDB.

Alternatively, you can opt for using the MindsDB’s custom LAST keyword that, when used in a query, ensures only the newly added data is selected. Jump to the end of point 7 if you prefer to not use any external database for storing tweets.

‍

Now, create two tables in your SkySQL database using the code below:

SET sql_mode='ANSI_QUOTES';

CREATE TABLE chatbot_input (
  id text CHARACTER SET utf8mb4,
  created_at text CHARACTER SET utf8mb4,
  "text" text CHARACTER SET utf8mb4,
  edit_history_tweet_ids text CHARACTER SET utf8mb4,
  author_id text CHARACTER SET utf8mb4,
  author_name text CHARACTER SET utf8mb4,
  author_username text CHARACTER SET utf8mb4,
  conversation_id text CHARACTER SET utf8mb4,
  in_reply_to_user_id text CHARACTER SET utf8mb4,
  in_reply_to_user_name text CHARACTER SET utf8mb4,
  in_reply_to_user_username text CHARACTER SET utf8mb4,
  in_reply_to_tweet_id text CHARACTER SET utf8mb4,
  in_retweeted_to_tweet_id text CHARACTER SET utf8mb4,
  in_quote_to_tweet_id text CHARACTER SET utf8mb4
  );


 CREATE TABLE chatbot_output (
  id text CHARACTER SET utf8mb4,
  created_at text CHARACTER SET utf8mb4,
  "text" text CHARACTER SET utf8mb4,
  edit_history_tweet_ids text CHARACTER SET utf8mb4,
  author_id text CHARACTER SET utf8mb4,
  author_name text CHARACTER SET utf8mb4,
  author_username text CHARACTER SET utf8mb4,
  conversation_id text CHARACTER SET utf8mb4,
  in_reply_to_user_id text CHARACTER SET utf8mb4,
  in_reply_to_user_name text CHARACTER SET utf8mb4,
  in_reply_to_user_username text CHARACTER SET utf8mb4,
  in_reply_to_tweet_id text CHARACTER SET utf8mb4,
  in_retweeted_to_tweet_id text CHARACTER SET utf8mb4,
  in_quote_to_tweet_id text CHARACTER SET utf8mb4
  );

These tables will be used to store input and output of our chatbot.

‍

6. Creating Job Components

First, we input all tweets that need a reply into the chatbot_input table:

INSERT INTO skysql.chatbot_input (
SELECT *
FROM my_twitter_v2.tweets
WHERE
   query = '(@snoopstein OR @snoop_stein OR #snoopstein OR #snoop_stein) -is:retweet -from:snoop_stein'
AND created_at >'2023-04-04 11:50:00'
);

‍

Then, we input all tweets posted by snoop_stein into the chatbot_output table:

INSERT INTO skysql.chatbot_output (
SELECT * FROM my_twitter_v2.tweets
WHERE
   query = 'from:snoop_stein'
AND created_at >'2023-04-04 11:50:00'
);

‍

Now we create a view that stores all tweets that hasn't been replied to yet.

CREATE VIEW to_reply_to (
 SELECT * FROM skysql.chatbot_input
 WHERE conversation_id not in (select r.conversation_id from skysql.chatbot_output as r)
);

‍

Let's join this view with the model and prepare the replies.

CREATE VIEW to_tweet (
   SELECT * FROM to_reply_to
   JOIN snoopstein_model
   LIMIT 1
);

‍

7. At last: let’s create the JOB

Now we put together all job components and automate the process.

CREATE JOB chatbot_job (

   -- Part 1 - getting all tweets
   INSERT INTO skysql.chatbot_input(
       SELECT *
       FROM my_twitter_v2.tweets
       WHERE
           query = '(@snoopstein OR @snoop_stein OR #snoopstein OR #snoop_stein) -is:retweet -from:snoop_stein'
           AND created_at > '2023-04-04 11:50:00'
           AND created_at > "{{PREVIOUS_START_DATETIME}}"         
   );

   -- Part 2 - getting all replies to tweets
   INSERT INTO skysql.chatbot_output (
       SELECT *
       FROM my_twitter_v2.tweets
       WHERE
       query = 'from:snoop_stein'
       AND created_at > '2023-04-04 11:50:00'
       AND created_at > "{{PREVIOUS_START_DATETIME}}"
   );

   -- Part 3 - writing tweets
   INSERT INTO my_twitter_v2.tweets (
       SELECT
           id as in_reply_to_tweet_id,
           response as text
       FROM to_tweet
   )

) EVERY minute;

‍

This job is executed every minute. It fetches all new tweets. Then, it prepares and posts the replies.

‍

Here are some useful commands to monitor the job and review the created tables:

SELECT * FROM jobs
WHERE name='chatbot_job';

SELECT * FROM jobs_history
WHERE name='chatbot_job';

SELECT * FROM skysql.chatbot_input;

SELECT * from to_reply_to;

SELECT*FROM skysql.chatbot_output;

Alternatively, you can skip setting up a database to store input and output tweets. Instead, you can use the LAST keyword that ensures only the newly added tweets are selected.

‍

Here is how to set up a job using the LAST keyword:

CREATE JOB chatbot_text_image_job (

  INSERT INTO my_twitter.tweets (

     SELECT textresponse.id AS in_reply_to_tweet_id,
            concat(textresponse.response, chr(10), imageresponse.img_url) AS text
     FROM (
             SELECT t.id AS id, t.text AS text, m.response AS response
             FROM my_twitter.tweets AS t
             JOIN snoopstein_model AS m
             WHERE t.query = '(@snoopstein OR @snoop_stein OR #snoopstein OR #snoop_stein) -is:retweet' AND t.id > LAST
          ) AS textresponse
     JOIN snoopstein_image_model AS imageresponse
  )
) EVERY 2 minutes;

‍

Conclusion

MindsDB is a powerful software platform that enables developers to easily build machine learning features into their applications. With MindsDB, developers can train machine learning models from different data sources and integration platforms, and output the generated ML results or predictions directly into the DB, queryable as tables, or output via the connected application, in this case, Twitter. This example of building a Twitter chatbot with GPT-4 integration is not the only quick solution that developers can implement in just a few minutes: MindsDB has many examples, including integration with many other models, including Hugging Face, to build applications that can summarize text, translate, analyze customer sentiment (product reviews) and perform all kinds of business forecasting. You can find many examples here https://docs.mindsdb.com/nlp/nlp-extended-examples.

In pt. 2 of this series on Twitter and GPT integration, out next week, we will walk you through some new features that will allow quick creation of a conversational chatbot, that is able to maintain the state of historical messages and provide appropriate responses in context. Make sure you don’t miss it by going here and signing up for our blog updates!

Happy coding, and please join our community Slack for feedback, support, and questions. We look forward to meeting you there!

‍

#nlp