Interactive Vector Embeddings & Semantic Search Lab | Large Language Models & Generative AI

Vector RAG Lab: Interactive Lab

LLMs & RAGLesson / Article Lab

Calculates normalized cosine similarity scores across candidate knowledge chunks.

Semantic Vector Retrieval & Cosine Similarity Ranking

import numpy as np

def cosine_similarity(a, b):
    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

# 4-dimensional normalized concept embeddings
knowledge_base = {
    "Chunk 1: Python Memory & Allocators": np.array([0.92, 0.12, 0.08, 0.15]),
    "Chunk 2: Deep Learning Backpropagation": np.array([0.10, 0.88, 0.85, 0.20]),
    "Chunk 3: Vector Embeddings & HNSW Indexes": np.array([0.18, 0.72, 0.91, 0.40]),
    "Chunk 4: Prompt Engineering Techniques": np.array([0.35, 0.40, 0.50, 0.82])
}

# User Query: "How does vector nearest-neighbor search index passages?"
query_vector = np.array([0.20, 0.70, 0.95, 0.38])

results = []
for title, vector in knowledge_base.items():
    score = cosine_similarity(query_vector, vector)
    results.append((title, score))

results.sort(key=lambda x: x[1], reverse=True)

print("Ranked Retrieval Results:")
for rank, (title, score) in enumerate(results, 1):
    print(f"  #{rank} [Score {score:.4f}] {title}")

print(f"Top context selected: '{results[0][0]}' for LLM prompt augmentation.")

Pyodide WasmCPython 3.12 • NumPy

Interactive Challenge: Modify code inputs, click Run Code to execute live in WebAssembly.

+25 XP Reward

Wasm Terminal Output

Click Run Code to execute this algorithm in the browser sandbox.

Welcome to the Vector Embeddings & Semantic Retrieval Lab.

Retrieval Augmented Generation (RAG) bridges private knowledge stores with foundational Large Language Models. At the center of every RAG system lies vector semantic search: converting raw text passages into high-dimensional geometric embeddings where semantic similarity maps directly to angular proximity.

Lab Objectives

Understand vector normalization and calculate dot products vs cosine distances.
Explore Approximate Nearest Neighbor (ANN) index behaviors (HNSW graphs and IVF partitions).
Evaluate chunking strategies: how document overlap and token boundaries influence embedding fidelity.

Interactive Vector Search & Embeddings Playground

Use the live Python code sandbox below to compute document embeddings, query vectors, and rank retrieved chunks by relevance score.

import numpy as np

def cosine_similarity(a, b):
    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

# Simulated 4-dimensional semantic embeddings
docs = {
    "Doc 1 (Python Allocator)": np.array([0.9, 0.1, 0.05, 0.2]),
    "Doc 2 (Transformer Attention)": np.array([0.15, 0.85, 0.9, 0.3]),
    "Doc 3 (Vector DB HNSW Index)": np.array([0.2, 0.7, 0.88, 0.45]),
}

# User Query: "How does vector index nearest neighbor retrieval work?"
query = np.array([0.18, 0.75, 0.92, 0.4])

print("Query-Document Relevance Scores:")
for name, vec in docs.items():
    score = cosine_similarity(query, vec)
    print(f"  {name}: {score:.4f}")

Production RAG Best Practices

Reciprocal Rank Fusion (RRF): Combine dense vector search with sparse keyword search (BM25) to avoid catastrophic misses when users search for specific error codes or part numbers.
Rerankers: Pass top-50 vector results into a Cross-Encoder reranker (such as bge-reranker-large) to produce higher precision top-5 context windows for the LLM.