SAG Agents

SAG (Small Agentic Group) agents are deterministic pre-processing agents that handle common tasks without using LLM calls. This makes responses faster and more cost-effective.

Overview

SAG agents use rule-based logic, keyword matching, and regex patterns to:

• Classify user intent
• Detect sentiment and urgency
• Find products and FAQ matches
• Handle greetings/farewells directly
• Extract appointments and tasks

The 6 SAG Agents

Agent	Purpose	LLM?
IntentSAG	Classify message intent	No
SentimentSAG	Detect mood/urgency	No
PersonaSAG	Get persona context	No
BusinessSAG	Find products/FAQ	No
TaskSAG	Extract tasks/reminders	No
GreetingSAG	Handle social messages	No

IntentSAG

Classifies user intent into categories:

Greetings & Social

GREETING  → "salut", "buna", "hello"
FAREWELL  → "pa", "bye", "la revedere"
THANKS    → "multumesc", "mersi", "thanks"

Questions

QUESTION_PRICE        → "cat costa", "pret"
QUESTION_PRODUCT      → "aveti", "exista", "caut"
QUESTION_AVAILABILITY → "in stoc", "disponibil"
QUESTION_HOURS        → "program", "deschis"
QUESTION_SHIPPING     → "livrare", "transport"
QUESTION_RETURNS      → "retur", "schimb"

Actions

ORDER     → "comanda", "vreau sa cumpar"
COMPLAINT → "reclamatie", "problema"
SCHEDULE  → "programare", "rezervare"
CANCEL    → "anuleaza", "renunt"
HUMAN     → "operator", "om", "agent"

SentimentSAG

Detects emotional state and urgency:

Sentiment	Indicators
Positive	"super", "excelent", "multumesc mult"
Negative	"dezamagit", "nasol", "nu merge"
Urgent	"urgent", "acum", "imediat"
Frustrated	Multiple punctuation, caps, complaints

GreetingSAG

Handles greetings directly without LLM:

User: "Salut!"
→ GreetingSAG responds: "Bună! Cu ce te pot ajuta?"

User: "Pa pa"
→ GreetingSAG responds: "La revedere! O zi frumoasă!"

This saves LLM tokens for actual questions.

BusinessSAG

Matches user queries against:

• Product catalog (name, description, keywords)
• FAQ entries
• Business policies (returns, shipping, hours)

User: "Aveti laptop-uri gaming?"
→ BusinessSAG: Finds matching products
→ Returns: [Product A, Product B, Product C]

TaskSAG

Extracts structured data from messages:

User: "Vreau o programare maine la 14:00"
→ TaskSAG extracts:
   {
     type: "appointment",
     date: "2026-04-21",
     time: "14:00"
   }

How SAG Works in Flow

1. Message arrives
2. SAG agents analyze in parallel:
   - IntentSAG → What does user want?
   - SentimentSAG → How do they feel?
   - BusinessSAG → Any product/FAQ matches?
   - GreetingSAG → Is it just a greeting?

3. If GreetingSAG matches → Respond directly (no LLM)
4. Otherwise → Pass enriched context to LLM

Performance Benefits

Metric	With SAG	Without SAG
Greeting response	~50ms	~800ms
Intent detection	~10ms	~500ms
LLM calls saved	~30%	0%
Token usage	Lower	Higher

Multi-Language Support

SAG agents support Romanian and English patterns:

// Romanian
"cat costa" → QUESTION_PRICE
"vreau sa cumpar" → ORDER

// English
"how much" → QUESTION_PRICE
"I want to buy" → ORDER

Configuration

SAG agents are enabled by default. Customize in widget settings:

{
  "sag": {
    "greeting_responses": true,
    "intent_classification": true,
    "sentiment_tracking": true,
    "fallback_to_llm": true
  }
}