Closed robknapen closed 4 months ago
I have listed a few in the README.md file for this component. But (as discussed today) there are still more options that can be considered. It goes a bit hand-in-hand with what purpose do we see for the NLQ component, i.e. what type of soil knowledge related questions do we want it to be able to answer? (What is the information / knowledge retrieval that needs to be done?)
Following the (Brugge) architecture, the LLM component only interacts with the triple store, using SPARQL queries. So at first it can work with text and relations it can retrieve from there. After the user selects one or more documents these can be retrieved and used as input for RAG. Importing documents into a vector store can take a few minutes (depending on size of the documents, embedding model used, and hardware configuration). After that the user can 'talk' with these documents to have more detailed access to the knowledge contained within them.
To avoid spending a lot of time perfecting the generation of SPARQL queries by the LLM, we can use known fixed queries and have the LLM use them via "tool calling".
(self) answered.
What different types of NLQ implementations exist?