New Demo: Q&A with your PDF documents 🎬

Questioning and Answering

Today, I would like to share a demo video I made showcasing how Q, ChatGPT for Slack, can help reference PDF documents for your company's products and operations, using NVIDIA's latest GPU manual as an example.

Challenges in the background

Imagine the various manuals and documents your company uses - for products, services, or even internal processes.

Often, these essential materials remain underutilized. Employees may struggle to locate them, or find them too complex to navigate. Consequently, a few individuals with extensive knowledge become the default go-to for questions, slowing down their productivity and the overall workflow.

Or consider remote teams and new hires who might hesitate to ask for help, resulting in them either over-researching or not researching at all. Again, productivity suffers. But by harnessing Q, ChatGPT for Slack, to answer these questions with your PDF manuals, we might be able to help address these problems.

Step by Step: Here's How It Works

Our recent demo showed how 'Q, ChatGPT for Slack' can help alleviate these issues, by directly accessing the content of your PDF manuals to answer queries.

Let's dive in step by step! 🚀

Step 1: Upload the PDF

You can upload a PDF simply by sending it in a direct message tab with Q. The time it takes to process the PDF will depend on its size, but for our NVIDIA GPU manual of 160 pages as an example, it was almost instantaneous. Additionally, you can upload two or more PDFs at the same time.

Step 2: Activate Your Plugin

Go to the Home tab, where you'll find a section for plugins. Here, you can select the plugins you want to use. For our purpose today, turn on the "Pre-Uploaded Document Search". This will enable Q to search through your uploaded documents such as PDFs, URLs, and even GitHub repositories. Today, we'll focus on PDFs.

Now, you're all set to begin asking questions!

Step 3: Begin Questioning and Answering

In the demo, I asked the following questions:

Question 1: "Do you know NVIDIA GeForce RTX 4090?"

ChatGPT's answer? Of course it was unable to provide an answer.

Q&A about your company's products or operations will most likely not be answered in principle. Also, documents available on the web but published after September 2021 are not covered as well.

For this demonstration, we had no choice but to use PDFs published on the web, so we uploaded the latest manuals for NVIDIA's GPUs released after September 2021. Then we fed it to Q, and voilà! Q responded with details such as what accessories come with the GPU and even provided the file name and page number for reference.

Question 2: "Is there anything required?"

Again, Q was able to provide a response, such as the type of motherboard slot required and the need to use the included adapter. There were, however, some missed details like the power rating and OS requirements. But this could be due to the vague nature of the question. This opens up opportunities for improvement and hacks in PDF search capabilities.

Question 3: "How do I install it?"

While Q provided a detailed installation process, it did miss out on some specifics mentioned in the manual. The point here is not perfection but a significant step towards the goal.

Key Takeaways: How to Make This Works for You

Although we're not yet at a stage where the AI can perfectly answer all manual-based queries, it's definitely progressing in that direction. And there are two key areas where you, as a user, can influence the AI's performance: the way you ask questions and the quality of the PDF document you're using.

Refine the way you ask questions

You should start with a short, broad or vague question as it's easier for you. But if you're not getting the answers you want, it may be worth trying the following:

• Rephrase your questions
• Be specific about what you want to know.
• Replace pronouns with specific nouns
• Make sure your sentences have a clear subject, predicate, and object
• Or even switch the thread of conversation in Slack

The last point, in particular, is more straightforward. Q uses the entire conversation thread to perform vector similarity searches on PDFs that are split into smaller pieces. Therefore, by switching threads, you can refresh the search.

As for the other three points, they are also important basic techniques that are useful when talking to AI in the first place.

Refine the quality of your PDF documents

In terms of document quality, if you don't mind re-creating the PDFs, it may be worth trying the following:

• Remove unnecessary line breaks
• Label important sections explicitly
• And include the chapter name on each page

The current OpenAI GPT-3.5 and GPT-4 have a limit on the number of tokens, which limits the amount of text the AI can read or write at one time. Therefore, in the case of long text, such as a PDF manual, it is split into chunks small enough to be referenced. Since the main separator of that division is line breaks in PDF, it is effective to exclude unnecessary line breaks.

It is also effective to clearly state that notes are notes; what is unreadable to the AI is anything that is emphasized by visual effects such as red text, bold text, red borders, etc., and that information is lost to the AI. Strictly speaking, the contents of a PDF file are stored in a binary format, not as text, and such visual information is lost when it's read.

Also, for cross-page sections, it is better to have a string for that section on each page, since PDF, while visually readable for humans, is full of unnecessary line breaks for AI, and sections are not as clear as markdown, for example.

Wrapping Up

If this all sounds exciting, I encourage you to try Q.

Furthermore, if you use Slack Connect to invite info@suchica.com, you can also chat with our customer support team on Slack. Slack Connect is completely free for direct messages. Some of our customers are already using this service instead of email.

As with any new tool, there may be some initial challenges and learning curves, but once you get the hang of it, you'll see how Q can revolutionize the way you work. No more hunting for information in long, complex documents - simply ask Q, and you'll get your answer. It's that straightforward.

If there's an instance where Q was unable to successfully answer questions regarding your PDF documents, please let us know. If you can provide us with the PDF and the question in question, we can examine and improve it in greater detail!