Skip to content

QUICKSTART.md

Latest commit

 

History

History
220 lines (151 loc) · 4.62 KB

QUICKSTART.md

File metadata and controls

220 lines (151 loc) · 4.62 KB

SNS-Core Quick Start Guide

Get started with SNS-Core in 5 minutes!

Method 1: Use the SNS Converter (Easiest)

Step 1: Copy model.sns

Open model.sns and copy the entire contents.

Step 2: Give to Your LLM

Paste the model.sns content into any LLM (ChatGPT, Claude, Gemini, etc.).

Step 3: Convert Your Prompts

Ask the LLM:

Convert this prompt to SNS notation:

[Your natural language prompt here]

Step 4: Use the SNS Output

Copy the SNS notation and use it in your AI system!

Example

You provide:

Analyze the user query and extract keywords. 
Then classify the intent as question, complaint, or request.
Return structured results.

LLM converts to:

q → kw_extract → kw
q → classify(["question","complaint","request"]) → intent
→ {kw, intent}

Token savings: ~50 tokens → ~15 tokens (70% reduction)

Method 2: Learn SNS Manually

Step 1: Learn Core Patterns (5 minutes)

Read Core Patterns to understand:

  • Flow: a → b → c
  • Pipeline: a | b | c
  • Conditional: x ? y : z

Step 2: Try an Example (2 minutes)

Natural Language:

Filter the documents to keep only those with score above 0.7, 
then sort them by relevance, and return the top 5

Convert to SNS:

docs | filter(score > 0.7) | sort(relevance) | top(5)

Test it with your LLM—it works!

Step 3: Convert Your First Prompt (10 minutes)

Take an internal AI-to-AI prompt from your system and convert it:

  1. Identify operations (extract, classify, filter, etc.)
  2. Connect with arrows (→) or pipes (|)
  3. Use abbreviations (query → q, keywords → kw)
  4. Remove filler words ("please", "carefully", "then")
  5. Test with your LLM

Real-World Example

Your RAG Orchestrator

Before (Natural Language):

You are the orchestrator in a RAG pipeline. Analyze the user's query 
to extract keywords and classify the intent into one of these categories: 
informational, transactional, or navigational. Then expand the keywords 
into search terms and infer the relevant document categories. Finally, 
return a structured object containing the search terms, categories, 
intent, and keywords.

Token count: ~150 tokens
Cost (at 10K queries/day): ~$45/day

After (SNS):

q → kw_extract → kw
q → classify(["info","transactional","nav"]) → intent
(kw + q) → expand_q → terms
intent → infer_cats → cats
→ {terms, cats, intent, kw}

Token count: ~30 tokens
Cost (at 10K queries/day): ~$9/day
Savings: $36/day = $1,080/month 💰

Testing Your SNS

Quick Test

Give this to ChatGPT/Claude (no context needed):

q → kw_extract → kw
q → classify(["positive","negative","neutral"]) → sentiment
→ {kw, sentiment}

q = "I love this product!"

Expected output:

{
  "kw": ["love", "product"],
  "sentiment": "positive"
}

It works! 🎉

Common Patterns Cheat Sheet

# Extract and analyze
text → extract_keywords → keywords

# Classify
text → classify(categories) → category

# Pipeline
data | step1 | step2 | step3

# Conditional
score > 0.7 ? keep : discard

# Filter collection
[items] | filter(condition)

# Multiple operations
input → op1 → result1
input → op2 → result2
→ {result1, result2}

# Combine inputs
(input1 + input2) → process → output

# Check and branch
valid ? process : reject

Next Steps

Immediate (Today)

  • Convert 1 internal AI prompt to SNS
  • Test it with your LLM
  • Measure token savings

This Week

  • Convert all high-volume internal prompts
  • Calculate monthly cost savings
  • Share results with team

This Month

  • Consider training an SLM on SNS (see SLM Training Guide)
  • Build SNS into your deployment pipeline
  • Contribute your use case back to the community

Get Help

Tips for Success

  1. Start small: Convert one prompt first
  2. Use model.sns: Let the LLM do the conversion
  3. Verify quality: Compare outputs to natural language version
  4. Measure savings: Track token counts
  5. Share learnings: Contribute back to the community

Ready to save 60-85% on your AI costs? Let's go! 🚀

Repository: github.com/EsotericShadow/sns-core