The problem SUPPLAI solves

THE PROBLEM STATEMENT

Traditional rags are very token expensive and as the context and files increases the LLM model eventually collapses the context and starts hallucination. So, it is context expensive can take upto 4000-7000 token in a single prompt and it can also hallucinate. Also all the files are chuncked so model have to parse all the files again and again for every prompt given to it.

THE SOLUTION

1. Build a knowledge system with nodes where each node represent a data point connected to other nodes which are related data points connect by weighted graph data structure. Now whenever any prompt is given the model first create isolated nodes then started connecting nodes and then increases the weights, So the knowledge graph for each user is unique and each user gets there specific response, the model on parses the important nodes.

WHAT IT SOLVES, SCALABILITY AND APPLICATION

1. Traditional RAG systems are not token effiecient it is 70%-80% more token efficient than traditional RAGs
2. It can be implemented to any traditional RAG application which halluciates due to huge context, in the architecture the model knows just two thing from which node I came from and on which node I am on, and decides which node to go next based on the weights of the graph so that the user query can be solved with minimun tokens and appropriate answer.

WHAT WE HAVE BUILT

1. An AI based B2B procurement system, where you can do anything in the website from searching seller, to product to adding order to cart to placing order, the only thing user need to do is payment.

2. A user based customized knowledge graph, where each seller, product, category and chat thread is a node and based on this node types and parameters such as number of times a product is queried, number of time the same product brought and all the model decides which node to jump and increases the weight of the path while jumping which make the path more important, So next time when the model is again on the same node it knows which node to go next.

Challenges we ran into

Initially in the knowledge graph we used undirected graph, which would have been a bottle neck as the application scaled becase the model will have to travel the whole graph in worst case, but after taking suggestions from a mentor we have implemented a directed graph now the problem of attention and decision making is solved, this transition from undirected to directed graph was biggest challenge to implement.