SupportAgent.NET: An AI Customer Support Copilot

Open-source project Public repository, under active development

SupportAgent.NET is my open-source project demonstrating how AI can work inside a customer-support application. It combines .NET business services with document retrieval, model tool calling, and draft replies that remain under human review.

The Problem

Support agents often answer the same questions, such as "Where is my order?", by switching between customer records, order systems, and policy documents. A standalone chatbot cannot answer these questions reliably because it has no safe access to the business data. SupportAgent.NET explores a different design: the language model acts as an intelligence layer over normal .NET services, and a human agent stays in control of what is sent to the customer.

My Role

I designed and built the project end to end: the ASP.NET Core API, the React support workspace, the SQL Server data model, the AI gateway, the tools, and the retrieval pipeline.

Example Workflow

This workflow uses the repository's sample development data, not real customer information.

1. A support agent opens ticket 1001, "Where is my order?", from sample customer John Smith, and asks the assistant for the order status.
2. The model decides it needs data and makes a native tool call to GetOrderStatus, rather than relying on keyword routing.
3. The .NET order service returns only the records the signed-in user is permitted to see.
4. For policy questions, SearchKnowledgeBase retrieves relevant company knowledge, such as the shipping policy.
5. The assistant drafts a reply grounded in those results. The agent reviews and edits it; nothing is sent automatically.

Implemented Scope

• Provider-independent AI gateway supporting Ollama (local models) and OpenAI
• Native tool calling over .NET services: GetCustomer, GetOrderStatus, and SearchKnowledgeBase
• Retrieval over company documents: chunking, embeddings, and hybrid semantic and lexical search, stored in SQL Server
• React support workspace with tickets, conversations, and an AI assistant panel
• Structured draft replies for human review, with AI conversation state persisted
• Role-based access and tenant-scoped data enforced on the server, never accepted from the browser or a tool call
• Read-only suggested actions, such as viewing an invoice, that a human must initiate

Technology Used

C#, ASP.NET Core Web API, Entity Framework Core, SQL Server, React, TypeScript, Vite, Ollama or the OpenAI API for the model, nomic-embed-text embeddings, and xUnit tests.

Limitations

• It is a public demonstration and starter project, not a deployment for a paying client.
• Embeddings are stored in SQL Server rather than a dedicated vector database.
• Tool calling requires a model with native function calling, and answer quality depends on the chosen model.
• Connecting it to a real ERP or CRM requires reviewing that system's data and permissions first.

Evidence

The source code, setup instructions, and documentation are public on GitHub: rahmatafridi/SupportAgent.NET.

Discuss a Similar Integration