I’m exploring ways to optimize prompt efficiency with GPT-4, specifically aiming to reduce the number of tokens required. One approach we’re considering is building a knowledge graph of our product specifications, including features and other details, and leveraging it during prompting to validate information against this base data.
Our main objectives are:
To build a comprehensive knowledge graph that encompasses all product specifications and details.
To integrate this knowledge graph with GPT-4 prompt in a way that allows for efficient and accurate validation of information with minimal token usage.
Any insights, techniques, or examples of similar implementations would be greatly appreciated!
It’s called Retrieval Augmented Generation (or RAG) - while most people associate that with Embedding retrievals, you can use pretty much any search and retrieval method (e.g. SQL, Cypher, google) to enhance your output
This sounds interesting. In the same sort of way as representing all the rows of data in a data table as sentences is inefficient, there are some kinds of information which are more concisely and clearly represented in data structures.
Can you give a small example of a Knowledge Graph representation as text in what you’re planning? I’m interested in this as well, and happy to discuss.
