Following on the previous issue we now need to generate a nice mermaid chart to supplement the output summary to enhance the user's understanding of the summary.

We are building the chart in two steps:

1. Generating it.
2. Displaying it.

The problem is in generating it, check the generate_chart(transcript), it calls ChatGPT to generate it. The problem is that the result is not always accurate (tested on GPT-4) sometime it is a correct chart, sometimes it is not.
To test the chart, run the output on Mermaid's live editor.

Observation: GPT seems to work better with very short videos.

What needs to be done:

1. Prompting: Play with the prompt a bit to see if you can reach consistent working charts.
2. Different models: Test against different LLM models.
3. Finetuning: Can you get hands on a dataset that consists of blocks of texts and their equivalent mermaid charts? Then either provide us with this dataset or finetune any LLM model and share your code if the results are better than GPT.

The chart is then displayed via st_mermaid(mermaid_code), no issues with that.

User Experience:

As the user is watching the video, they can chat with AiDA about it. Their questions, might require an in depth response aided by a nice chart or it might be a simple response. Hence, a chart is not always necessary in this case.

What needs to be done:

To mitigate this redundancy, a nice way is to classify the user's intent, as a start if they are looking for a summary then yes generate a chart and audio file (text-to-speech) as well.

For that check function classify_intent(user_input) it attempts to classify whether the user is requesting a summary or something else via a ChatGPT call. The responses are inconsistent and not always accurate.

To address this, the following steps need to be implemented:

1. Prompt: We need to make sure ChatGPT response is precisely "True" or "False" to treat the output as a boolean with ast.literal_eval(response.content). It might be that GPT is responding with something like "true" or "true.." or "yes, it is a summary" or the like. Keep playing with the prompt to reach consistent output.

2. Function Calling: Use a json schema and function calling to return the response to a specific json key, read this.

   Bonus Points: Figure out Function Calling in LangChain!

3. Testing Robustness: Need to be tested sufficiently to confirm that the intent is correctly classified as "True" if request relates to a summary or "False" for anything else.