Generated graph should be make sense not beautiful
graph accuracy is imporant (more graph metrics required reflecting
just retry for invalid generation
유저가 이상하면 새로 생성함
named entity linking 으로 tool response 같은거 segmentation 해서 해당 부분 index range replacing
In production perspective, clients are continual user, reuse agent graph from original method
- In research, just a single shot is important but for production we can use updating strategy
- Ask the agent graph update or not since it is important
- Reference method is more suitible for realtime trace tracking but not for production due to the data storing issue, temporal or spatial changes chunk
Chunking
how about text spliting process done not chunker
another evaluation for ground truth chunking (unit of the trace is the LLM Call)
Agent Graph Ideas
add entity classifiers
진짜 agentic 하게 가려면 fetch 부터 해서 metadata 나 length 보고 좋은 trace 골라서
Seonglae Cho