Ask HN: Does piping LLM output into a RAG stack sound like a good idea?
I would like to know if this can improve agent memory or adherence to coding guidelines and policies in large codebases over time.
I'm also interested in this as a general concept.
Comments
Could you explain a little more? Are you trying to fix the problem of the LLM ignoring instructions?
Not exactly agents ignoring instructions, but I'm wondering if (selective) inclusion of LLM responses in a RAG stack might be suitable as a sort of long-term memory for "accepted" LLM contributions to code for example. This way, unwanted novel or alternate solutions to repeated patterns might be avoided?