RAG & Vector Databases · Lesson 2

Vector Stores: Pinecone, Qdrant, pgvector, Chroma, FAISS

Comparing managed and self-hosted stores, HNSW and IVF index types, running Qdrant in Docker and pgvector in Postgres, when to choose what.

30 min read4 questions in quizReady prompt includedIn progress

Practical exercise

What to do after this lesson

Run Qdrant in Docker and pgvector locally. Load 1000 vectors into each, measure top-5 query latency and compare. Create an HNSW index in pgvector and re-measure — how much faster did it get?

Ready-to-use prompt

Template for this lesson

Copy and adapt to your context. Text in angle brackets should be replaced.

Help me choose a vector store.
Volume: <number of vectors>, dimension: <…>
Infrastructure: <do you have Postgres / k8s / managed-only>
Requirements: <privacy, filtering, budget, latency>

Compare Pinecone / Qdrant / pgvector / Chroma along these axes and give a recommendation with index type (HNSW/IVF) and parameters.

Common mistakes

What people get wrong

Using FAISS as a "database" and being surprised it has no filters or networking — it's an index library.
Forgetting to create an HNSW index in pgvector — search does a seq scan and crawls on large tables.
The index metric (cosine) doesn't match how the embeddings are normalized.
Choosing Pinecone for private data without accounting for data leaving to the vendor.

Pinecone (managed)

import os from pinecone import Pinecone, ServerlessSpec pc = Pinecone(api_key=os.environ["PINECONE_API_KEY"]) pc.create_index( name="docs", dimension=1536, metric="cosine", spec=ServerlessSpec(cloud="aws", region="us-east-1"), ) index = pc.Index("docs") index.upsert(vectors=[{"id": "d1", "values": vec, "metadata": {"src": "faq"}}])

Pros: zero infrastructure, auto-scale. Cons: cost, data at the vendor, vendor lock-in.

Qdrant (self-hosted in Docker)

docker run -p 6333:6333 qdrant/qdrant

from qdrant_client import QdrantClient from qdrant_client.models import Distance, VectorParams q = QdrantClient(url="http://localhost:6333") q.recreate_collection( collection_name="docs", vectors_config=VectorParams(size=1536, distance=Distance.COSINE), )

Open source, powerful payload filtering, hybrid mode. Great when data can't leave your perimeter.

pgvector (Postgres extension)

CREATE EXTENSION IF NOT EXISTS vector; CREATE TABLE chunks (id bigserial PRIMARY KEY, body text, embedding vector(1536)); CREATE INDEX ON chunks USING hnsw (embedding vector_cosine_ops); SELECT body FROM chunks ORDER BY embedding <=> '[...]'::vector LIMIT 5;

Ideal when you already run Postgres: transactions, JOINs with business data, a single backup. <=> is cosine distance, <-> is L2, <#> is negative inner product.

Index types

HNSW (nearest-neighbor graph) — high recall, fast search, expensive memory and build. The default almost everywhere. Params: m (connectivity), ef_construction, ef_search (recall vs query speed).

IVF (inverted lists + clusters) — more compact, searches only the closest nprobe clusters; slightly lower recall, better for very large datasets with memory limits.

Сообщить об ошибке

Vector Stores: Pinecone, Qdrant, pgvector, Chroma, FAISS

Квиз — 4 вопроса

Pinecone (managed)

Qdrant (self-hosted in Docker)

pgvector (Postgres extension)

ChromaDB and FAISS

Index types

How to choose