Extracting core concepts with ChatGPT 4 from OCR of scanned sample examination PDF - Feb 2024

I think this is an interesting example of what works and doesn't work on the personal AI front in early 2024.

My son was given a printed practice exam in microeconomics. I wanted ChatGPT 4 to extract and summarize the core concepts. This turned out to require two steps, one of which only worked with Google.

Step One: OCR and download text file

I scanned the document in ScanSnap and produced a scan PDF. I tried getting ChatGPT to do the OCR but it abandoned that task. I then tried Gemini and it told me it didn't do OCR. Next I tried Microsoft Lens, but it seemed to only do OCR from a local image, I couldn't see how to use it with a OneDrive PDF. ChatGPT claimed that I could open a OneDrive PDF in Office 365 Word but that did not work with the web version (perhaps it works with full Word?). ChatGPT did not know of a way to do PDF OCR on Sonoma.

The only thing that worked was Google Drive. It allowed me to open the PDF in Google Docs and then export a .txt version.

Step Two: ChatGPT 4 analysis

I asked ChatGPT 4 to extract the key concepts from the .txt file. It provided a plausible set and then proceeded to answer some of the exam questions. Concepts captured were:

... equilibrium price, consumer's surplus, producer's surplus, total surplus, efficient output levels, negative externalities, deadweight loss (DWL), price ceilings, and the impact of taxes on market outcomes ...

I don't think it added much to the textbook chapter topic loss but it did provide a plausible set of topics to emphasize in my son's studying. I was primarily interested in the workflow today. It will be interesting to look back on this in a year and see what's different.

Rendering ChatGPT output in readable form in a Juypter Notebook

Update: This post is still useful, but there's also a way to enable line wrap in Visual Studio Code's Jupyter extension. You can also use the Python Print function instead of the Display example in my original post. For example:

output = client.completions.create(

    model="gpt-3.5-turbo-instruct",

    prompt="List the days of the week: ",

    max_tokens=100,

    stop = "Saturday", #put this in for fun

)

print(output.choices[0].text)

Note these print parameters are specific to this particular object's structure. (I think JSON but I'm a newbie.)

Original below

--------

This was a bit of a revelation. I don't know Python but I've been working through a ChatGPT / LLM tutorial using Visual Studio Code and a Juypter Notebook on macOS. In a Jupyter cell the output renders below the cell and it looks like this:

Completion(id='cmpl-....', choices=[CompletionChoice(finish_reason='stop', index=0, logprobs=None, text='1...

All in one unreadable line with \n as a paragraph deliver and no line wrap.

I asked ChatGPT 4 to help. Over a series of interactions I tried different things and got various error messages I passed to ChatGPT 4. In turn it analyzed my error message and suggested fixes.

This is what I ended up with in about 15 minutes, here added to a cell that ran a simple prompt query

from IPython.display import display, HTML

output = client.completions.create(

    model="gpt-3.5-turbo-instruct",

    prompt="write me a poem",

    max_tokens=100,

    n=3

)

text_content = output.choices[0].text if output.choices else ""

html_output = text_content.replace('\n', '<br>')

display(HTML(html_output))

This is what the output looks like now (the poetry is greeting card quality and mildly painful):

A poem, a weave of words and rhyme

A tapestry of thoughts and time

A magic spell from the poet's pen

A story of love, of loss, of when

The stars above, they guide my hand

As I write of distant lands

Of fiery sunsets and ocean tides

Of moments we hold and let slip by ...   

[Adolescent poetry truncated] 

This screenshot shows it best ...

We are in a new world.