AI & LLM Optimization

1. Structured Data in Context

When providing tabular data (users, products, logs) as context to an LLM:

{
  "users": [
    {"id": 1, "name": "Alice", "email": "alice@example.com"},
    {"id": 2, "name": "Bob", "email": "bob@example.com"}
  ]
}

!def User id name email
---
users [
  !use User
  1 Alice "alice@example.com"
  2 Bob "bob@example.com"
]

💡 Pro Tip: Use the playground to paste your JSON and see instant token savings.

2. Few-Shot Examples

When providing training examples for in-context learning:

You are a sentiment classifier. Here are examples:

!def Example text sentiment confidence
"Great product!" positive 0.95
"Terrible service" negative 0.92
"It's okay" neutral 0.78

Now classify: "Amazing experience!"

Instead of verbose JSON objects, schemas make examples compact and the pattern clear.

3. Tool Calling / Function Results

When returning database query results or API responses to the LLM:

// Database query result
!def Order id customer_id total status created_at
12345 789 1299.99 shipped "2025-01-15"
12346 790 49.99 pending "2025-01-16"
12347 789 299.99 delivered "2025-01-14"

// The LLM can easily parse this and generate insights

4. RAG (Retrieval-Augmented Generation)

When embedding retrieved documents in the context:

!def Doc id title snippet relevance
101 "Tauq Basics" "Tauq is a token-efficient..." 0.94
205 "Schema Patterns" "Use !def to define schemas..." 0.89
387 "Streaming API" "Process large files with..." 0.76

Fit more retrieved documents in your context window = better answers.

5. Multi-Turn Conversations with History

Maintain conversation context efficiently:

!def Message role content timestamp
user "What's the weather?" "2025-01-16 10:00"
assistant "It's sunny, 72°F" "2025-01-16 10:01"
user "Perfect for a run!" "2025-01-16 10:02"

Keep longer conversation histories within your token budget.

📊 Use Schemas for Tabular Data

Arrays of objects with consistent fields are Tauq's sweet spot. Define a schema once, reap 60%+ token savings.

🗜️ Minify for Production

Use tauq minify or the minify button in the playground to compress to a single line:

tauq minify data.tqn -o data.min.tqn

🔄 Convert at Runtime

Use language bindings to convert your app's data structures on-the-fly:

import tauq

# Fetch from database
users = db.query("SELECT id, name, email FROM users LIMIT 100")

# Convert to Tauq for LLM context
tauq_context = tauq.dumps(users)

# Send to LLM
response = openai.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "user", "content": f"Analyze these users:\\n{tauq_context}"}
    ]
)

📈 Monitor Token Usage

Track before/after metrics to quantify savings. Most LLM APIs return token counts in response metadata.