I keep seeing people fine-tune LLMs for use cases where they probably don’t need to. In most doc/product scenarios, you don’t need another fine-tuned model—you just need retrieval-augmented generation (RAG).
Why I think RAG wins in most cases:
Fine-tuning is expensive, slow, and brittle.
Most use cases don’t require “teaching” the model, just giving it the right context.
With RAG, you keep things fresh: update your docs → update your embeddings → done.
I built a small proof-of-concept to test this: a documentation assistant where docs are chunked + embedded, user queries are matched with cosine similarity, and GPT answers with the relevant context injected. Every query is logged, which turns out to be valuable—surfacing missing docs, common user struggles, and even feature requests.
Demo: https://intlayer.org/doc/chat
Write-up + code: https://intlayer.org/blog/rag-powered-documentation-assistan...
My question:
Do you see fine-tuning + RAG coexisting for these types of tasks?
Or is RAG simply the obvious solution for 80% of real-world doc/product use cases?
I keep seeing people fine-tune LLMs for use cases where they probably don’t need to. In most doc/product scenarios, you don’t need another fine-tuned model—you just need retrieval-augmented generation (RAG). Why I think RAG wins in most cases: Fine-tuning is expensive, slow, and brittle. Most use cases don’t require “teaching” the model, just giving it the right context. With RAG, you keep things fresh: update your docs → update your embeddings → done. I built a small proof-of-concept to test this: a documentation assistant where docs are chunked + embedded, user queries are matched with cosine similarity, and GPT answers with the relevant context injected. Every query is logged, which turns out to be valuable—surfacing missing docs, common user struggles, and even feature requests. Demo: https://intlayer.org/doc/chat Write-up + code: https://intlayer.org/blog/rag-powered-documentation-assistan... My question: Do you see fine-tuning + RAG coexisting for these types of tasks? Or is RAG simply the obvious solution for 80% of real-world doc/product use cases?