Interactive tools and visualizations that explain how modern AI systems work — from retrieval pipelines and embeddings to search, reasoning, and system design.
This space contains hands-on AI engineering experiments built to make complex concepts easier to understand through visualization and interaction.
Retrieval-Augmented Generation, step by step
Learn how Retrieval-Augmented Generation works step-by-step by exploring chunking, embeddings, vector search, retrieval, prompt construction, and answer generation.
Compare prompts and see how outputs shift
Experiment with system prompts, few-shot examples, and constraints to understand how prompt structure changes model behavior.
See text projected into vector space
Project sentences into a 2D map to build intuition for how embeddings cluster semantically similar text.
Watch similarity ranking happen live
Inspect cosine similarity, top-k retrieval, and how index parameters trade off recall against latency.
Keyword vs. semantic retrieval
Compare lexical and semantic search side by side to see where each one wins on real queries.
Trace tool-calling and reasoning loops
Step through an agent's plan, tool calls, and observations to understand how multi-step reasoning is orchestrated.
Budget tokens across a prompt
Visualize how system prompts, retrieved context, and history compete for a finite context window.
A quick primer on the building blocks these tools make tangible. For a full, interactive walkthrough, run the RAG Explorer.
Retrieval-Augmented Generation (RAG) pairs a search system with a language model. Instead of answering only from memorized training data, the model first retrieves relevant passages from a knowledge source and answers from that evidence — keeping responses current, accurate, and traceable.
An embedding turns text into a vector of numbers that captures meaning. Texts with similar meaning get similar vectors, which is why embeddings power semantic search: they match ideas rather than exact keywords, finding the right passage even when the wording is different.
Vector search ranks stored vectors by how close they are to a query vector, usually with cosine similarity. A vector database compares the question against every chunk and returns the closest matches in milliseconds — the retrieval step that makes real-time RAG possible.
A production assistant ingests documents ahead of time — chunking, embedding, and storing them. At query time it embeds your question, retrieves the most relevant chunks, and builds a grounded prompt so the model answers from real sources instead of guessing.
