Pinecone

Create Index on Pinecone Console

Sign up for an account and create a new index with the correct vector dimension suitable for the embedding model you will use.

Go to Pinecone Console > Create Index, name the index, select dimension = 1536 (if using OpenAI text-embedding-3-small) and metric = cosine.

Remember the exact dimension of the embedding model — incorrect dimension will cause data write failures.

Obtain API Key and Configure SDK

Retrieve the API key from the console and initialize the Pinecone client in your application code.

Install `pip install pinecone-client`, then initialize `pc = Pinecone(api_key="YOUR_KEY")` and `index = pc.Index("index-name")`.

Store the API key in an environment variable, do not hardcode it directly into the code.

Create Embedding for Product Data

Convert the description of each product into a vector using the embedding model before writing to Pinecone.

For each product, call the embedding API (e.g., OpenAI) to obtain the vector, along with metadata such as name, price, and category.

Combine the product name and detailed description into the same text block before creating the embedding to improve accuracy.

Write Data to Index in Batches

Upsert all product vectors into Pinecone in batches to avoid timeouts.

Call `index.upsert(vectors=[(id, vector, metadata), ...])` in batches of about 100 products at a time.

Use the real product ID (SKU) as the vector ID to make it easier to update when the product changes.

Query Semantic Search and Combine Filter

Build a search API that accepts customer query sentences, converts them into vectors, and queries Pinecone.

Call `index.query(vector=query_embedding, top_k=10, filter={"category": "furniture"}, include_metadata=True)`.

Combine filters based on metadata (category, price range) with semantic search to ensure results are both relevant and within the customer's desired scope.