Vector Databases

import chromadb
from chromadb.utils import embedding_functions

client = chromadb.Client()
embed_fn = embedding_functions.SentenceTransformerEmbeddingFunction(
    model_name="all-MiniLM-L6-v2"
)
collection = client.create_collection(name="faq", embedding_function=embed_fn)

collection.add(
    documents=[
        "You can reset your password from the account settings page.",
        "Refunds are processed within 5–7 business days.",
        "Our customer support team is available Monday to Friday, 9am–6pm EST.",
    ],
    ids=["faq1", "faq2", "faq3"]
)

results = collection.query(
    query_texts=["How do I change my login credentials?"],
    n_results=2
)
print(results["documents"])

[['You can reset your password from the account settings page.',
  'Our customer support team is available Monday to Friday, 9am–6pm EST.']]

The query 'change my login credentials' semantically matched 'reset your password' with zero keyword overlap.

# Step 1: Embed and upsert document chunks (run once at ingestion)
for chunk in chunks:
    embedding = openai.Embedding.create(
        input=chunk["text"], model="text-embedding-3-small"
    )["data"][0]["embedding"]
    index.upsert([(chunk["id"], embedding, {"text": chunk["text"]})])

# Step 2: Query at runtime
user_query = "How many vacation days do I get?"
query_embedding = openai.Embedding.create(
    input=user_query, model="text-embedding-3-small"
)["data"][0]["embedding"]

results = index.query(vector=query_embedding, top_k=2, include_metadata=True)
context = " ".join([r["metadata"]["text"] for r in results["matches"]])

# Step 3: Pass context to LLM
response = openai.ChatCompletion.create(
    model="gpt-5.5",
    messages=[
        {"role": "system", "content": "Answer using only the provided context."},
        {"role": "user", "content": f"Context: {context}\n\nQuestion: {user_query}"}
    ]
)

Employees are entitled to 20 days of paid leave per year.

This is the full RAG loop: embed → store → retrieve → generate. The LLM never hallucInates because it's grounded in retrieved context.

import faiss
import numpy as np

# Simulated CLIP embeddings for 5 products (512-dim)
product_embeddings = np.random.rand(5, 512).astype("float32")
faiss.normalize_L2(product_embeddings)  # Normalize for cosine similarity

# Build FAISS flat index
index = faiss.IndexFlatIP(512)  # Inner product = cosine similarity after normalization
index.add(product_embeddings)

# Simulate user uploading a photo → CLIP produces query embedding
query_embedding = np.random.rand(1, 512).astype("float32")
faiss.normalize_L2(query_embedding)

# Search for top 3 similar products
distances, indices = index.search(query_embedding, k=3)
print("Top 3 similar product indices:", indices[0])
print("Similarity scores:", distances[0])

Top 3 similar product indices: [2, 0, 4]
Similarity scores: [0.973, 0.961, 0.948]

The three most visually similar products are retrieved in order of cosine similarity. In production, these indices map back to product IDs and catalog metadata.