Need Help Implementing Supervisor

[jonaslouras]

Hey everyone,
I'm working on a Supervisor with LangGraph for a company internship. My mentor has asked me to create three Agents: "Question Agent", "Answer Agent", and "Summarizer Agent".
The input is a PDF, which I need to split by page and add each page to a vectorial database for later use. Each agent will also save its outputs in the vectorial POSTGRES database. Here's a rough idea of the structure:
Questions Table
id (Primary Key)
question (Text)
embedding (Vector)
document_id (Integer)
page_number (Integer)
Answers Table
id (Primary Key)
answer (Text)
embedding (Vector)
document_id (Integer)
page_number (Integer)
question_id (Foreign Key to Questions table)
Summaries Table
id (Primary Key)
summary (Text)
embedding (Vector)
document_id (Integer)
page_number (Integer)
Documents Table
id (Primary Key)
summary (Text)
document_id (Integer)
number_of_pages (Integer)

The workflow is something like this:

• Load the document (sanitize the text, embed it, save in "Documents")
• Make a summary of each page (save in "Summaries")
• Generate questions for each page (save in "Questions")
• Answer all the questions generated by the Question Agent, considering the context of the page (save in "Answers")

My biggest question is: what tools and agents should I implement for this? Most resources I've found online use tools like Tavily Search and Python REPL, which aren't really helpful for my case. I need to use the Supervisor since it's a project requirement, and I'm a bit confused about the implementation details, since this would be very easy to implement with simple chains, and the only solution I could come up with is tooless agents...?
Any advice or pointers would be greatly appreciated! Thanks!

[hwchase17]

"tooless agents" isnt a terrible idea here! one thought exercise i always like to do is how would i do this, as a human. I think for this task I could probably complete just with with (a) the document in question, (b) my logic - i don't think I'd rely on a search engine, calculator, etc. A lot of it will probably come down to the "flow" you set up between agnets, and the prompts you use to instruct the agent how to behave