🚀 RAG-TUI v0.0.3 Beta

The Terminal App That Makes Chunking Actually Fun

"I used to stare at my RAG pipeline wondering why it sucked. Then I found RAG-TUI and realized my chunks were the size of War and Peace." - A Developer, Probably

---

🎭 What Even Is This?

RAG-TUI is a beautiful terminal-based debugger for your Retrieval-Augmented Generation (RAG) pipelines. It's like having X-ray vision for your text chunking.

You know that feeling when your LLM hallucinates because your retrieval returned garbage? Yeah, this fixes that.

┌─────────────────────────────────────────────────────────────────┐
│  RAG-TUI v0.0.3 Beta                                            │
├─────────────────────────────────────────────────────────────────┤
│  📝 Input   🎨 Chunks   🔍 Search   📊 Batch   ⚙️ Settings      │
│                                                                 │
│  Your text, but now with ✨ colors ✨ and 📊 metrics 📊        │
│                                                                 │
│  Chunk Size: [◀] [200] [▶]  ████████░░░░░░  tokens              │
│  Overlap:    [◀] [10]  [▶]  ██░░░░░░░░░░░░  %                   │
│                                                                 │
│  ⚡ 5 chunks                                                    │
└─────────────────────────────────────────────────────────────────┘

📸 Screenshots

Input Tab	Chunks Tab

Search Tab	Chat Tab

---

🤔 Why Should I Care?

The Problem

You're building a RAG app. You chunk your documents. You embed them. You search. And then...

User: "What's the company's refund policy?"
LLM: "Based on the context, your refrigerator appears to be running."

The Solution

See exactly how your text is being chunked. Tweak parameters in real-time. Test queries. Export settings. Actually understand what's going on.

---

⚡ Quick Start (30 Seconds, I Promise)

Install

pip install rag-tui

Run

rag-tui

That's It

No really. You're done. Press L to load sample text and start playing.

---

🎨 Features (The Good Stuff)

1. 🧩 Six Chunking Strategies

Because one size definitely does NOT fit all.

Strategy	Best For	Vibe
Token	General text	"I count tokens for breakfast"
Sentence	Articles, docs	"Periods are sacred"
Paragraph	Structured text	"Double newline gang"
Recursive	Code, mixed	"I'll try everything"
Fixed	Speed demons	"Just cut every 500 chars lol"
Custom	You, apparently	"I know better" (you might!)

2. 🔌 Four LLM Providers

Switch between providers like you switch between tabs (too often).

# Ollama (Free! Local! Private!)
ollama serve
rag-tui

# OpenAI (When you need that GPT juice)
export OPENAI_API_KEY="sk-..." 
rag-tui

# Groq (FAST. LIKE, REALLY FAST.)
export GROQ_API_KEY="gsk_..."
rag-tui

# Google Gemini (Free tier FTW)
export GOOGLE_API_KEY="AI..."
rag-tui

3. 📁 Load Any File

PDFs? ✅ Markdown? ✅ Python? ✅ That random .txt file from 2019? ✅

Supported: .txt, .md, .py, .js, .json, .yaml, .pdf, and 10 more!

4. 📊 Batch Testing

Test 50 queries at once. See which ones fail. Cry. Fix. Repeat.

📊 Batch Test Results
━━━━━━━━━━━━━━━━━━━━
Total Queries: 50
Hit Rate (>0.5): 78%
Avg Top Score: 0.72

You're doing better than average!
(The average is made up, but still, congrats!)

5. ⚡ Built-in Presets

Don't know what settings to use? We got you.

Preset	Size	Overlap	For
Q&A Retrieval	200	15%	Chatbots, search
Document Summary	500	5%	Long docs
Code Analysis	300	20%	Source code
Long Context	800	10%	GPT-4-128k users
High Precision	100	25%	When you NEED accuracy

6. 📋 Export Settings

Take your carefully tuned settings and use them in production.

# LangChain Export
from langchain.text_splitter import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(
    chunk_size=800,
    chunk_overlap=80,
)

# LlamaIndex Export  
from llama_index.core.node_parser import SentenceSplitter

parser = SentenceSplitter(
    chunk_size=800,
    chunk_overlap=80,
)

---

🎮 How To Use It

The Interface

┌─────────────────────────────────────────────────────────────────┐
│  Strategy: [Token ▼]  │  File: [path...]  │  [📂 Load]         │
├─────────────────────────────────────────────────────────────────┤
│  📝 Input │ 🎨 Chunks │ 🔍 Search │ 📊 Batch │ ⚙️ Settings │ 💬 │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│                    (Your content here)                          │
│                                                                 │
├─────────────────────────────────────────────────────────────────┤
│  Chunk Size: [◀] [200] [▶]  │  Overlap: [◀] [10] [▶] %          │
│  ⚡ 5 chunks                                                    │
└─────────────────────────────────────────────────────────────────┘

The Tabs

Tab	What It Does	When To Use It
📝 Input	Paste or load your document	First
🎨 Chunks	See colorful chunk cards	To see the magic
🔍 Search	Query and see what comes back	Testing retrieval
📊 Batch	Test many queries at once	Before production
⚙️ Settings	Export config, custom code	When you're done
💬 Chat	Talk to your chunks	For fun

Keyboard Shortcuts

Key	Action	Pro Tip
Q	Quit	When you're done procrastinating
L	Load sample	Start here if confused
R	Rechunk	After changing params
D	Dark/Light mode	We default to dark (obviously)
E	Export config	Save your precious settings
F1	Help	When this README isn't enough
Tab	Next tab	Navigate like a pro

---

🔧 The Workflow (AKA What To Actually Do)

Step 1: Load Your Document

Either:

• Type/paste in the 📝 Input tab
• Enter a file path and click 📂 Load
• Press L for sample text (recommended for newbies)

Step 2: Pick Your Strategy

Use the dropdown at the top. If unsure:

• Text/articles? → Sentence
• Code? → Recursive
• Don't know? → Token (it's the safe choice)

Step 3: Adjust Parameters

In the 🎨 Chunks tab, use the sliders:

• Chunk Size: How big each chunk should be (in tokens)
• Overlap: How much chunks should share (prevents context loss)

The Golden Rule: Smaller chunks = more precise, less context. Bigger chunks = more context, less precise.

Step 4: Test Your Queries

Go to 🔍 Search tab:

1. Type a question
2. Click Search
3. See what chunks come back
4. Cry or celebrate accordingly

Step 5: Batch Test (The Pro Move)

Go to 📊 Batch tab:

1. Enter multiple queries (one per line)
2. Click Run Batch Test
3. See your hit rate
4. Adjust until it's good enough™

Step 6: Export

Go to ⚙️ Settings tab:

• Click JSON, LangChain, or LlamaIndex
• Copy the generated code
• Paste in your project
• Deploy
• Profit???

---

🧪 Custom Chunking (For The Brave)

Don't like our strategies? Roll your own!

Go to ⚙️ Settings tab, paste a function like:

def chunk_by_headers(text, chunk_size, overlap):
    """Split on markdown headers."""
    import re
    sections = re.split(r'\n(?=#{1,3} )', text)
    return [(s, 0, len(s)) for s in sections if s.strip()]

Click ⚡ Apply Custom Chunker and watch the magic.

---

🤖 Provider Setup

Ollama (Recommended for Privacy)

# Install Ollama
brew install ollama  # macOS
# or download from ollama.ai

# Pull required models
ollama pull nomic-embed-text  # For embeddings
ollama pull llama3.2:1b       # For chat (small & fast)

# Start the server
ollama serve

OpenAI

export OPENAI_API_KEY="sk-your-key-here"

Uses: text-embedding-3-small + gpt-4o-mini

Groq (Free Tier!)

export GROQ_API_KEY="gsk_your-key-here"

Uses: llama-3.1-8b-instant (NO embeddings - pair with Ollama)

Google Gemini (Also Free Tier!)

export GOOGLE_API_KEY="your-key-here"

Uses: text-embedding-004 + gemini-1.5-flash

---

📈 Understanding The Metrics

In Chunks Tab

📊 5 chunks | Avg: 180 chars | Total: 900 chars | ~225 tokens

• 5 chunks: Your document was split into 5 pieces
• Avg: 180 chars: Each chunk is ~180 characters
• Total: 900 chars: Your whole document size
• ~225 tokens: Estimated token count (chars ÷ 4)

In Search Tab

#1 ████████░░ 0.89  "The refund policy states..."
#2 ██████░░░░ 0.72  "For returns within 30 days..."
#3 ████░░░░░░ 0.45  "Our customer service team..."

• #1, #2, #3: Ranking by relevance
• ████████░░: Visual similarity bar
• 0.89: Cosine similarity score (0-1, higher = better)

In Batch Tab

Hit Rate (>0.5): 78%
Avg Top Score: 0.72

• Hit Rate: % of queries where top result scored > 0.5
• Avg Top Score: Average of all top-1 scores

As a rule of thumb:

• Hit Rate > 80% = Great
• Hit Rate 60-80% = Acceptable
• Hit Rate < 60% = Time to tune

---

🐛 Troubleshooting

"Ollama not available"

# Is Ollama running?
ollama serve

# Did you pull the models?
ollama pull nomic-embed-text
ollama pull llama3.2:1b

"No chunks"

• Is your text too short?
• Is chunk size bigger than your text?
• Try lowering chunk size to 50

"Search returns garbage"

• Check if embeddings were created (needs Ollama/OpenAI)
• Try a different chunking strategy
• Lower chunk size for more precision

"App looks weird"

# Reset terminal
reset

# Try a different terminal (iTerm2, Warp, etc.)

---

🎓 Chunking 101 (The Theory)

Why Chunk At All?

LLMs have context limits. Your document is bigger than the limit. So we split it up, find the relevant parts, and only send those.

Your 50-page PDF → Split into 100 chunks → Search → Top 3 sent to LLM → Answer!

The Size-Precision Tradeoff

Chunk Size	Precision	Context	Best For
Small (50-100)	High ✅	Low ❌	Specific facts
Medium (200-400)	Medium	Medium	General Q&A
Large (500-1000)	Low ❌	High ✅	Summaries

The Overlap Question

Overlap = how many tokens chunks share at boundaries.

• 0% overlap: Chunks are completely separate (risk: losing context at boundaries)
• 10-20% overlap: Goldilocks zone (recommended)
• 50% overlap: Lots of redundancy (wastes tokens but very safe)

---

📦 Programmatic Usage

Don't want the TUI? Use the library directly:

from rag_tui.core import ChunkingEngine, StrategyType

# Create engine
engine = ChunkingEngine()
engine.set_strategy(StrategyType.SENTENCE)

# Chunk some text
chunks = engine.chunk_text(
    "Your document here...",
    chunk_size=200,
    overlap=20
)

for text, start, end in chunks:
    print(f"[{start}:{end}] {text[:50]}...")

Use Providers Directly

from rag_tui.core.providers import get_provider, ProviderType

# Get Ollama provider
provider = get_provider(ProviderType.OLLAMA)

# Check connection
if await provider.check_connection():
    # Embed text
    embedding = await provider.embed("Hello world")
    
    # Generate response
    response = await provider.generate("What is RAG?")

---

🤝 Contributing

Found a bug? Have an idea? Want to add support for Claude/Anthropic?

1. Fork the repo
2. Create a branch
3. Make your changes
4. Submit a PR
5. Get famous (in our small community)

---

📜 License

MIT License - Do whatever you want, just don't blame us if your RAG app becomes sentient.

---

🙏 Credits

Built with:

• Textual - The TUI framework that makes terminals beautiful
• Chonkie - Token-based chunking
• Usearch - Blazing fast vector search
• Ollama - Local LLM inference

---

💭 Final Words

RAG is hard. Chunking is an art. But with RAG-TUI, at least you can see what you're doing wrong.

Now go forth and chunk responsibly! 🎯

---

Made with ❤️ and too much ☕ for RAG developers everywhere

"May your chunks be small and your retrieval be accurate."