Hello I am working on building a strong customer service chatbot for a ventilation company. I have experimented with Pinecone Vector Stores and Supabase. Both work great, however they lack in technical areas and being able to provide very accurate answers. I believe it comes down to the chunking of the documents. I have compiled a document that outlines all the necessary area such as installation procedure, compatible components, technical information etc but when It goes into the text splitter, it gets split into chunks based off characters and not headings so a lot of the sections overlap with each other which I think leads to inaccurate answers. Any recommendation for a workflow with complex knowledge? Has anyone done something similar?
1 Like
Hi @tainoooo
It might help if you store information in a tabular format instead. Here’s a RAG agent that has a tool to query the database to find information. Doing it this way returns a much smaller piece of data for the llm as rows and columns.
if this post helps solve your issue, please mark it as the solution so others can benefit from it
Yea this is exactly how my agent looks. Still struggles sometimes with getting good answer. Would this just be the knowledge base I have provided.
1 Like
What type of data are you putting into supabase? Doc files, pdfs? Or data more like excel.
You could also spit the agent into multiple. Instead of one big agent, have one agent that determines which agent would be the best at answering the question, and funnel it down to the smaller agents.
Oh yea that’s a good idea. I am just using google docs, uploading the file through google drive.
1 Like
Its going to work much better if you store the data in csv/sheets format. Then the “Query Document Rows” can write a query to return the relevent data.
If this helped you, click the heart to show it’s useful to others.
1 Like