Pinecone for Production RAG at Scale

Why Pinecone?

• Fully managed (no ops)
• Scales to billions of vectors
• 50ms p95 latency
• Built-in hybrid search
• SOC 2 compliant

Setup (November 2025)

DEVELOPERpython
from pinecone import Pinecone

pc = Pinecone(api_key="YOUR_API_KEY")

# Create index
pc.create_index(
    name="rag-production",
    dimension=1536,  # OpenAI text-embedding-3-small
    metric="cosine",
    spec=ServerlessSpec(
        cloud="aws",
        region="us-east-1"
    )
)

index = pc.Index("rag-production")

Upserting Documents

DEVELOPERpython
from openai import OpenAI

client = OpenAI()

def upsert_documents(documents):
    vectors = []
    
    for i, doc in enumerate(documents):
        # Generate embedding
        embedding = client.embeddings.create(
            model="text-embedding-3-small",
            input=doc['text']
        ).data[0].embedding
        
        vectors.append({
            "id": f"doc_{i}",
            "values": embedding,
            "metadata": {
                "text": doc['text'],
                "source": doc['source'],
                "date": doc['date']
            }
        })
    
    # Batch upsert (max 100 per batch)
    for i in range(0, len(vectors), 100):
        batch = vectors[i:i+100]
        index.upsert(vectors=batch)

Querying

DEVELOPERpython
def search(query, top_k=10):
    # Embed query
    query_embedding = client.embeddings.create(
        model="text-embedding-3-small",
        input=query
    ).data[0].embedding
    
    # Search
    results = index.query(
        vector=query_embedding,
        top_k=top_k,
        include_metadata=True
    )
    
    return [match['metadata']['text'] for match in results['matches']]

Metadata Filtering

DEVELOPERpython
# Filter by source
results = index.query(
    vector=query_embedding,
    filter={"source": {"$eq": "wikipedia"}},
    top_k=10,
    include_metadata=True
)

# Date range
results = index.query(
    vector=query_embedding,
    filter={
        "date": {"$gte": "2025-01-01", "$lte": "2025-12-31"}
    },
    top_k=10
)

Namespaces (Multi-tenancy)

DEVELOPERpython
# Separate customer data
index.upsert(
    vectors=[...],
    namespace="customer_123"
)

# Query specific namespace
results = index.query(
    vector=query_embedding,
    namespace="customer_123",
    top_k=10
)

Hybrid Search (Sparse + Dense)

DEVELOPERpython
# Upsert with sparse vectors
index.upsert(
    vectors=[{
        "id": "doc1",
        "values": dense_vector,  # Dense embedding
        "sparse_values": {
            "indices": [10, 45, 123],  # BM25 indices
            "values": [0.5, 0.3, 0.2]
        },
        "metadata": {"text": "..."}
    }]
)

# Hybrid query
results = index.query(
    vector=dense_query,
    sparse_vector={
        "indices": [10, 45],
        "values": [0.6, 0.4]
    },
    top_k=10,
    alpha=0.7  # Dense weight
)

Cost Optimization

Serverless pricing (Nov 2025):

• $0.09 per million read units
• $2.00 per million write units
• $0.00025 per GB-hour storage

Tips:

1. Use serverless for variable load
2. Batch upserts (100 per request)
3. Cache frequent queries
4. Delete old data

Monitoring

DEVELOPERpython
# Index stats
stats = index.describe_index_stats()
print(f"Total vectors: {stats['total_vector_count']}")
print(f"Dimension: {stats['dimension']}")

Pinecone is production-grade. Use it when you need scale, reliability, and zero ops.