Vector Databases: Pinecone, Weaviate, Chroma

In traditional apps, we use SQL (PostgreSQL) or NoSQL (MongoDB). We query for exact matches: WHERE user_id = 123.

In AI apps, we need to query for meaning: WHERE content is similar to 'happy dog'.

To do this, we use Vector Databases.

What is a Vector (Embedding)?

An LLM converts text into a long list of floating-point numbers, typically 1,536 dimensions (for OpenAI) or 768 (for open source). This list represents the semantic meaning of the text.

[0.001, -0.23, 0.88, ...]

• “Dog” and “Puppy” will have vectors that are mathematically close (high cosine similarity).
• “Dog” and “Tax Return” will be far apart.

The Players

1. Pinecone

The Cloud Native Choice.

• Fully managed (SaaS).
• Fast, scalable, expensive.
• Good for enterprise teams who don’t want to manage servers.

2. Weaviate

The Hybrid Choice.

• Open source, but has a cloud offering.
• Supports “hybrid search” (Vector + Keyword) out of the box very well.
• Has a strong schema definition (like GraphQL).

3. Chroma (ChromaDB)

The Developer Choice.

• Pure open source, runs locally easily.
• pip install chromadb.
• Perfect for prototyping and small-to-medium apps.

4. pgvector (PostgreSQL)

The Pragmatic Choice.

• It’s just a Postgres extension!
• If you already use Postgres, just add a vector column.
• Pro: One database to manage. ACID compliance.
• Con: Doesn’t scale to billions of vectors as well as specialized DBs, but fine for 99% of apps.

How to Choose?

• Prototyping: Chroma or generic FAISS.
• Production (Small/Mid): pgvector (Keep your stack simple).
• Production (Huge Scale/Enterprise): Pinecone or Weaviate Cloud.

The Code (Python + Chroma)

import chromadb

client = chromadb.Client()
collection = client.create_collection("my_docs")

# Add documents (automatically embedded)
collection.add(
    documents=["This is a document about cats", "This is about dogs"],
    metadatas=[{"source": "wiki"}, {"source": "wiki"}],
    ids=["id1", "id2"]
)

# Query
results = collection.query(
    query_texts=["animal"],
    n_results=2
)
# Returns both documents because both are animals.

Conclusion

Vector Databases are the long-term memory of AI. Choosing the right one depends on whether you value simplicity (pgvector) or raw performance (Pinecone).

---

Next: Running LLMs Locally — Take back control.