LumenData’s Director of Analytics, Andrew Crider, highlights how the focus on chatbots in Generative AI is misplaced and recommends leveraging large language models to build tools to solve real-world problems.
I recently returned from an Enterprise Technology conference, and I can only say I was flabbergasted about the number of Natural Language Chat features on the floor. Since this was a Data Conference, you can only imagine what the options were:
- How many Customers do I have that have a positive return on my investment?
- Please show me a bar graph of how much my revenue was over the past year.
- What is in this 10-K filing? How much revenue do they expect to make?
Chatbots have been the go-to example in GenAI for the past few months. However, the experience of using a ChatBot to retrieve information could be better, as anyone who has used Siri, Alexa, or Google Assistant can tell you. The results can be sporadically accurate and precise, which erodes the user’s trust and faith in the bot. This results in chatbots being used for timers and to play music. The models have gotten much better about providing information, but have they gotten more accurate? Recent outputs from Google’s attempt to add information via LLM would say no.
Original Image: https://www.threads.net/@petergyang/post/C7S6fzINqZj by Gyang Peter
The funny thing was that I was watching some YouTube Shorts this week, and I saw a “Secrets of Advertising” video in which the videographers used glue to make cheese-less tacky and more film-worthy. Google is no doubt indexing and training models with YouTube data, and since I have seen at least a million versions of this video, this probably indicates that they need to clean up their data.
So, if we stop using chatbots as information sources, how can we use LLM models to improve our lives? The LLM developer should focus on reducing the cost of gathering verifiable data and summarizing small data packets to present them to the user more naturally.
I want to present three tools I use to do just that.
Image Classifications
Images available to author via unique access to Midjourney; the author assumes responsibility for the authenticity.
I frequently use Midjourney to create images. These images fill my Digital Art Frame, are used in Ambient Videos for YouTube, and bring character and life to my blog posts and PowerPoint decks. I don’t have a way to classify them, but I can use GenAI to extract information from over 1,000 images I have created in minutes.
Tools: Python, Google Gemini, llama_index
Within that code, I utilize a pydantic class. These classes allow me to dictate what the LLM is looking for and what format I want the response to be in.
Results
Images available to author via unique access to Midjourney; the author assumes responsibility for the authenticity.
The Code to the right is from the Python Code in the above section.
Images available to author via unique access to Midjourney; the author assumes responsibility for the authenticity.
The Code to the right is from the Python Code in the above section.
After running this code, I get information on the images, including color, style, and content. With this, I can create a vector search and include it in an AI assistant’s toolkit, create a function that changes the images on my Digital Art Frame based on a mood or keyword, or use it as a basis for a simple search feature.
Midjourney’s Consistent Character opens up a Wide Range of Possibilities.
Creating consistent characters became possible with Midjourney’s new feature.
Summarize Your Podcasts
Images available to author via unique access to Midjourney; the author assumes responsibility for the authenticity.
I recently published an article on how to build your own podcast to kickstart your morning. The content I used for this podcast includes weather reports, daily news, my calendar, and my tasks stored in Todoist.
Create a Personal Podcast for less than a Latte a Month
While everyone is busy creating chatbots LLMs and NLP, create a daily audio summary for your day!
I subscribe to many podcasts but only listen to specific feeds if the topic interests me. This code summarizes my favorite podcasts every morning to let me know if I should listen to the episode:
Tools: Python, Google Gemini, feedparser
The great thing is that this code can be adapted to anything with an RSS feed! GenAI is excellent for summarizing content, and this processing can help you efficiently drill down into all of your media consumption to get the information that is important to you.
Objection Handler
At the conference I mentioned above, I presented for LumenData (LumenData inc) on how to efficiently use Informatica Master Data Management and Snowflake to prepare data for GenAI. You can find out more about that here. The result of the demo was an LLM call to Snowflake’s Arctic LLM that resulted in a series of prompts for Customer Services Reps to overcome objections to items that weren’t purchased.
Tools: Snowflake SQL, Snowflake Arctic
With a little bit of data cleaning, we were able to demonstrate how we could get the specific attributes of a customer (age, gender, etc.) and use that to build customized scripts on how to overcome objections:
Image is from a StreamLit Dashboard after running the above SQL Code.
To make this more effective, you can include RAG pipelines to further influence the LLM’s generation, i.e., proven sales techniques, detailed information on the objects, feedback from other users, etc.
Chatbots are not the only utilization of GenAI!
Chatbots can be impressive delivery mechanisms to the end user. However, we need to focus more on developing the underlying tools that power them. By spending extra time developing tools that can be called via ChatBots, we can provide focused and valuable GenAI applications. Systems like LangChain and LlamaIndex are great places to start creating these tools. Go out and solve multiple problems. Once you empower the chatbots, then you can show them off.
Author
Andrew Crider
Andrew Crider is the Director of Analytics at LumenData. With over ten years of experience in the analytics space, he has helped multiple Fortune 500 companies get the most value out of their data innovations.
