I’m researching the idea of connecting an LLM to ShotGird Database that uses RAG system to add precise answers to the user questions about site data. Since the topic is hot, there are a lot of ideas and starting points how to start the project, but since everything is new no workflow, technology, app-structure and also the documentations are standardized.
I started with Langchain, by first getting to load all the ShotGrid data, downloaded from a project (one csv per entity) and querying the data. The next step is to query the api, but with the relevant fields, that needs to the answer. The problem comes when you have a question like: Which Shot had the Task, that took the longest time to finish. The answer needs two tables at least with all of its data, get the Task to each one, and it’s queried time log data field. This could create a huge json query, where the json answer schema is not standarised, since the question determines, which connections should be there between the fields, for it to be relevant information for the answer. A.k.a the chunking is not happening by the API query.
The better option is to create an offline representation of the whole ShotGrid site, an offline databse, that can be chunked, and vectorized to that the RAG system can get relevant information from the preprocessed data.
This offline ShotGrid database refreshing seems to me like a whole mess, but relying on just the api query, and fine tuning that, seems messy too me as well.
My question would be:
Has anyone had a working solution for connecting the ShotGrid data to an LLm, and build a RAG system out of it.
I’m curious for your insights and your solutions! :))