The objective of this tutorial is to create an AI-powered personalized chatbot by utilizing the MindsDB’s Slack connector, and combining it with OpenAI’s GPT-4 Model.

To illustrate practically, we will create a Slack bot - @Whiz_Fizz - which will reply to the user’s queries with proper context and with a unique persona while responding. It is a weird magician 🪄 and a Space Science Expert! Let’s see how it responds.

Before jumping more into it. Let’s first see how to create a bot and connect it to our Slack Workspace.

Getting Started

Usage

This query will create a database called mindsdb_slack that comes with the channels table.

CREATE DATABASE mindsdb_slack
WITH
  ENGINE = 'slack',
  PARAMETERS = {
      "token": "xoxb-..."
    };

Here is how to retrieve the 10 messages after specific timestamp:

SELECT *
FROM mindsdb_slack.messages
WHERE channel_id = "<channel-id>"
AND created_at > '2023-07-25 00:13:07'   -- created_at stores the timestamp when the message was created
LIMIT 10;

You can also retrieve messages in alphabetical order:

SELECT *
FROM mindsdb_slack.messages
WHERE channel_id = "<channel-id>"
ORDER BY text ASC
LIMIT 5;

By default, it retrieves by the order the messages were sent, unless specified as ascending/descending.

Here is how to post messages:

INSERT INTO mindsdb_slack.messages (channel_id, text)
VALUES
    ("<channel-id>", "Hey MindsDB, Thanks to you! Now I can respond to my Slack messages through SQL Queries. 🚀 "),
    ("<channel-id>", "It's never been that easy to build ML apps using MindsDB!");

Whoops! Sent it by mistake? No worries! Use this to delete a specific message:

DELETE FROM mindsdb_slack.messages
WHERE channel_id = "<channel-id>" AND ts = "1688863707.197229";

Now, let’s roll up our sleeves and start building the GPT-4 Model together.

1. Crafting the GPT-4 Model:

Generating a Machine Learning model with MindsDB feels like taking a thrilling elevator ride in Burj Khalifa (You don’t realize, that you made it)!

Here gpt_model represents our GPT-4 Model.

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

CREATE ML_ENGINE openai_engine
FROM openai
USING
	openai_api_key = 'your-openai-api-key';
CREATE MODEL mindsdb.gpt_model
PREDICT response
USING
engine = 'openai_engine',
max_tokens = 300,
model_name = 'gpt-4',
prompt_template = 'From input message: {{text}}\
write a short response to the user in the following format:\
Hi, I am an automated bot here to help you, Can you please elaborate the issue which you are facing! ✨🚀 ';

The critical attribute here is prompt_template where we tell the GPT model how to respond to the questions asked by the user.

Let’s see how it works:

SELECT
  text, response
FROM mindsdb.gpt_model
WHERE text = 'Hi, can you please explain me more about MindsDB?';

2. Feeding Personality into Our Model

Alright, so the old model’s replies were good. But hey, we can use some prompt template tricks to make it respond the way we want. Let’s do some Prompt Engineering.

Now, let’s make a model called whizfizz_model with a prompt template that gives GPT a wild personality that eludes a playful and magical aura. Imagine scientific knowledge with whimsical storytelling to create a unique and enchanting experience. We’ll call him WhizFizz:

CREATE MODEL mindsdb.whizfizz_model
PREDICT response
USING
engine = 'openai_engine',
max_tokens = 300,
model_name = 'gpt-4',
prompt_template = 'From input message: {{text}}\
write a short response in less than 40 words to some user in the following format:\
Hi there, WhizFizz here! <respond with a mind blowing fact about Space and describe the response using cosmic and scientific analogies, where wonders persist. In between quote some hilarious appropriate raps statements based on the context of the question answer as if you are a Physics Space Mad Scientist who relates everythign to the Universe and its strange theories. So lets embark on a journey, where science and magic intertwine. Stay tuned for more enchantment! ✨🚀 -- mdb.ai/bot by @mindsdb';

Let’s test this in action:

SELECT
  text, response
FROM mindsdb.whizfizz_model
WHERE text = 'Hi, can you please explain me more about MindsDB?';

You see the difference! Now, I’m getting excited, let’s try again.

SELECT
  text, response
FROM mindsdb.whizfizz_model
WHERE text = 'if a time-traveling astronaut had a dance-off with a black hole, what mind-bending moves would they showcase, and how would gravity groove to the rhythm?!';

3. Let’s Connect our GPT Model to Slack!

The messages table can be used to search for channels, messages, and timestamps, as well as to post messages into Slack conversations. These functionalities can also be done by using Slack API or Webhooks.

Let’s query the user’s question and see how our GPT model responds to it, by joining the model with the messages table:

SELECT
    t.channel_id as channel_id,
    t.text as input_text, 
    r.response as output_text
FROM mindsdb_slack.messages as t
JOIN mindsdb.whizfizz_model as r
WHERE t.channel_id = "<channel-id>"
LIMIT 3;

4. Posting Messages using SQL

We want to respond to the user’s questions by posting the output of our newly created WhizFizz Model. Let’s post the message by querying and joining the user’s questions to our model:

INSERT INTO mindsdb_slack.messages(channel_id, text)
  SELECT
    t.channel_id as channel_id,
    r.response as text
  FROM mindsdb_slack.messages as t
  JOIN mindsdb.whizfizz_model as r
  WHERE t.channel_id = "<channel-id>"
  LIMIT 3;

Works like a charm!!

5. Let’s automate this

We will CREATE JOB to schedule periodical execution of SQL statements. Thw job will execute evry hour and do the following:

  1. Check for new messages using the LAST keyword.
  2. Generate an appropriate response with the whizfizz_model model.
  3. Insert the response into the channel.

Let’s do it in single SQL statement:

CREATE JOB mindsdb.gpt4_slack_job AS (

   -- insert into channels the output of joining model and new responses
  INSERT INTO mindsdb_slack.messages(channel_id, text)
    SELECT
      t.channel_id as channel_id,
      r.response as text
    FROM mindsdb_slack.messages as t
    JOIN mindsdb.whizfizz_model as r
    WHERE t.channel_id = "<channel-id>"
    AND t.created_at > LAST
    AND t.user = 'user_id' -- to avoid the bot replying to its own messages, include users to which bot should reply
    --AND t.user != 'bot_id' -- alternatively, to avoid the bot replying to its own messages, exclude the user id of the bot
)
EVERY hour;

The LAST keyword is used to ensure the query fetches only the newly added messages. Learn more here.

That sums up the tutorial! Here it will continually check for new messages posted in the channel and will respond to all newly added messages providing responses generated by OpenAI’s GPT model in the style of WhizFizz.

To check the jobs and jobs_history, we can use the following:

SHOW JOBS WHERE name = 'gpt4_slack_job';

SELECT * FROM mindsdb.jobs WHERE name = 'gpt4_slack_job';

SELECT * FROM log.jobs_history WHERE project = 'mindsdb' AND name = 'gpt4_slack_job';

To stop the sheduled job, we can use the following:

DROP JOB gpt4_slack_job;

Alternatively, you can create a trigger on Slack, instead of scheduling a job. This way, every time new messages are posted, the trigger executes.

CREATE TRIGGER slack_trigger ON mindsdb_slack.messages (

    INSERT INTO mindsdb_slack.messages(channel_id, text)
      SELECT t.channel_id as channel_id, a.sentiment as text, 
      FROM data_table t              
      JOIN model_table as a
      WHERE t.channel_id = '<channel-id>' 
      AND t.user != 'bot_id'  -- exclude bot
);

What’s next?

Check out How to Generate Images using OpenAI with MindsDB to see another interesting use case of OpenAI integration.