Summary
We have already created a solution within Bubble.io that is designed to provide a chatbot/agent to answer questions from large PDF based research reports. The solution unfortunately has a number of limitations which do not meet our clients requirements and we are seeking a developer with a combination of code base and AI experience to create the functionality we require or coach our team to complete.
Background and Context
The current solution is built in bubble.io. At present the solution retrieves a URL with a unique key to connect to a single large PDF file. We then vectorize and chat with that PDF via API using ChatPDF (an external AI service). We also display the PDF next to the chat, with the help of the “PDF Viewer” plugin, to allow the user to navigate to any section. Although this allows us to chat with the document it has two major limitations we need to overcome:
- It does not return all the queries correctly, and often refers to other points of data - which is a limitation of ChatPDF. As we have no control over how it chunks the file when it is vectorized, or can we provide any additional prompt training to reduce errors
- We need the responses to reference the correct pages within the document either by page number or in the ideal world show that section on the rendered PDF
Project Goals and Objectives
The goal of this project is to use the existing Bubble.io framework but to develop the additional functionality as outlined above. We are open to different approaches including using OpenAI agents/GPTS and/or API based plugins (it should be noted that we have explored a wide variety of different plugins including Chatbase, ChatPDF & Chatdoc, but they do not provide the flexibility we need.
Scope of Work
- Can be embedded within the existing bubble.io solution
- Be able to maintain client brand and tone
- Be capable of vectorizing a PDF on the fly and keep the load time under 2 minutes
- Have some control over the vectorization and management of chunks so that the queries are returning the right fidelity of answers
- Be able to provide some prompt training or coaching such as “You are a research analyst….”
- Be capable of supporting multiple concurrent users (ideally up to 10)
- Provide chat capabilities to query the document
- Query returns should show the source information (page number as a minimum)
- Nice to have - Reference location is highlighted on rendered PDF (i.e. the PDF viewer will be scrapped and a new/interactive PDF viewer will be utilized).
- Nice to have - Ability to choose the underlying model eg. ChatGPT 3.5 or ChatGPT 4