LLM Patterns

LLM Patterns Skill

Load this skill when working with LLMs, RAG systems, embeddings, or AI integrations.

---

Rule

LLM INTEGRATIONS MUST BE ROBUST AND COST-EFFECTIVE!

• Always handle streaming and errors
• Optimize tokens for cost
• Use appropriate models for tasks

---

Model Selection Guide

Task	Model	Why
Complex reasoning	Claude Opus 4.5 / GPT-4	Best quality
General tasks	Claude Sonnet 4.5 / GPT-4o	Balance
Simple tasks	Claude Haiku 4.5 / GPT-4o-mini	Fast & cheap
Embeddings	text-embedding-3-small	Cost-effective
Classification	Fine-tuned small model	Fastest

---

RAG Architecture

Basic RAG Flow

Query → Embed Query → Vector Search → Retrieve Chunks → Augment Prompt → LLM → Response

Implementation

async function ragQuery(query: string): Promise<string> {
  // 1. Embed the query
  const queryEmbedding = await embedText(query);

  // 2. Vector search for relevant chunks
  const chunks = await vectorStore.similaritySearch(queryEmbedding, {
    topK: 5,
    minScore: 0.7,
  });

  // 3. Build context from chunks
  const context = chunks.map((c) => c.content).join('\n\n---\n\n');

  // 4. Augment prompt
  const prompt = `Answer based on the following context:

Context:
${context}

Question: ${query}

Answer:`;

  // 5. Get LLM response
  return await llm.complete(prompt);
}

---

Chunking Strategies

Size Guidelines

Document Type	Chunk Size	Overlap
Dense text (legal, technical)	256-512 tokens	50 tokens
General content	512-1024 tokens	100 tokens
Conversational	1024-2048 tokens	200 tokens

Chunking Methods

// 1. Fixed size chunking
function fixedChunks(text: string, size: number, overlap: number): string[] {
  const chunks = [];
  for (let i = 0; i < text.length; i += size - overlap) {
    chunks.push(text.slice(i, i + size));
  }
  return chunks;
}

// 2. Semantic chunking (by paragraph/section)
function semanticChunks(text: string): string[] {
  return text
    .split(/\n\n+/)
    .filter((chunk) => chunk.trim().length > 50);
}

// 3. Recursive character splitting (LangChain style)
const splitter = new RecursiveCharacterTextSplitter({
  chunkSize: 1000,
  chunkOverlap: 200,
  separators: ['\n\n', '\n', '. ', ' ', ''],
});