SudoAll.com

A vector database stores data as high-dimensional numeric arrays, embeddings, instead of rows and columns. When you search, you don’t match keywords; you find items that are semantically close in that space. Two sentences can share zero words and still be nearest neighbours if they mean the same thing. That’s why vector databases are the backbone of RAG (Retrieval-Augmented Generation): you embed a user’s question, retrieve the most relevant document chunks from the database, and hand those chunks to the LLM as context. The LLM answers from evidence rather than from weights alone, grounded, citable, and up to date with your own data.

How embeddings work

An embedding model (like OpenAI’s text-embedding-3-small) reads a piece of text and outputs a list of floats, typically 1,536 of them. Each float encodes some aspect of meaning learned during training. The distance between two vectors (cosine similarity is standard) measures semantic closeness: 1.0 means identical in meaning, 0 means unrelated, negative means opposite. The magic is that this geometry is compositional: “Paris” minus “France” plus “Italy” lands near “Rome”. You don’t program these relationships, they emerge from training on billions of documents.

The full pipeline

The pipeline has two phases: ingestion (offline, runs once or on update) and retrieval (online, runs per query). During ingestion you chunk your documents, embed each chunk, and upsert into Pinecone. During retrieval you embed the query, call Pinecone’s query endpoint, and get back the top-K most similar chunks. Those chunks become the context you inject into the LLM prompt.

Step 1: Setup

npm install @pinecone-database/pinecone openai

Create a free Pinecone index at pinecone.io. Set dimension to 1536 (matches text-embedding-3-small) and metric to cosine. Store both keys in .env:

PINECONE_API_KEY=your-pinecone-key
OPENAI_API_KEY=your-openai-key

// client.js
import { Pinecone } from '@pinecone-database/pinecone';
import OpenAI from 'openai';

export const pinecone = new Pinecone({ apiKey: process.env.PINECONE_API_KEY });
export const openai   = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
export const index    = pinecone.index('documents');

Step 2: Chunk your documents

Embedding models have token limits (8,191 for text-embedding-3-small). More importantly, a chunk that’s too large drowns the signal, a paragraph about your answer might be buried inside a 5,000-word chunk. Aim for 200–500 words per chunk with a small overlap so context doesn’t get cut at chunk boundaries.

// chunker.js

/**
 * Split text into overlapping chunks.
 * @param {string} text        - Full document text
 * @param {number} chunkSize   - Target chars per chunk (default 1200 ≈ 300 words)
 * @param {number} overlap     - Overlap chars between chunks (default 200)
 * @returns {string[]}
 */
export function chunkText(text, chunkSize = 1200, overlap = 200) {
  const chunks = [];
  let start = 0;

  while (start < text.length) {
    const end = Math.min(start + chunkSize, text.length);
    chunks.push(text.slice(start, end).trim());
    if (end === text.length) break;
    start += chunkSize - overlap; // step back by overlap amount
  }

  return chunks.filter(c => c.length > 50); // drop tiny trailing chunks
}

Step 3: Embed in batches and upsert

OpenAI’s embeddings endpoint accepts up to 2,048 inputs per call. Batching is faster and cheaper than one call per chunk. After embedding, upsert to Pinecone with a stable ID and metadata. The metadata is what comes back in query results, store everything you’ll need to display or cite the source.

// ingest.js
import { openai, index } from './client.js';
import { chunkText }     from './chunker.js';

const EMBED_MODEL  = 'text-embedding-3-small';
const BATCH_SIZE   = 100; // chunks per OpenAI call

/**
 * Embed an array of strings in batches.
 */
async function embedBatch(texts) {
  const res = await openai.embeddings.create({
    model: EMBED_MODEL,
    input: texts,
  });
  return res.data.map(d => d.embedding); // array of float[]
}

/**
 * Ingest a document into Pinecone.
 * @param {string} docId    - Stable identifier for this document
 * @param {string} text     - Full document text
 * @param {object} meta     - Extra metadata (title, url, date, etc.)
 */
export async function ingestDocument(docId, text, meta = {}) {
  const chunks  = chunkText(text);
  const vectors = [];

  // Embed in batches of BATCH_SIZE
  for (let i = 0; i < chunks.length; i += BATCH_SIZE) {
    const batch      = chunks.slice(i, i + BATCH_SIZE);
    const embeddings = await embedBatch(batch);

    embeddings.forEach((embedding, j) => {
      const chunkIndex = i + j;
      vectors.push({
        id: `${docId}__chunk_${chunkIndex}`,
        values: embedding,
        metadata: {
          ...meta,
          docId,
          chunkIndex,
          text: chunks[chunkIndex], // store chunk text for retrieval
        },
      });
    });
  }

  // Upsert to Pinecone (max 100 vectors per call)
  for (let i = 0; i < vectors.length; i += 100) {
    await index.upsert(vectors.slice(i, i + 100));
  }

  console.log(`Ingested ${vectors.length} chunks for doc: ${docId}`);
}

// --- Usage example ---
await ingestDocument(
  'nodejs-docs-v20',
  await fs.readFile('nodejs-docs.txt', 'utf8'),
  { title: 'Node.js v20 Docs', url: 'https://nodejs.org/docs/v20/' }
);

Step 4: Query and retrieve

At query time: embed the user's question, search Pinecone for the top-K nearest chunks, and return them. Pinecone also supports metadata filtering so you can narrow results to a specific document, date range, or tag, without re-embedding.

// retrieve.js
import { openai, index } from './client.js';

const EMBED_MODEL = 'text-embedding-3-small';

/**
 * Retrieve the top-K most relevant chunks for a query.
 * @param {string} query    - User's natural language question
 * @param {number} topK     - Number of results (default 5)
 * @param {object} filter   - Optional Pinecone metadata filter
 * @returns {Array<{text, score, meta}>}
 */
export async function retrieve(query, topK = 5, filter = {}) {
  // 1. Embed the query
  const res = await openai.embeddings.create({
    model: EMBED_MODEL,
    input: query,
  });
  const queryVector = res.data[0].embedding;

  // 2. Search Pinecone
  const results = await index.query({
    vector:          queryVector,
    topK,
    includeMetadata: true,
    filter:          Object.keys(filter).length ? filter : undefined,
  });

  // 3. Return clean objects
  return results.matches.map(m => ({
    text:  m.metadata.text,
    score: m.score,         // cosine similarity 0–1
    meta:  m.metadata,
  }));
}

// --- With metadata filter (only search a specific doc) ---
const chunks = await retrieve(
  'How do I use async iterators?',
  5,
  { docId: { $eq: 'nodejs-docs-v20' } }
);

Step 5: Wire it into a RAG answer

Now put it all together: retrieve relevant chunks, build a prompt, call the LLM. The key is to pass the chunks as explicit context and instruct the model to answer only from them, this is what gives you grounded, citable responses instead of hallucinations.

// rag.js
import { openai }  from './client.js';
import { retrieve } from './retrieve.js';

/**
 * Answer a question using RAG over your Pinecone index.
 * @param {string} question
 * @param {object} filter    - Optional metadata filter
 * @returns {string}         - LLM answer
 */
export async function ragAnswer(question, filter = {}) {
  // 1. Retrieve relevant chunks
  const chunks = await retrieve(question, 5, filter);

  if (chunks.length === 0) {
    return "I couldn't find relevant information in the knowledge base.";
  }

  // 2. Build context block
  const context = chunks
    .map((c, i) => `[${i + 1}] (score: ${c.score.toFixed(3)})n${c.text}`)
    .join('nn---nn');

  // 3. Build prompt
  const systemPrompt = `You are a helpful assistant. Answer the user's question
using ONLY the context provided below. If the answer is not in the context,
say so. Cite the chunk number [1], [2], etc. where relevant.

CONTEXT:
${context}`;

  // 4. Call the LLM
  const completion = await openai.chat.completions.create({
    model:    'gpt-4o-mini',
    messages: [
      { role: 'system',  content: systemPrompt },
      { role: 'user',    content: question },
    ],
    temperature: 0.2, // low temperature = more faithful to context
  });

  return completion.choices[0].message.content;
}

// --- Example ---
const answer = await ragAnswer(
  'What changed in the streams API in Node.js v20?',
  { docId: { $eq: 'nodejs-docs-v20' } }
);
console.log(answer);

Practical tips

• Delete old chunks when you update a document, use a consistent naming scheme like docId__chunk_N so you can delete all chunks for a doc before re-ingesting: index.deleteMany({ docId: { $eq: docId } }) (requires a paid Pinecone plan for filter-based delete; alternatively delete by ID prefix).
• Score thresholds, don't blindly pass all top-K to the LLM. Filter out chunks with score below ~0.75; they're probably noise. Low scores mean your query is outside what the index knows.
• Namespace isolation, use Pinecone namespaces to keep multi-tenant data separate: index.namespace('tenant-abc').upsert(...). Free plans have no namespace limit.
• Hybrid search, Pinecone's sparse-dense hybrid mode lets you combine BM25 keyword matching with vector similarity. Useful when exact terms matter (product codes, names) alongside semantic meaning.
• Cost, embedding 1M tokens with text-embedding-3-small costs $0.02. A typical 1,000-page knowledge base (~500K words) costs under $2 to embed. Re-embedding only changed documents keeps costs minimal.

nJoy 😉