Support Ticket Triage

Classify intent, extract entities and sentiment in parallel, then draft a grounded response.

A support ticket comes in. The pipeline classifies the intent, extracts entities and sentiment in parallel, then drafts a response grounded in the classification.

Tree

Flow: "ticket-triage"
├── Block: "classify-intent" (llm)
├── Container (parallel)
│     ├── Block: "extract-entities" (llm)
│     └── Block: "score-sentiment" (llm)
└── Block: "draft-response" (llm)

Blocks

classify-intent (llm)

Classify the support ticket into one of: billing, bug_report, feature_request, account_access, other.

Ticket: {{message}}

Output schema:

{
  "type": "object",
  "properties": {
    "intent": { "type": "string", "enum": ["billing", "bug_report", "feature_request", "account_access", "other"] },
    "confidence": { "type": "number" }
  }
}

extract-entities (llm) — runs in parallel with sentiment

Extract any order numbers, product names, error codes, and dates from the ticket. Return empty arrays if none.

Ticket: {{message}}

score-sentiment (llm) — runs in parallel with entity extraction

Score the sentiment of the ticket as a number from -1 (furious) to 1 (delighted), and identify the dominant emotion in one word.

Ticket: {{message}}

draft-response (llm)

You are a support agent. Draft a reply to the ticket below.

- Intent: {{classify_intent.intent}}
- Entities: {{extract_entities}}
- Sentiment: {{score_sentiment}}

Match the tone to the sentiment. If the sentiment score is below -0.3, lead with an apology.

Ticket: {{message}}

Why this shape?

Intent classification first, because the response draft depends on it. Everything downstream is conditioned on knowing what the ticket is about.
Entities and sentiment in parallel because they don't depend on each other. Sequential would just add latency for no reason.
Final block sees all three signals. Naming each upstream block clearly (classify_intent, extract_entities, score_sentiment) lets the response prompt reference them explicitly without a code block to reshape the data.
The original ticket is re-passed to draft-response because the response needs the actual customer's words, not just structured signals. This is one of the few times re-passing raw input is correct.

Support Ticket Triage

Tree

Blocks

Why this shape?

On this page