Customer Support RAG Triage Agent

Overview

A retrieval-grounded support triage workflow with observable steps, citations, evaluation, and fallback behavior.

Problem

Support triage combines classification, retrieval, response drafting, and escalation decisions that need to remain inspectable and safely bounded.

Solution

Built a typed workflow for ticket normalization, intent and urgency handling, retrieval, grounded answer generation, source checks, and safe fallback behavior.

Technical Decisions

• A fixed workflow replaces an open-ended agent loop.
• Provider failures degrade to bounded output rather than bypassing guardrails.

Outcome

The project provides an API contract, a deterministic offline path, evaluation artifacts, guardrails, and a public reviewer demo.

Key Features

• Typed workflow state keeps triage behavior inspectable.
• Grounding and source checks constrain generated responses.
• Offline evaluation and deterministic fallbacks preserve a stable review path.

Architecture

1. 01

   Ticket input

2. 02

   Intent and urgency

3. 03

   Retrieval

4. 04

   Ranking

5. 05

   Grounded generation

6. 06

   Source checks

7. 07

   Human action

Tech Stack

• Python
• FastAPI
• RAG
• LangGraph
• Qdrant
• React
• TypeScript
• Evaluation

Verification

• Offline deterministic demo
• Retrieval and grounding evaluation
• API contract checks

Security & Privacy

• Provider credentials remain server-side.
• Guardrails and escalation preserve human review for uncertain output.

Limitations

• The bundled support corpus is a review fixture, not a private production knowledge base.
• Generated drafts still require policy-aware human review.

Future Improvements

• Add stronger reranking and multilingual retrieval.
• Capture structured reviewer feedback for evaluation.