Ticket_resolver_crewai

<div align="center"> <img src="https://capsule-render.vercel.app/api?type=waving&color=gradient&customColorList=6,11,20&height=230&section=header&text=SupportCrew%20AI&fontSize=64&fontColor=ffffff&fontAlignY=38&desc=Multi-agent%20support%20with%20a%20QA%20feedback%20loop%20%E2%80%A2%20100%25%20local&descAlignY=60&descSize=18&animation=fadeIn" alt="SupportCrew AI banner" width="100%"/> <a href="https://github.com/Mahee0117/Ticket_resolver_crewai"> <img src="https://readme-typing-svg.demolab.com?font=Fira+Code&weight=600&size=20&pause=1200&color=8B5CF6&center=true&vCenter=true&width=680&lines=Triage+%E2%86%92+Route+%E2%86%92+Resolve+%E2%86%92+QA;Rejected+answers+get+feedback+and+a+retry;Runs+fully+local+with+Ollama+%2B+Qwen2" alt="Typing animation" /> </a> <br/>

Overview · What's new in v2 · How it works · Agents · Quick start · Roadmap

</div>

---

📌 Overview

SupportCrew AI is a multi-agent customer support system built with CrewAI.

A ticket comes in. A Triage Agent classifies it and sets its priority. The ticket is routed to the right specialist agent, which drafts a customer-friendly response. Then a QA Agent reviews that response, and if it isn't good enough, the specialist gets feedback and tries again.

💡 Why multi-agent? Each agent has one clear job. Specialists give focused answers, and a separate reviewer catches what the writer missed.

---

🆕 What's new in v2

<table> <tr> <td width="33%" valign="top">

🧪 QA Agent

A new reviewer checks every specialist response for relevance, correctness and usefulness before it goes out.

</td> <td width="33%" valign="top">

🔁 Feedback & retry

Rejected responses go back to the specialist together with the QA feedback, with a bounded retry limit.

</td> <td width="33%" valign="top">

🏠 Fully local

Moved from Hugging Face to Ollama + Qwen2 7B. No API credits or tokens required.

</td> </tr> </table>

| | v1 | v2 | |---|---|---| | Flow | Ticket → Triage → Specialist → Response | Ticket → Triage → Specialist → QA → Final response or retry | | QA review | ❌ | ✅ structured QAResult | | Feedback loop | ❌ | ✅ up to 2 retries | | LLM | Hugging Face (hosted) | Ollama qwen2:7b (local) | | Code layout | Single file | Modular: agents, tasks, models, config, router, tickets |

---

🧠 How it works

flowchart TD
    A([🎫 Customer Ticket]) --> B[🔎 Triage Agent]
    B --> C{{📦 Structured Triage Result<br/>category · priority · issue}}
    C --> D{🧭 Python Router}
    D -->|Billing| E[💳 Billing Agent]
    D -->|Technical| F[🛠️ Technical Agent]
    D -->|Account| G[👤 Account Agent]
    D -->|General| H[💬 General Support Agent]
    E --> I[📝 Customer Response]
    F --> I
    G --> I
    H --> I
    I --> J[🧪 QA Agent]
    J -->|✅ APPROVED| K([📨 Final Response])
    J -->|❌ REJECTED + feedback| L{Retries left?}
    L -->|Yes| D
    L -->|No| K

    style A fill:#8b5cf6,color:#fff,stroke:none
    style B fill:#6366f1,color:#fff,stroke:none
    style C fill:#0ea5e9,color:#fff,stroke:none
    style D fill:#f59e0b,color:#fff,stroke:none
    style J fill:#ec4899,color:#fff,stroke:none
    style K fill:#22c55e,color:#fff,stroke:none

<table> <tr> <td width="20%" align="center"><h3>1️⃣</h3><b>Ticket</b><br/>_{A customer describes the problem in plain language}</td> <td width="20%" align="center"><h3>2️⃣</h3><b>Triage</b><br/>_{Category, priority and issue summary are extracted}</td> <td width="20%" align="center"><h3>3️⃣</h3><b>Route</b><br/>_{Plain Python logic picks one specialist, with no extra LLM call}</td> <td width="20%" align="center"><h3>4️⃣</h3><b>Resolve</b><br/>_{The specialist drafts a response}</td> <td width="20%" align="center"><h3>5️⃣</h3><b>QA</b><br/>_{Approve it, or send it back with feedback}</td> </tr> </table>

🔁 The QA retry loop

The loop is bounded on purpose. An open-ended while not approved could run forever, so the system stops after a fixed number of retries.

MAX_RETRIES = 2   # 1 first attempt + up to 2 retries = max 3 attempts

flowchart LR
    A1[Attempt 1] --> Q1[QA]
    Q1 -->|Reject| A2[Attempt 2]
    A2 --> Q2[QA]
    Q2 -->|Reject| A3[Attempt 3]
    A3 --> Q3[QA]
    Q3 -->|Approve or limit reached| Z([Stop])
    Q1 -->|Approve| Z
    Q2 -->|Approve| Z

On a retry, the specialist receives its previous response and the QA feedback, then improves the draft:

PREVIOUS RESPONSE:
...
QA FEEDBACK:
...
Improve the previous response based on the QA feedback.

Illustrative example of what a rejection could look like:

QAResult(
    approved = False,
    feedback = "Do not assume the customer uses Windows. The response is unnecessarily verbose."
)

📦 Structured outputs

Both the triage and QA agents return validated Pydantic models instead of free text.

class TriageResult(BaseModel):
    category: str
    priority: str
    issue: str

class QAResult(BaseModel):
    approved: bool
    feedback: str

---

🧪 Tested tickets

Four tickets ran through the full Triage → Specialist → QA pipeline on local Qwen2 7B:

| # | Ticket | Category | Priority | Result | |:-:|---|:-:|:-:|:-:| | 1 | Password reset problem | Account | Medium | ✅ Approved | | 2 | App crashes on PDF upload | Technical | High | ✅ Approved | | 3 | Profile picture issue | Account | Low | ✅ Approved | | 4 | Unauthorized access and fraudulent transactions | Account | Urgent | ✅ Approved |

🔎 Honest note: all four responses were approved on the first attempt, so the reject → feedback → retry path is implemented but not yet demonstrated in a real run.

<!-- 📸 Add a terminal screenshot or GIF of a real run here, for example: <p align="center"><img src="docs/demo.gif" alt="SupportCrew AI demo" width="80%"/></p> -->

---

🦸 Meet the crew

<table> <tr> <td width="20%" align="center"><h1>🔎</h1><b>Triage Agent</b><br/>_{<i>Senior Customer Support Triage Specialist</i>}</td> <td> Classifies the ticket, assigns a priority and summarises the issue.<br/><br/> <b>Categories:</b> <code>Billing</code> <code>Account</code> <code>Technical</code> <code>General</code><br/> <b>Priorities:</b> <code>Low</code> <code>Medium</code> <code>High</code> <code>Urgent</code> </td> </tr> <tr> <td align="center"><h1>💳</h1><b>Billing Agent</b></td> <td>Payments · Refunds · Duplicate charges · Subscriptions · Billing problems</td> </tr> <tr> <td align="center"><h1>🛠️</h1><b>Technical Agent</b></td> <td>Application crashes · Bugs · Errors · Technical failures · Troubleshooting</td> </tr> <tr> <td align="center"><h1>👤</h1><b>Account Agent</b></td> <td>Login problems · Password issues · Account access · Profile problems · Account security</td> </tr> <tr> <td align="center"><h1>💬</h1><b>General Support Agent</b></td> <td>General questions · Information requests · How-to questions · Basic customer assistance</td> </tr> <tr> <td align="center"><h1>🧪</h1><b>QA Agent</b><br/>_{<i>Customer Support Quality Assurance Specialist</i>}</td> <td> Reviews each specialist response and returns a structured verdict.<br/><br/> <b>Checks:</b> Does it address the issue? · Is it relevant? · Is it useful? · Does it ignore important information? </td> </tr> </table>

---

🏗️ Architecture

main.py is now a thin orchestrator. The work is split into small functions.

main.py
│
├── run_triage()         → creates the triage task + crew, returns TriageResult
├── run_specialist()     → routes by category, runs the matching specialist
├── run_quality_check()  → creates the QA task + crew, returns QAResult
└── main()               → ticket → triage → specialist → QA → retry if needed

The router (category → agent) is deterministic Python, not another LLM call.

---

🛠️ Tech stack

| Layer | Tools | |---|---| | 🐍 Language | Python | | 🤝 Agent framework | CrewAI | | 🧠 LLM runtime | Ollama | | 🤖 Model | qwen2:7b (local) | | ✅ Data validation | Pydantic | | 📦 Packaging | uv |

---

📁 Project structure

Ticket_resolver_crewai/
│
├── src/
│   └── supportcrew_ai/
│       ├── __init__.py
│       ├── main.py       # orchestration + QA retry loop
│       ├── config.py     # local Ollama LLM setup
│       ├── models.py     # TriageResult, QAResult
│       ├── agents.py     # all six agents
│       ├── tasks.py      # all tasks, incl. QA task
│       ├── router.py     # category → specialist
│       └── tickets.py    # sample tickets
│
├── .gitignore
├── .python-version
├── pyproject.toml
├── uv.lock
└── README.md

🧩 Agents and tasks each live in a single file on purpose. At this size, one file per agent would be needless fragmentation.

---

🚀 Quick start

Prerequisites: uv and Ollama

1. Clone the repository

git clone https://github.com/Mahee0117/Ticket_resolver_crewai.git
cd Ticket_resolver_crewai

2. Install dependencies

uv sync

3. Pull the local model and make sure Ollama is running

ollama pull qwen2:7b

Ollama serves on http://localhost:11434 by default.

4. Run it

uv run python src/supportcrew_ai/main.py

✅ No API keys needed. v2 runs entirely on your machine.

Using a different model? Change the model in config.py:

llm = LLM(
    model="ollama/qwen2:7b",
    base_url="http://localhost:11434",
)

---

⚠️ Known limitations

• Verbose answers. Qwen2 7B sometimes writes long, speculative troubleshooting for simple tickets. For example, the PDF crash ticket got a far more extensive reply than it needed.
• Retry path untested in practice. QA approved everything on the first attempt in the latest run.
• Lenient QA. The review criteria are basic and can be tightened (conciseness, less speculation).

These are planned improvements, not blockers for v2.

---

🗺️ Roadmap

SupportCrew AI is built incrementally. Each version adds one new agentic capability.

| Version | Milestone | Status | |:---:|---|:---:| | v1.0 | Multi-agent triage and routing | ✅ Done | | v2.0 | QA Agent + feedback retry loop + local Ollama | ✅ Done | | v3.0 | Tools + knowledge base (RAG) | 🗓️ Planned | | v4.0 | CrewAI Flow + human escalation | 🗓️ Planned | | v5.0 | FastAPI + PostgreSQL + React | 🗓️ Planned |

<details> <summary>✅ <b>v1.0: Multi-Agent Triage &amp; Routing</b> (completed)</summary> <br/>

• Triage, Billing, Technical, Account and General Support agents
• Category and priority classification
• Pydantic structured output
• Category-based routing and specialist task execution
• Multiple ticket processing

</details> <details> <summary>✅ <b>v2.0: QA &amp; Feedback Retry</b> (completed)</summary> <br/>

• QA Agent with structured QAResult (approved, feedback)
• Specialist tasks accept previous_response and qa_feedback
• Bounded retry loop (MAX_RETRIES = 2)
• Modular code: run_triage(), run_specialist(), run_quality_check()
• Switched to local Ollama (qwen2:7b), no hosted API needed

</details> <details> <summary>🟠 <b>v3.0: Tools &amp; Knowledge Base</b> (planned)</summary> <br/>

Specialists gain tools and can look things up in company documentation, so they work with real data instead of only the ticket text.

get_customer()   get_order()   get_payment()   get_account()   search_logs()

Potential knowledge sources: FAQs · Refund policies · Account recovery docs · Product docs · Troubleshooting guides

flowchart LR
    T[🎫 Ticket] --> S[Specialist Agent] --> K[(📚 Knowledge Base)] --> D[Relevant Docs] --> R([✅ Accurate Resolution])

</details> <details> <summary>🟣 <b>v4.0: CrewAI Flow &amp; Human Escalation</b> (planned)</summary> <br/>

Move orchestration into a structured CrewAI Flow, and escalate to a human for:

• High-risk security issues
• Complex unresolved tickets
• Tickets that still fail QA after the retry limit
• Anything that needs human judgement

</details> <details> <summary>🚀 <b>v5.0: Production Application</b> (planned)</summary> <br/>

flowchart LR
    A[⚛️ React Frontend] --> B[⚡ FastAPI Backend] --> C[🤝 CrewAI Workflow] --> D[🧰 Agents + Tools + RAG] --> E[(🐘 PostgreSQL)]

</details>

🚧 Everything from v3.0 onward is a plan, not a feature that exists today.

---

📚 Learning journey

This project doubles as a hands-on way of learning CrewAI, built one layer at a time:

Single Agent → Multiple Agents → Routing → Structured Outputs → QA + Feedback
      → Tools + RAG → Workflow Orchestration → Full Application

Each version marks a new stage of understanding and implementation.

---

👨‍💻 Author

<div align="center">

Mahesh M S K <br/> _{Computer Science student · Agentic AI · Multi-Agent Systems · Generative AI · DevOps · Cloud · Full-Stack}

<br/>

<br/>

Current version: v2.0 · 🚧 Actively being developed

⭐ If you find this project interesting, consider giving it a star!

<img src="https://capsule-render.vercel.app/api?type=waving&color=gradient&customColorList=6,11,20&height=110&section=footer" alt="footer" width="100%"/> </div>