Agent Reference¶

LLM-Driven vs Rule-Based¶

Understanding where LLM reasoning is applied versus deterministic logic is important for cost modeling, debugging, and enterprise justification.

Component	Module	Model	Calls per Invocation
Quadrant classification	`orchestrator.py`	Groq / Llama 3.3 70B	1 per run
Specialist recommendations	`subagents.py` (×5 parallel)	Groq / Llama 3.3 70B	1 per agent per run
Synthesis narrative	`orchestrator.py` — synthesizer node	Groq / Llama 3.3 70B	1 per run
Confidence scoring	`llm_tools.py` — embedded in recommendation call	Groq / Llama 3.3 70B	Part of subagent call
Intent routing (ambiguous)	`update_agent.py` — `classify_input()`	Groq / Llama 3.3 70B	0–1 per command
Chart complexity classification	`chart_agent.py`	Groq / Llama 3.3 70B	1 per chart request
Chart spec / figure generation	`chart_agent.py`	Groq / Llama 3.3 70B	1 per chart request
Registry command interpretation	`update_agent.py`	Groq / Llama 3.3 70B	1–2 per command
Cross-item RCA	`rca_agent.py`	Groq / Llama 3.3 70B	1 per RCA request
5 Whys causal chain	`whys_agent.py`	Groq / Llama 3.3 70B	1 per category analyzed
Predictive quadrant preview	`quadrant_preview.py`	Groq / Llama 3.3 70B	1 per preview (deduped)
Completeness scoring	`completeness_agent.py`	Groq / Llama 3.3 70B	1 per score (deduped)
Document intake	`intake_agent.py`	Gemini 2.5 Flash-Lite	1 per uploaded document
Spreadsheet analytics + correlation	`analytics_agent.py`	Gemini 2.5 Flash-Lite	2 per uploaded file
Schema metric discovery	`schema_agent.py`	Gemini 2.5 Flash-Lite	1 per discovery run

Component	Logic
Stale detection	`days_since_update >= 14` — no LLM
Trigger → category filter	Keyword map in `orchestrator.py` — no LLM
HU/HI-only mode	Trigger keyword list — no LLM
HITL deferral filtering	Set operations on item IDs — no LLM
Item ID generation	Sequential counter per category prefix — no LLM
Command intent (unambiguous)	Regex + keyword heuristic in `update_agent.py` — no LLM
Chart rendering	Plotly, derived from classified state — no LLM
PDF export	reportlab template — no LLM