I think this might be the issue:
> I think it is outputting the exact chunk in the embedding vs using the knowledge vector database to come up with and output the best answer.
The vector database will only output the embedding data and any metadata you store with it (likely in your case the transcript) FOR THAT CHUNK. If you need a natural language text output based on those results, then you will need to push that through an LLM. I think if I'm understanding everything correctly, you've done the retrieval step of RAG and now you need to use those results to do the generation part. See this diagram:
https://www.pinecone.io/learn/retrieval-augmented-generation/#What-is-Retrieval-Augmented-Generation