Appendix D

File to run NeMo Agent Toolkit sizing against AI-Q, aiq-test-rtxpro6k.yml

general:
 use_uvloop: true
 telemetry:
   tracing:
     phoenix:
       _type: phoenix
       endpoint: http://10.184.203.86:6006/v1/traces
       project: default
 front_end:
   _type: fastapi
   endpoints:
     - path: /generate_query
       method: POST
       description: Creates the query
       function_name: generate_query
     - path: /generate_summary
       method: POST
       description: Generates the summary
       function_name: generate_summary
     - path: /artifact_qa
       method: POST
       description: Q/A or chat about a previously generated artifact
       function_name: artifact_qa
     - path: /aiqhealth
       method: GET
       description: Health check for the AIQ AIRA service
       function_name: health_check
     - path: /default_collections
       method: GET
       description: Get the default collections
       function_name: default_collections
     # Add profiler endpoint
     - path: /analyze_performance
       method: POST
       description: Analyze performance metrics using profiler agent
       function_name: profiler_agent


llms:
 instruct_llm:
   _type: openai
   model_name: meta/llama-3.3-70b-instruct
   temperature: 0.0
   base_url: http://10.184.203.82:8000/v1
   api_key: not-needed
   stream: false
   max_retries: 1  # Default is 3
   timeout: 120
  nemotron:
   _type: openai
   model_name: nvidia/llama-3.3-nemotron-super-49b-v1.5
   temperature: 0.0
   base_url: http://10.184.203.83:8000/v1
   stream: false
   api_key: not-needed
   max_retries: 1  # Default is 3
   timeout: 120

functions:
 generate_query:
   _type: generate_queries
   llm_name: nemotron

 generate_summary:
   _type: generate_summaries
   rag_url: http://10.184.203.84:8081/v1
   timeout: 60

 artifact_qa:
   _type: artifact_qa
   llm_name: instruct_llm
   rag_url: http://10.184.203.84:8081/v1
   timeout: 60

 # Add profiler agent and related tools
 profiler_agent:
   _type: profiler_agent
   llm_name: instruct_llm
   max_iterations: 4
   max_retries: 3
   tools:
     - px_query
     - flow_chart
     - token_usage
     - response_composer
   output_dir: .tmp/aiq/aira/profiler_agent/  # Add persistent storage

 px_query:
   _type: px_query
   phoenix_url: http://10.184.203.86:6006
   time_window_seconds: 600000
   default_project_name: default

 flow_chart:
   _type: flow_chart

 token_usage:
   _type: token_usage

 response_composer:
   _type: response_composer



workflow:
 _type: ai_researcher
 timeout: 600

eval:
 general:
   output_dir: .tmp/eval/aiq-test-rtxpro6k
   dataset:
     _type: json
     file_path: .tmp/research_assistant_dataset.json
   profiler:
     base_metrics: true
     token_uniqueness_forecast: false
     workflow_runtime_forecast: false
     compute_llm_metrics: true
     csv_exclude_io_text: true
     prompt_caching_prefixes:
       enable: false
       min_frequency: 0.1
     bottleneck_analysis:
       enable_nested_stack: false
  # # Evaluators for quality assessment during sizing calculations
 # evaluators:
 #   research_quality_evaluator:
 #     _type: tunable_rag_evaluator
 #     llm_name: instruct_llm  # Using your existing LLM
 #     judge_llm_prompt: |
 #       You are an expert evaluator for AI research assistant outputs. Assess the quality of the generated research content based on:
 #       1. Coverage: How comprehensively does it address the research question?
 #       2. Correctness: How accurate and factual is the information?
 #       3. Relevance: How relevant is the content to the specific research domain?
 #       4. Citation Quality: How well are sources cited and referenced?
       
 #       Rate each aspect from 0.0 to 1.0.
 #     default_scoring: true
 #     default_score_weights:
 #       coverage: 0.3
 #       correctness: 0.3
 #       relevance: 0.2
 #       citation_quality: 0.2