Creating Summaries of Podcast Segments

Podcast transcripts are retrieved! 

(If you haven't seen yesterday's post, you can review it [here](https://smoothing-learning-curve.beehiiv.com/p/saving-3-hours-chatgpt))

Today, we will summarize podcast transcripts using the Langchain library.

This is where we're currently in the process:

What is Langchain?

Langchain is a practical tool designed to automate interactions with large language models using Python. It simplifies the task of integrating AI functionalities into applications, offering a way to utilize language models.

Capabilities of Langchain:

• Prompt Management: Manage and adjust prompts.

• Chaining: Set up sequences of questions and answers for more detailed queries.

• Integration with Agents: Connect with other platforms, such as Google search, to gather and process data.

• Memory in Conversations: The AI can remember and refer back to previous interactions during a conversation.

• Summarization: Efficiently condense large amounts of text or data into concise summaries.

Today, we'll use the last capability from the list - text summarization.

Why summarize?

• Structuring information. Podcast transcripts can be messy so this is a way of cleaning and standardizing information.

• Summaries will be used for semantic search. We'd like each element in the database to contain roughly the same amount of information. We want to avoid bias in our search results.

Creating summaries in Langchain

To create a summary in Langchain, first, we need to make sure we have OPENAI_API_KEY set up:

• get OpenAI API Key (if you don't know how, watch this video

• add it to your .env file (OPENAI_API_KEY variable)

Then, we get to the summarization part. We will use load_summarize_chain with stuff chain. That means that each transcript will be summarized once.

We have over 5000 files to process!

That's a lot.

This is why we need to do this efficiently. Langchain supports asynchronous calls to OpenAI API with chain.arun method.

This is exactly what we need!

There's one quirk, though.

We cannot exceed the maximal token rate per minute. For the model, will be using (gpt-3.5-turbo-16k), which is 180,000 TPM (tokens per minute).

This is why we will batch our async calls (batch size = 30 transcripts) to avoid exceeding the limit.

The code that achieves that is available here.

And this is the summarization prompt, you might find useful:

You are a podcast summarization expert. Your outputs are easy to read-through, concise and actionable.


As an input you will receive a transcript of the podcast section.

 
Your task is to create concise and in-depth summary based on the transcript that you will receive.

Compress as little information as possible.

If there is any information that should be added, please do so.


Your answer will be presented to an audience interested in science, technology, artificial intelligence and self-improvement.


Here is the transcript:
  

{text}


SUMMARY IN BULLET POINTS (UP TO 15 POINTS):

Btw. The asynchronous code was generated based on my non-asynchronous implementation, documentation of Langchain's Async API, and ChatGPT:

I just gave ChatGPT the documentation and asked it to make adjustments!

Results

The results look pretty good!

However, it already cost me over $20, and I still have 1500 segments to process:

This is why I don't advise you to run the code.

Tomorrow I'll share the whole summary database for FREE!

Stay tuned, it will be on my Twitter!

This is the eighth day of the 30-day AI challenge.

Over the next month, I will be building the Lex Fridman AI engine with you!

If you're reading this, I assume you'd like to build things. If you stick to this newsletter you will have a running project after a month and know the necessary technology to build AI apps.

I've recently built PodcastGPT and want to share the process with the community. If you haven't seen the app yet, you can get access here: PodcastGPT

This is all for now! See you tomorrow.

Stay focused!

Luke