XenorsVector Databases: ChromaDB, Pinecone and FAISS
Keyword search misses meaning. Vector databases store embeddings, numeric fingerprints of text or images, and find the closest ones to a query in milliseconds.
How vector search works
An embedding model maps text to a vector so that similar meanings sit close together. At query time the question is embedded and compared against stored vectors. Exact search compares every vector. Approximate nearest neighbour (ANN) indexes such as HNSW or IVF trade a little accuracy for large speed gains.
| Tool | Type | Best for |
|---|---|---|
| FAISS | Library | Fast local search, research |
| ChromaDB | Embedded database | Prototypes and small apps |
| Pinecone | Managed cloud service | Production without operating infrastructure |
Semantic search with FAISS
Normalise vectors and use inner product to get cosine similarity.
# pip install faiss-cpu sentence-transformers
import faiss, numpy as np
from sentence_transformers import SentenceTransformer
docs = ["How to reset my password", "Refund policy for annual plans",
"Train a random forest in Python", "Fix a CUDA out of memory error"]
model = SentenceTransformer("all-MiniLM-L6-v2")
vecs = model.encode(docs, normalize_embeddings=True).astype("float32")
index = faiss.IndexFlatIP(vecs.shape[1])
index.add(vecs)
q = model.encode(["GPU memory error when training"], normalize_embeddings=True).astype("float32")
scores, ids = index.search(q, 2)
for s, i in zip(scores[0], ids[0]):
print(round(float(s), 3), docs[i])Store and filter with ChromaDB
Metadata filters combine semantic matching with structured conditions.
import chromadb
client = chromadb.PersistentClient(path="chroma_data")
col = client.get_or_create_collection("faq", metadata={"hnsw:space": "cosine"})
col.add(ids=["1", "2", "3"],
documents=docs[:3],
embeddings=vecs[:3].tolist(),
metadatas=[{"topic": "account"}, {"topic": "billing"}, {"topic": "ml"}])
res = col.query(query_embeddings=q.tolist(), n_results=2, where={"topic": "ml"})
print(res["documents"])Use Pinecone as a managed service
The index must be created with the same dimension as your embedding model. Check the Pinecone documentation for current index options and regions.
# pip install pinecone
import os
from pinecone import Pinecone, ServerlessSpec
pc = Pinecone(api_key=os.environ["PINECONE_API_KEY"])
if "faq" not in pc.list_indexes().names():
pc.create_index("faq", dimension=384, metric="cosine",
spec=ServerlessSpec(cloud="aws", region="us-east-1"))
idx = pc.Index("faq")
idx.upsert([(str(i), vecs[i].tolist(), {"text": docs[i]}) for i in range(len(docs))])
print(idx.query(vector=q[0].tolist(), top_k=2, include_metadata=True))Choosing and tuning
Start with ChromaDB or FAISS locally. Move to a managed or server-based database when you need replication, access control or very large collections. Use the same embedding model for indexing and querying, and re-embed everything if you change it.
Hands-on practice
Run the examples in the downloadable Jupyter notebook. The extra checks below focus on mistakes that appear in real projects, not just syntax.
Compare cosine similarity
import numpy as np
a=np.array([1.,0.,1.]); b=np.array([.8,.1,.9])
cos=a@b/(np.linalg.norm(a)*np.linalg.norm(b))
print(round(cos,3))~0.994
High cosine similarity means similar direction, not necessarily identical semantics.
Apply metadata filtering
records=[('doc1','policy'),('doc2','manual'),('doc3','policy')]
print([r for r in records if r[1]=='policy'])[('doc1', 'policy'), ('doc3', 'policy')]Metadata filters can improve relevance before vector ranking.
Self-check
- Can you explain why the first practice result is trustworthy rather than just plausible?
- What would you monitor or test before using this Vector Databases workflow in production?
- Which assumption in the example would break first on a larger or messier dataset?
Frequently asked questions
What is an embedding?
A list of numbers produced by a model that represents the meaning of text, an image or other data.
Cosine similarity or Euclidean distance?
For normalised text embeddings they rank results identically. Use the metric your embedding model recommends.
Do I need a vector database for RAG?
Small projects can use FAISS or an in-memory index. A vector database helps with persistence, filtering and scale.
Primary references
Use these official sources when exact syntax, defaults or definitions matter beyond the examples here.
