Anatomy Bench

raw query → query_prep → dialog resolver → multi-path retrieval → selected agent → saved report
user id
topic id
Topic overviewstate snapshot
[no topic selected]
topK bm25
topK vector
max chunks
prompt tokens
agent
answer tokens
temperature
LLM
Selected agentregistry metadata

      

    

    
Agents are loaded from the registry file and refreshed before each run.


    

      

        

          
Answermarkdown

          

        

        

          
Topic summary20k users × 20 requests/day

          
Run a query or load a topic to see aggregate cost, cache and focus.

        

      

      

        
Run statslast request

        

      

    


    

      
Topic historymessages + turn usage + structured hints

      
No topic loaded yet.

    


    

      
Last report JSONtruncated for debugging