MultiAgent - E-Commerce AI Customer Support- 🤖🛍️

An AI-powered customer-service crew that ingests policy documents, classifies support tickets, and produces fully-cited, policy-compliant resolutions automatically.

Fast ingest Smart RAG CrewAI Gemini Python313+

---

Table of Contents

• Why This Repo?
• Architecture
• Project Structure
• Project Structure
• Outputs
• Usage
• Python API Usage
• Evaluation Suite
• Extending Agents or Policies
• Contributing
• Troubleshooting
• License

---

Why This Repo?

Online stores face thousands of repetitive support requests.
This repo automates first-line responses:

• Understands tickets via LLM triage
• Retrieves only relevant policy chunks
• Writes fully-cited replies or escalates intelligently
• 20 real-world test-cases baked in (≈ 85 % accuracy)

No GPUs needed—everything runs on Google's Gemini embeddings , groq/llama-3.1-8b-instant .

---

Architecture

1. Triage Agent (crewai) – classify ticket, detect missing info
2. Retriever Agent – Policy Search Tool → exact Chroma collection (refund, dispute, shipping, …) via Gemini embedding-001
3. Writer Agent – strict prompt: every claim must cite doc, else escalate
4. Compliance Agent – reject hallucinations (citations < 1, fuzzy language)

All results conform to app.schemas.io.Output.

---

Project Structure

.
├── app/                       # Core source
│   ├── agents/                # CrewAI definitions
│   ├── core/                  # Gemini LLM wrapper
│   ├── graph/                 # Crew → workflow
│   ├── rag/                   # Embedding, ingestion, retrieval
│   ├── evaluation/            # Accuracy scoring
│   └── schemas/               # Pydantic models
├── data/
│   └── policies/              # Plaintext knowledge base
├── tests/
│   ├── test_cases.json        # 20 golden examples
│   └── test_e2e.py            # Pytest harness
├── ingestion_trigger.py       # One-click ingestion
└── pyproject.toml             # PEP-621 deps

---

Outputs

---

Installation

Prerequisites

Python 3.13+
Groq API key Gemini API key

uv (recommended)

uv init 
uv add -r requirements.txt

All dependencies listed in pyproject.toml.

---

Usage

SET env variables 

# (1) Make sure Chroma is populated first
uv run ingestion_trigger.py

# (2) Run an interactive loop
uv run python -m app.main
     # default ticket in file

Change the hard-coded ticket string in app/main.py or:

python -c "
from app.main import workflow
print(workflow.run('Received damaged laptop – screen cracked, order #12345'))"

Sample output

{'classification':'refund',
 'confidence':0.93,
 'decision':'approve',
 'rationale':'Policy §Damaged Item allows refund/replace – evidence required...',
 'response':'We’re sorry... full refund processed.',
 'citations':['data/policies/disputes.txt', 'data/policies/returns_refunds.txt'],
 'notes':'APPROVED by compliance'}

---

Python API Usage

from app.main import workflow
payload = "Package lost, marked delivered but never arrived."

result = workflow.run(payload)
print(result.decision)   # → 'approve'

To re-use individual Crew agents (advanced):

factory = AgentFactory(tools)
triage_agent  = factory.triage()

---

Evaluation Suite

Evaluate on 20 curated tickets (accuracy, citation-score, escalation-rate):

pytest -q

Results summary printed per test-case from tests/test_cases.json.

---

Extending Agents or Policies

1. Drop new *.txt files into data/policies/
2. Re-ingest: python ingestion_trigger.py
3. Add new classification in RAGPipeline.get_collection() mapping
4. Re-run tests: pytest -q
5. (Opt.) tweak Crew prompts in app/agents/crew_agents.py for stricter leniency

---

Contributing

We ❤️ pull requests.
Please:

1. Fork & branch (feature/, fix/, docs/)
2. Add a unit or integration test (tests/)
3. Ensure pytest green
4. Update this README for user-facing changes
5. Submit PR with concise description

Code style = Black ≥ 24, isort, pep-8 naming.

---

Troubleshooting

• ValueError: GOOGLE_API_KEY not found ➜ fill .env
• chromadb.errors on first run ➜ ensure you've run ingestion
• Reply too cautious (many escalations) ➜ lower Compliance Agent strictness or enrich policy docs
• Embeddings slow ➜ Gemini free tier limits; consider batching

---

License

MIT – see LICENSE file. No warranties; commercial use at your own risk.