# 🧠 RAG Document API**A modern, local-first Retrieval-Augmented Generation (RAG) REST API** built with **.NET 8/9**, **ASP.NET Core**, **Entity Framework Core**, **Wolverine**, and **LM Studio**.
Store documents → generate embeddings → semantic search → contextual LLM answers — **100% local, no cloud dependency**.
## ✨ Features- Upload and store documents in SQL Server
- Generate embeddings using Nomic Embed Text v1.5 (via LM Studio)
- Perform cosine-similarity-based semantic search
- Generate accurate, context-aware answers using a local LLM
- Clean layered architecture (Domain-Driven Design inspired)
- OpenAPI/Swagger documentation out of the box
## 🏗 Architecture Overview

Client ──→ ASP.NET Core API ──→ Wolverine Handlers ──→ RAG Pipeline ──→ LM Studio (LLM + Embeddings) └───────→ SQL Server (documents + embeddings)

### RAG Flow
1. **Ingestion** Text → Embedding (Nomic) → Store document + vector in SQL Server
2. **Query** Question → Embedding → Cosine similarity search → Top-K chunks → Build prompt (context + question) → LLM → Answer
## 🛠 Tech Stack
| Layer | Technology |
|--------------------|-----------------------------------|
| Web API | ASP.NET Core |
| Command Handling | Wolverine |
| ORM / Migrations | Entity Framework Core |
| Database | Microsoft SQL Server |
| LLM & Embeddings | LM Studio (OpenAI-compatible API) |
| Embedding Model | nomic-embed-text-v1.5 |
| Similarity | Cosine similarity (in-memory) |
| Documentation | Swashbuckle / OpenAPI |
## 📂 Project Structure

src/ ├── RAGDocument.API # ASP.NET Core Web API + Controllers ├── RAGDocument.Application # Commands, Queries, Handlers, Services ├── RAGDocument.Domain # Entities, Interfaces, Value Objects └── RAGDocument.Infrastructure # EF Core DbContext, Repositories, Persistence

## ⚙️ Configuration
1. Copy the example configuration file:
```bash
cp appsettings.example.json appsettings.json
# or on Windows:
copy appsettings.example.json appsettings.json

2. Review / edit appsettings.json (example):

{
"Logging": {
"LogLevel": {
"Default": "Information",
"Microsoft.AspNetCore": "Warning"
}
},
"ConnectionStrings": {
"RAGDbConn": "Server=localhost;Database=RAGDocumentDB;Trusted_Connection=True;TrustServerCertificate=True;"
},
"LLMS": {
"EndPointUrl": "http://localhost:1234",
"LLMModel": "qwen2.5-14b-instruct",
"EmbeddingModel": "nomic-embed-text-v1.5",
"MaxRetries": 3,
"TimeoutSeconds": 600,
"Temperature": 0.7,
"MaxTokens": 2048,
"TopK": 4
},
"AllowedHosts": "*"
}

Security note: Never commit appsettings.json or appsettings.Development.json containing real credentials. Use .gitignore, User Secrets, or Azure Key Vault in production.

🚀 Getting Started

Prerequisites

• .NET 8 or .NET 9 SDK
• SQL Server (local or Docker)
• LM Studio running with:
  • LLM model loaded (e.g. Qwen 2.5 14B Instruct)
  • Embedding model: nomic-embed-text-v1.5
  • Local inference server enabled (default: http://localhost:1234)

Steps

1. Clone the repository

git clone https://github.com/yourusername/RAGDocument_API.git
cd RAGDocument_API

2. Restore dependencies

dotnet restore

3. Apply database migrations

# Option A - from any project folder
dotnet ef database update --project src/RAGDocument.Infrastructure --startup-project src/RAGDocument.API
# Option B - if already in the API project folder
dotnet ef database update

4. Start LM Studio server (with both LLM and embedding model loaded)

5. Run the API

dotnet run --project src/RAGDocument.API

Swagger UI will be available at:
👉 https://localhost:5001/swagger or http://localhost:5000/swagger

📡 API Endpoints (Summary)

Method	Endpoint	Description
POST	/api/document/AddDocument	Upload text document
POST	/api/document/Query	Ask question → get RAG answer

Example – Add Document (multipart/form-data)

POST /api/document/AddDocumentContent-Type: multipart/form-dataContent: "The quick brown fox jumps over the lazy dog..."

Example – Query

POST /api/document/QueryContent-Type: multipart/form-dataQuestion: What does the fox do?

🧩 Core Components

• EmbeddingService – generates vectors using LM Studio embedding endpoint
• RAGSystem – orchestrates retrieval + prompt building
• LMStudioService – calls chat completions endpoint
• UnitOfWork / Repositories – clean database access

⚠️ Current Limitations

• In-memory cosine similarity (not suitable for >50k documents)
• No automatic document chunking / splitting
• No hybrid (BM25 + vector) search yet
• No rate limiting or authentication (add in production)

🔮 Planned Improvements

• Vector indexing (SQL Server 2022 vector type or pgvector)
• Intelligent chunking & metadata filtering
• Hybrid retrieval (keyword + semantic)
• Response streaming
• Authentication & authorization
• Caching (Redis / in-memory)
• Evaluation metrics & reranking

📄 License

MIT License

Feel free to use, modify, and contribute!

❤️ Contributing

Pull requests, bug reports, and feature suggestions are welcome!