Building with BLAZE — Creating LLM-powered Webex Bots for Interacting with Meetings

In our first blog post, Introducing BLAZE (hyperlinked), Cisco Research proudly presented BLAZE – Build Language Applications Easily, a flexible, standardized, no-code, open-source platform to easily assemble, modify, and deploy various NLP models, datasets, and components. In our second blog post, we will walk through creating an NLP pipeline using Meeting Transcripts, BERT, BART, GPT, and the Webex Chatbot Interface. We'll show how BLAZE makes the creation of NLP solutions simple, efficient, and flexible — as easy as Build, Execute, Interact, or 1-2-3!

Installing BLAZE

To install BLAZE, we can follow the instructions in the README on BLAZE's GitHub page, shown below:

(Cisco Open Source's GitHub - BLAZE Repo) https://github.com/cisco-open/Blaze#installation

There are many options to install BLAZE, including, but not limited to: Docker, PyEnv, Conda, Shell.

Once the installation is ready to go, we can get started with creating our Webex Chatbot!

Waltzing with Webex

Cisco Webex enables seamless communication in the digital landscape, allowing users to meet virtually with a suite of intelligent tools at their disposal. One of these tools is Webex's Meeting Assistant, which automatically records, transcribes, and stores transcripts for each Webex meeting.

However, with dozens of weekly meetings, one concern is the utility of these stored meeting transcripts. Given their lengthy durations as well as their numerous quantity, it is infeasible to quickly get a recap of each meeting or locate important information buried under hundreds of transcripts.

While Webex includes search features through its website, we want to make them more conversational and searchable semantically on the Webex Messaging App. Our solution can also be extended easily to other meeting or messaging platforms. 

To address these issues, we can quickly design a Webex Bot using BLAZE that takes in these meeting transcripts and enables semantic search, summarization, and actionable extraction across all of them!

Before we begin, we need to retrieve our Webex Access Token and Bot Access Token. 

Building our Bot — Webex Access Token

• First, we can visit https://developer.webex.com/docs/getting-started

• Upon logging in (top right button), we can select the copy button underneath the subheading "Your Personal Access Token." This string is our Webex Access Token, valid for a period of 12 hours.

Building our Bot — Bot Access Token

• First, we can visit developer.webex.com, which will take us to the webpage below:

• Upon logging in (top right button), we can select our profile and click "My Webex Apps."

• Next, we can create a new App by selecting "Create a New App" followed by "Create a Bot."

• Then, we can fill in the required details on the provided page. The "Bot name" and "Bot username" should be unique. Please make sure to keep this information handy for later! 

• Keep track of the "Bot Access Token" for later - this will be used within BLAZE.

• Finally, navigate to Cisco Webex and search for your bot using its username (in our case, for example, testBlaze@webex.bot). You'll find your bot listed in your Webex directory.

Booting up the Bot

Our journey continues with BLAZE, which operates in three steps: Build, Execute, and Interact. 

Build - Specifying the Pipeline

In BLAZE, the YAML file is the "recipe" containing the pipeline's specifications. This includes models, datasets, processing components, and UI. While BLAZE has a block-based drag-and-drop UI for users to build their own YAML recipes, we'll use a pre-existing recipe in this case. 

This YAML file can be found in the yaml folder and is titled 05_search_summary_webex.yaml:

In this case, we see the input data specified, namely WebEx. Furthermore, we can see the specified two tasks, search and summarization, with models ElasticBERT and BART, respectively.

Interestingly, we see that the actionables task has not been specified. This is due to our usage of Dynamic APIs for prompt-based tasks. The inclusion of module: ['openai'] indicates to BLAZE that we have one or more tasks that require an OpenAI LLM. The user can specify the corresponding prompts for these downstream tasks, giving them greater flexibility over specific functionalities. 

We'll take a closer look at Dynamic APIs in our next blog post!

This YAML file will be used to create a solution-specific REST API server in the "Execute" stage.

Execute - Launching the Pipeline

Now, BLAZE can utilize the YAML file chosen at the preceding stage to establish a server, fetching and hosting the appropriate models, data, and components as specified with the press of a button.

However, before doing this, we must first utilize our Webex Access Token and Bot Access Token.

To pass these tokens into BLAZE, we can simply add them as environment variables, detailed below:

export WEBEX_BOT_TOKEN=<YOUR_WEBEX_BOT_TOKEN>
export WEBEX_ACCESS_TOKEN=<YOUR_WEBEX_ACCESS_TOKEN> 

echo $WEBEX_BOT_TOKEN
echo $WEBEX_ACCESS_TOKEN 

B) Windows Environment Variables

To add an environment variable in Windows, we can open PowerShell and run the following commands:

$Env:WEBEX_BOT_TOKEN = '<YOUR_WEBEX_BOT_TOKEN>'
$Env:WEBEX_ACCESS_TOKEN = '<YOUR_WEBEX_ACCESS_TOKEN>'

Alternatively, we can use any of the following commands in Command Prompt, depending on our intent. The set command will only persist while the current Command Prompt is open. The setx command will add environment variables in user (or machine) registry, depending on the /m flag.

set WEBEX_BOT_TOKEN=<YOUR_WEBEX_BOT_TOKEN>
set WEBEX_ACCESS_TOKEN=<YOUR_WEBEX_ACCESS_TOKEN>

setx WEBEX_BOT_TOKEN = <YOUR_WEBEX_BOT_TOKEN>
setx WEBEX_ACCESS_TOKEN = <YOUR_WEBEX_ACCESS_TOKEN>

setx WEBEX_BOT_TOKEN = <YOUR_WEBEX_BOT_TOKEN> /m
setx WEBEX_ACCESS_TOKEN = <YOUR_WEBEX_ACCESS_TOKEN> /m

For more information on setting environment variables, please check out this awesome blog post!

After adding our environment variables, we can simply run the following command:

bash run.sh server yaml=yaml/05_search_summary_webex.yaml

In a few seconds, we should see our pipeline up and running!

Finally, we can kick off our Webex Bot. A sample bot with the following endpoints has been created in the webex_UI folder. Feel free to easily add and extend the current functionalities (detailed in "Interact").

To launch the Webex Bot (our frontend), we can simply run any the following command:

bash run.sh bot

And voila, we have our Webex bot and NLP solution up and running. Let's take it for a spin!

Interact - Using the Pipeline

The current functionalities supported by our sample Webex bot include:

• list - list all meetings
• search <query> - search across all meeting transcripts
• actionable s <meeting_id> - identify all actionables (todos) from the specified meeting
• summarize [<meeting_id>, ...] or all - summarize one, multiple, or all meeting transcripts

However, these can easily be modified and extended with minimal code. Let's take a closer look!

Underneath the Hood

Our Webex bot calls a number of endpoints hosted by our REST API server.

For example, the /summary endpoint is ideal for quickly understanding what a meeting was about. Similarly, the /search endpoint efficiently searches across all transcripts using our specified model.

Within webex_UI/webex_bot/main.py, we can see a number of commands added to the Webex bot. These commands are the functionalities that the Webex bot will have. For example, we see SummarAcross(), SearchAcross(), and Actionables() below.

Within webex_UI/webex_bot/cmds.py, we can see the specific inputs, outputs, and descriptions of each of these commands. For example, this is the SummarAcross() command class.

Finally, within webex_UI/webex_bot/help.py, we see the endpoints being called, enabling our Webex bot to be lightweight and extensible. It is very straightforward to add or modify a command!

Pretty neat! Now, let's give our Webex Bot a try.

Bringing it all Together

First, we can open our Webex chat interface and search for our bot using its username. 

Next, we can type 'help' to initiate a dialogue with the bot, getting a list of its supported commands.

From here, feel free to explore! Please bear in mind that the quality of the search and summarization depends solely on the models selected. Playing around with the specified models in the yaml file will lead to improvements in performance if desired. 

Here are some sample interactions with our Webex Bot:

First, we can say help to see our available commands.

Next, we can try out the list feature to quickly see all of our meetings.

Nice! Now, let's try out the Actionables feature on one of them.

Pretty cool! What about the summarize feature?

Ahh, so that's what that meeting was about! Finally, let's try a search.

Coming to a Close

In this blog post, we created a Webex bot capable of summarizing and semantically searching across meeting transcripts using BLAZE! In a few easy steps, we learned how to build a YAML file, execute our REST API server hosting our NLP pipeline, and interact with our solution through Webex.

Thank you so much for taking BLAZE for a spin!

We're always grateful for community involvement 😄. If you're interested, feel free to contribute on GitHub by starring, forking, or sharing the BLAZE repository 🥳! Thank you for your support!

The Road to Tomorrow

Here is a roadmap of our past, present, and future blog posts with BLAZE.

Next time, we'll harness BLAZE to build a live meeting assistant for Webex! Stay tuned! 🥳

• Post #1: Introducing BLAZE — Your Composable, Flexible, NLP Pipeline Solution
• Post #2: Building with BLAZE — Creating LLM-powered Webex Bots for Interacting with Meetings
• Post #3: Building with BLAZE — Creating LLM-powered Meeting Assistants for Webex (coming soon!)

Getting Started with BLAZE

To learn more about BLAZE, please take a look at our GitHub repository and documentation.  

• (Homepage) https://research.cisco.com/research-projects/blaze  
• (GitHub Repo) https://github.com/cisco-open/Blaze  
• (Docs/Usage) https://cisco-open.github.io/BLAZE/

If you have any questions, want to share your thoughts, or wish to contribute, feel free to reach out to our Group Mailing List (blaze-github-owners@cisco.com) — we’d love to hear from you! 

Acknowledgments

A lot of work went into building BLAZE — our team started this project last summer, and we have been meticulously working to ensure a seamless experience for users wishing to harness the promise of NLP.  

In addition to those acknowledged in our first blog post, we wanted to express our special thanks to:

• The Webex team, for their fantastic meeting and chat interface, helpful documentation, and impressive developer tools. BLAZE's integrations with Webex would not be possible without their curated open-source libraries, easy-to-follow demos, and meticulous blog posts 😄.