Build Your Own RAG System — From Scratch

// Too small (< 100 tokens):
"Our company's vacation policy"
// Problem: No context, meaningless fragment

// Too large (> 1000 tokens):  
[Entire 5-page policy document]
// Problem: Too much noise, key info gets lost

// Just right (200-500 tokens):
"Our company's vacation policy is designed to promote 
work-life balance. All full-time employees earn 15 days 
of paid vacation annually, starting from their hire date.
Vacation days must be requested at least two weeks in 
advance through our HR system..."
// Perfect: Complete thought with context

// Sample embedding vectors (simplified to 3 dimensions)
const embeddings = {
  "dog": [0.2, 0.8, 0.1],
  "puppy": [0.3, 0.7, 0.2],      // Close to "dog"
  "cat": [0.1, 0.6, 0.3],        // Pet-related, but different
  "vacation": [-0.5, 0.2, 0.8],  // Completely different topic
  "policy": [-0.3, 0.1, 0.9]     // Work-related
};

function similarity(vec1, vec2) {
  // Cosine similarity: how "parallel" are two vectors?
  let dotProduct = 0;
  let mag1 = 0, mag2 = 0;
  
  for (let i = 0; i < vec1.length; i++) {
    dotProduct += vec1[i] * vec2[i];
    mag1 += vec1[i] * vec1[i];
    mag2 += vec2[i] * vec2[i];
  }
  
  return dotProduct / (Math.sqrt(mag1) * Math.sqrt(mag2));
}

// "dog" vs "puppy" = 0.97 (very similar!)
// "dog" vs "vacation" = -0.23 (very different)

// Using OpenAI's embedding API:
async function getEmbedding(text) {
  const response = await fetch('https://api.openai.com/v1/embeddings', {
    method: 'POST',
    headers: {
      'Content-Type': 'application/json',
      'Authorization': `Bearer ${API_KEY}`
    },
    body: JSON.stringify({
      model: 'text-embedding-3-small',
      input: text
    })
  });
  
  const data = await response.json();
  return data.data[0].embedding; // Array of 1536 numbers
}

// Convert all chunks to embeddings
const chunkEmbeddings = [];
for (const chunk of chunks) {
  const embedding = await getEmbedding(chunk.text);
  chunkEmbeddings.push({
    text: chunk.text,
    embedding: embedding
  });
}

async function semanticSearch(query, chunkEmbeddings, topK = 3) {
  // 1. Embed the user's question
  const queryEmbedding = await getEmbedding(query);
  
  // 2. Calculate similarity to every chunk
  const scores = chunkEmbeddings.map(chunk => ({
    ...chunk,
    similarity: cosineSimilarity(queryEmbedding, chunk.embedding)
  }));
  
  // 3. Sort by similarity (highest first)
  scores.sort((a, b) => b.similarity - a.similarity);
  
  // 4. Return top-K most relevant chunks
  return scores.slice(0, topK);
}

function cosineSimilarity(vec1, vec2) {
  const dotProduct = vec1.reduce((sum, a, i) => sum + a * vec2[i], 0);
  const mag1 = Math.sqrt(vec1.reduce((sum, a) => sum + a * a, 0));
  const mag2 = Math.sqrt(vec2.reduce((sum, a) => sum + a * a, 0));
  return dotProduct / (mag1 * mag2);
}

// Vector database pseudocode:
const db = new VectorDB();

// Store embeddings (do this once)
await db.upsert(chunkEmbeddings);

// Search (do this for every query)
const results = await db.query({
  vector: queryEmbedding,
  topK: 5,
  filter: { source: "company-policies" } // Optional filtering
});

// Results come back ranked by similarity

async function ragQuery(question, documents) {
  // 1. Chunk the documents
  const chunks = chunkDocuments(documents);
  
  // 2. Create embeddings for all chunks  
  const chunkEmbeddings = [];
  for (const chunk of chunks) {
    const embedding = await getEmbedding(chunk);
    chunkEmbeddings.push({ text: chunk, embedding });
  }
  
  // 3. Embed the user's question
  const questionEmbedding = await getEmbedding(question);
  
  // 4. Find most relevant chunks
  const relevantChunks = semanticSearch(
    question, 
    chunkEmbeddings, 
    topK = 3
  );
  
  // 5. Build the prompt
  const context = relevantChunks.map(chunk => chunk.text).join('\n\n');
  const prompt = `
    Context: ${context}
    
    Question: ${question}
    
    Answer based only on the context above:`;
  
  // 6. Get the answer from LLM
  const response = await llm.complete(prompt);
  return response;
}

// Usage:
const answer = await ragQuery(
  "How many vacation days do I get?", 
  [vacationPolicyDoc, hrHandbookDoc]
);