LogDetect.AI

Design Document

Description:

Remember all the times you missed a meeting and had to ask multiple people just to know what was decided?

What if you had a dedicated assistant who you could ask anything about the meetings you missed, and it always gave you up-to-date and accurate answers?

LogDetect.AI is exactly that.

How it works is simple:

• You ask a question
• You provide a set of call logs (ordered by date)
• LogDetect parses the entire set of call logs and presents a curated set of facts that exactly answer your question.

Features:

• Ask any question
• Upload multiple call logs
• Re-order call logs based on your preference
• Generate facts that answer your question!
• Facts are generated by the OpenAI API and the GPT-4 model

Prompt Strategy:

• A Prompt Chaining Strategy was deployed
• For the first call log, all relevant facts to the question were extracted
• For all subsequent prompts:
  • The question was provided to maintain context
  • The output of the previous prompt was appended as known facts
  • The call log was provided
  • The model was asked to update the fact list based on the question and provided context
• The prompt strategy is better depicted below:

Design Decisions:

• Initially I experimented with providing all call-logs to the model in a single prompt. I decided against this solution for the following reasons:

  • Difficulty of maintaining context when call logs are either many in number or large
  • If the number of call logs is very large, it might breach the input token limit of gpt-4

• I decided to then use a prompt chaining technique as described above. I got decent results. Here are some of the tradeoffs I made in the design:

  • No. of API calls = No. of call logs. This means it is a more costly solution. However, I decided to optimize for accuracy and correctness, over costs for this implementation.
  • Priority on maintaining context. The method I employed does well in maintaining context across call logs, whereas in the case where all call logs are provided in a single prompt, the output quality will tend to decrease with an increase in the no. of call logs

Test Example:

Question:

What features should be included in our new software update?

Call Log 1:

1

00:01:05,300 --> 00:01:33,900

Alex: Good morning, team. To kick things off, I believe our new software update should definitely include enhanced security features. We've had feedback about vulnerabilities that need addressing.

2

00:01:34,000 --> 00:01:42,500

Tina: I agree with Alex on enhancing security. I also think we should integrate AI to personalize user experiences. It could analyze user data to provide customized suggestions.

3

00:01:43,600 --> 00:01:45,100

Raj: Both points sound crucial. Let's add these to our update. Additionally, I suggest including support for multiple languages to expand our market reach.

Call Log 2:

1

00:00:50,000 --> 00:01:20,000

Alex: Reflecting further, I think adding a dark mode is essential. Many users prefer it, especially for late-night use.

2

00:01:21,000 --> 00:01:30,000

Tina: Dark mode sounds like a great addition for user comfort. Let's ensure it's fully adjustable to suit various preferences.

3

00:01:31,000 --> 00:01:32,500

Raj: Agreed, dark mode will be part of this update.

Call Log 3:

1

00:00:55,000 --> 00:01:25,000

Alex: On re-evaluation, should we hold off on integrating AI? It might rush things, and we risk implementing it poorly.

2

00:01:26,000 --> 00:01:35,000

Tina: I think you're right, Alex. Let's focus on perfecting what's essential for now and revisit AI integration later.

3

00:01:36,000 --> 00:01:37,500

Raj: That sounds like a prudent decision. Let's proceed without AI for now.

Output:

Validations:

• You cannot submit without a question and at least one call log.
• The URL you specify must point to a .txt file
• You cannot add duplicate URLs
• You cannot submit unless all URL contents are accessible by the application
• You cannot submit unless you have saved the order in which you want the call logs processed

Future Scope:

The Future scope of this project is immense. Here are a few applications that this technique could apply to:

• Directly convert audio to log files for a multimodal functionality
• Real time transcript to fact generation
• Automatic Minutes of the Meeting Generator