RAG-CSharpAssistant

A cross-platform AI chat assistant for C# and .NET developers

Offline retrieval meets real-time streaming generation — so answers are grounded, fast, and honest about what they know.

---

What is this?

Most AI chat apps send your question straight to an LLM and hope for the best. This app does something smarter:

1. Validates — is this actually a C#/.NET question?
2. Retrieves — searches a local SQLite knowledge base (built from Stack Overflow C# data) for relevant chunks using BM25-ranked FTS5
3. Grounds — injects retrieved context into the prompt before generation
4. Streams — shows the Groq response token-by-token in real time

If no local context is found, the app still answers — but tells you so with a visible disclaimer. If the question is out of scope entirely, it says that too. No hallucinated confidence.

---

Screenshots

Splash	Empty State	Active Chat	No Context
App initialization with background data preloading	Guided entry with logo, tagline, and suggested prompts	Real-time streaming responses with grounding indicators	No relevant context detected — fallback answer generation

---

Architecture

User Input
    │
    ▼
┌─────────────────────────────────────┐
│         Scope Validation            │  llama-3.1-8b-instant, max_tokens=5
│   IsCSharpScopeAsync → YES / NO    │  Returns "YES" / "NO" only
└───────────────┬─────────────────────┘
                │ YES
                ▼
┌─────────────────────────────────────┐
│          FTS5 Retrieval             │  SQLite, BM25 ranking
│   AND search → OR fallback         │  Up to 3 chunks returned
│   RagSearchService.SearchAsync     │
└───────────────┬─────────────────────┘
                │
        ┌───────┴────────┐
        │                │
    Chunks found     No chunks
        │                │
        ▼                ▼
  Grounded prompt   Ungrounded prompt
  + disclaimer      + disclaimer
        │                │
        └───────┬────────┘
                ▼
┌─────────────────────────────────────┐
│         Groq API (streaming)        │  SSE, temperature=0.1
│   GenerateStreamAsync              │  HttpClient timeout = Infinite
└───────────────┬─────────────────────┘
                ▼
         Streaming Chat UI

Layer breakdown

Layer	What lives here
UI	MainPage.xaml, LoadingPage.xaml, AppShell.xaml — MAUI XAML + C# code-behind
ViewModel	MainPageViewModel.cs — MVVM, ObservableCollection<ChatMessage>, streaming pipeline orchestration
Services	RagSearchService, GroqClient, PromptBuilder, AppSecrets
Storage	csharp_knowledge.db — SQLite FTS5, bundled as MAUI raw asset
Data pipeline	DataPipeline/build_db.py — Python offline script, not shipped in app

---

Key design decisions

Startup: no white flash

The native splash (Maui.SplashTheme) and the windowBackground of Maui.MainTheme.Base are both set to #F5EFE6 (warm beige). This eliminates the white frame that typically appears between the OS splash and the first MAUI frame on Android. LoadingPage then holds for ~1.2 s while warming the DB, then swaps the window page to AppShell.

Singleton MainPage

MainPage is registered as a singleton in DI and pre-inflated during LoadingPage.OnAppearing. When AppShell resolves it via DataTemplate, it gets the same pre-warmed instance — chat history survives navigation and startup is perceptibly faster.

FTS5 query construction

Raw user input is sanitized (strip ' and ", split on non-word chars, drop tokens < 2 chars), then tokens are quoted ("token"*) to handle C#-specific characters like #, ., <>. AND precision search runs first; if empty, OR recall runs as fallback. On any SqliteException the service returns empty results gracefully.

Prompt strategy

Two prompt paths exist — grounded (context from DB) and ungrounded (general C#/.NET knowledge) — each with a user-visible disclaimer prepended to the streamed response. The scope-check prompt uses max_tokens: 5 and temperature: 0 to get a deterministic YES/NO. The generation prompt uses temperature: 0.1 with a single user message (no system role) for Flutter-parity behavior with llama-3.1-8b-instant.

Threading

Everything that touches the network or DB runs on the thread pool via Task.Run. Only UI mutations (bot.Text += chunk, scroll) are dispatched to the main thread via MainThread.InvokeOnMainThreadAsync. This avoids NetworkOnMainThreadException on Android and keeps the UI responsive during streaming.

---

Getting started

Prerequisites

• .NET 10 SDK with MAUI workload (dotnet workload install maui)
• A Groq API key (free tier available)
• Python 3.x — only if you want to (re)build the knowledge base

1. Clone

git clone https://github.com/softal55/RAG-CSharpAssistant.git
cd RAG-CSharpAssistant

2. Add your Groq API key

The app checks three locations in order — first non-empty wins:

Method	How
Environment variable	export GROQ_API_KEY=gsk_...
Raw asset file	Create Resources/Raw/groq.key containing just the key
App preferences	Set GroqApiKey in Preferences.Default at runtime

⚠️ groq.key and Resources/Raw/groq.key are in .gitignore. Never commit your key.

3. (Optional) Build the knowledge base

If you have the Stack Overflow C# dataset:

cd DataPipeline
pip install -r requirements.txt

# Place your dataset here:
# DataPipeline/stack_overflow_c#_data.jsonl

python build_db.py
# → produces csharp_knowledge.db (~100–150 MB)

Then rebuild the app so the new DB is bundled as a MAUI asset. Without a DB the app still works — RAG simply returns a "not ready" message and the LLM answers from general knowledge.

4. Run

# Android
dotnet build -t:Run -f net10.0-android

# Windows
dotnet build -t:Run -f net10.0-windows10.0.19041.0

# iOS / Mac Catalyst (macOS host only)
dotnet build -t:Run -f net10.0-ios
dotnet build -t:Run -f net10.0-maccatalyst

---

Data pipeline details

build_db.py converts raw Stack Overflow JSONL into a searchable SQLite FTS5 database.

stack_overflow_c#_data.jsonl
        │
        ▼
   Parse JSON lines
        │
        ▼
   clean_html_preserve_code()     ← BeautifulSoup, [CODE]...[/CODE] markers
        │
        ▼
   smart_chunk_text()             ← ~800 char chunks, respects code block boundaries
        │
        ▼
   MD5 deduplication              ← ~150,000 chunk cap
        │
        ▼
   INSERT into FTS5 qa_index      ← chunk, source_question, tags
   INSERT into metadata           ← rowid, score, is_accepted
        │
        ▼
   csharp_knowledge.db

Accepted answers are labelled [ACCEPTED SOLUTION] and ranked first at query time via ORDER BY metadata.is_accepted DESC.

---

Project structure

RAG-CSharpAssistant/
├── App.xaml(.cs)               # Application + merged resource dictionaries
├── MauiProgram.cs              # DI registration, fonts, builder
├── AppShell.xaml(.cs)          # Shell → MainPage route
├── MainPage.xaml(.cs)          # Chat UI, empty/chat toggle, bubble layout
├── LoadingPage.xaml(.cs)       # Splash warmup → AppShell swap
│
├── ViewModels/
│   └── MainPageViewModel.cs    # Observable state, streaming pipeline
│
├── Models/
│   └── ChatMessage.cs          # IsUser/IsBot, Text (INPC for streaming)
│
├── Services/
│   ├── IRagSearchService.cs
│   ├── RagSearchService.cs     # DB provisioning, FTS5 search
│   ├── IGroqClient.cs
│   ├── GroqClient.cs           # Scope check, SSE streaming
│   ├── PromptBuilder.cs        # All prompt strings and builders
│   └── AppSecrets.cs           # API key resolution (env → file → prefs)
│
├── DataPipeline/
│   ├── build_db.py             # Offline Python pipeline
│   ├── requirements.txt
│   └── README.md
│
├── Platforms/
│   ├── Android/                # MainActivity, styles.xml (no white flash)
│   ├── iOS/
│   ├── MacCatalyst/
│   └── Windows/
│
└── Resources/
    ├── Styles/                 # Colors.xaml, Styles.xaml
    ├── Raw/                    # groq.key (gitignored), csharp_knowledge.db
    ├── Fonts/                  # OpenSans Regular + Semibold
    └── Images/                 # csharp_logo.png, csharp_hex.png, etc.

---

Platform support

Platform	Target framework	Min OS
Android	net10.0-android	API 21 (Android 5.0)
iOS	net10.0-ios	iOS 15.0
macOS	net10.0-maccatalyst	macOS 15.0
Windows	net10.0-windows10.0.19041.0	Windows 10 build 17763

Windows uses unpackaged-style deployment (WindowsPackageType = None).

---

Roadmap

• Expand knowledge base with more C# / .NET topics
• Add a debugging assistant mode (paste stack trace → get explanation)
• In-app settings panel to swap API key without restarting
• Fine-tuned domain model for tighter scope classification
• Multi-language UI support
• Automated pipeline for keeping the DB up to date

---

Author

Sofiane Taleb — AI Student, University of Oran 1

GitHub · LinkedIn

---

_{Built with .NET MAUI · Powered by Groq · Knowledge from Stack Overflow}