Hello
I have been working on a Retrieval-Augmented Generation (RAG) workflow using Vectara’s semantic search APIs, and I’m curious to learn the best practices from the community.
My current pipeline integrates Vectara embeddings with prompts sent to an LLM, aiming to produce accurate and context-rich answers. 
Since I’m still refining the process, I would like to discuss common challenges and solutions others may have tried.
The key issues I’ve noticed so far are around how to structure prompt templates, how best to chunk large documents for indexing, and how to minimize hallucinations from the LLM. 
Checked Retrieval Augmented Generation: Everything you need to know guide for reference.
While Vectara’s retrieval quality is strong, combining it effectively with what is Generative AI for consistent outputs is still a challenge. I think there’s a lot of value in sharing real-world approaches here.
Has anyone explored specific techniques for balancing retrieved context with generation? I’d also be interested in evaluation strategies for retrieval quality before handing results over to the LLM.
Any practical tips, frameworks, or workflow examples would be greatly appreciated.
Thank you !!