Seonglae ChoVisualizaiton
Code Analysis visualzation
agent should be tree without loop → we may warn users to change request to prevent jailbreaking the prompt
how about make it linear since current graph is not very interprettable? → encapsualting agent graph with staring simplest version of garph visualization
from left to right the order of x axis should be time (input → agent/task/tool/feedback → output
shape algorithm flow chart
Graph Generation Optimization
improve by using scalable meethod (Single shot is cheap but not scalable)
context compreeion, document segmentation,
first using longest string matching with nlp select prompt candidates as an output → replace long candidates with variable with highly reduced context provided to llm
prompt extraction → replaced variable → agent graph construction
agent graph construction and prompt template extraction can be separated
may better to separate with agent graph construction and should be done with agent graph anlaysis
Post Feature Product
id 에서 해당 trace 만 가지고 id 하지말고 범용적으로 numbering 하거나 기능담는 id 로
If agent graph creates a loop then we should send clients to warn the agent loop might cause infinite loop
Analyze temporal difference analysis with sliding window triggered by user button
merging period button
Integrate knowledge graph with preexisting graph
Manual Knowledge graph Edit feature for eror correction
Enhance anomaly detection using network security domain's theory
This framework can be also applied to another graph genetaion such as service logs, network logs
같은 run observation input output 겹치는 부분 제거 output 에서 chat 특히 history 게속 같은거 포함