Can I do vector search with only embeddings and no vector DB?

Yes. You can absolutely do vector search with just embeddings and no vector database.

How it works

1. Generate embeddings for your documents/items.
2. Store them somewhere simple, like:
   • a JSON/CSV file
   • a database table
   • a Python list / NumPy array
3. At query time:
   • embed the user query
   • compute similarity against all stored embeddings
   • return the top matches

Common similarity methods

• Cosine similarity — most common for text embeddings
• Dot product — often equivalent if vectors are normalized
• Euclidean distance — sometimes used, less common for text search

Example in Python

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

docs = ["cat on a mat", "deep learning basics", "best pizza recipes"]
doc_embeddings = np.array([
    [0.1, 0.2, 0.3],
    [0.9, 0.8, 0.7],
    [0.2, 0.1, 0.4]
])

query_embedding = np.array([[0.15, 0.18, 0.28]])

scores = cosine_similarity(query_embedding, doc_embeddings)[0]
top_idx = np.argsort(scores)[::-1]

for i in top_idx:
    print(docs[i], scores[i])

When this is a good idea

• Small to medium datasets
• Prototyping
• Simple apps
• When you want to avoid extra infrastructure

When a vector DB helps

A vector DB becomes useful when you need:

• very large scale
• fast approximate nearest-neighbor search
• filtering/metadata queries
• persistence and replication
• managed indexing
• low-latency production workloads

Rule of thumb

• Up to tens of thousands of vectors: plain arrays/files/SQL are often fine
• Hundreds of thousands to millions: a vector DB or ANN index is usually better
• Very large or high-traffic systems: definitely consider a vector search engine

If you want, I can show you:

• a pure Python example,
• a SQL-based approach,
• or how to do it efficiently with NumPy / FAISS without a vector DB.