Dev Tools

If you've been living under a rock, vector databases are now boring infrastructure. The interesting question isn't "which is fastest" — it's "which costs the least engineer-hours to keep alive."

The Setup

The four contenders: Pinecone serverless (managed, zero ops), Qdrant (open source, run yourself), LanceDB (embedded, single-file), and pgvector on Supabase Postgres (the boring choice). At Rebuild Relief we ship pgvector inside Supabase and it has not once been the bottleneck.

# Pinecone serverless — fully managed
from pinecone import Pinecone, ServerlessSpec

pc = Pinecone(api_key="pc-...")
pc.create_index(
    name="claims",
    dimension=1536,
    spec=ServerlessSpec(cloud="aws", region="us-east-1"),
)
idx = pc.Index("claims")
idx.upsert([{"id": "1", "values": [0.1]*1536, "metadata": {"postcode": "4870"}}])
idx.query(vector=[0.1]*1536, top_k=5, filter={"postcode": "4870"})

The Money Pattern

If you're already on Postgres — and most of us are — pgvector is free, transactional, and joins to the rest of your schema. No second database to back up. No second auth model. Aiden uses this exact pattern across every Rebuild Relief tool.

# Qdrant — self-hosted, blazing fast, but you run it
from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, PointStruct

qd = QdrantClient(url="http://localhost:6333")
qd.recreate_collection(
    "claims",
    vectors_config=VectorParams(size=1536, distance=Distance.COSINE),
)
qd.upsert("claims", points=[
    PointStruct(id=1, vector=[0.1]*1536, payload={"postcode": "4870"})
])

The Catch

Pinecone is genuinely zero-ops but you pay for it — at scale the bill creeps. Qdrant is fast but you're now running another stateful service. LanceDB is brilliant for embedded use cases but doesn't fit a multi-tenant SaaS. pgvector hits a wall around 50M vectors without HNSW tuning.

The Verdict

Solo dev or small team on Supabase? pgvector, every day. Mid-sized team that wants ops-free? Pinecone serverless wins on developer hours. Big team with infra people who like Helm charts? Qdrant. Pick by team size, not by benchmark — the QPS gap doesn't matter at the stage you're actually at.