Visualizers

Learn by moving things

Interactive explainers for the ideas behind retrieval, embeddings, and agents. A formula tells you what is true; dragging a vector shows you why. Each one also appears inside the article it belongs to, and runs entirely in your browser, so drag, poke, and break it.

Portrait of Sachin Gupta rendered in binary

How a vectorless index is built

From: Vectorless RAG →

The full index-building algorithm, step by step, the way tree-index systems like PageIndex do it: parse pages, detect or generate the table of contents, split into overlapping page-groups, extract and number the hierarchy, locate and verify every section, then summarize each node. Each step is labelled as an LLM call or deterministic code. No embeddings, no chunking anywhere.

Input document

Employee Handbook.pdf

200 pages

each page tagged <physical_index_N>

Table-of-contents tree

(built after the structure is extracted)

step 1 / 8

Parse the document into pagesdeterministic code

The PDF is read into a list of pages. Each page is tagged with its number, like <physical_index_7>, so every later step can point back to exact pages. Nothing is embedded or chunked.

The text-analysis pipeline

From: Text Analysis →

Type a sentence and watch raw text become the terms an index stores: tokenize, lowercase, drop stopwords, then stem to a root. The same pipeline runs on your query too.

Type a sentence

1 · Tokenize

TheCatsareRunningquicklytoNYC

2 · Lowercase

thecatsarerunningquicklytonyc

3 · Remove stopwords

thecatsarerunningquicklytonyc

4 · Stem → index terms

cat ←catsrun ←runningquick ←quicklynyc

The dark chips are what actually gets indexed. The same pipeline must run on the query too, so “Running” in a document and “run” in a search still meet as “run”.

Retrieve then rerank

From: Rerankers →

First-stage retrieval ranks by a cheap score; a cross-encoder reranker rereads each result against the query and reorders. Press Rerank and watch an on-topic but keyword-poor result jump to the top.

Query: “how do I cancel my subscription?”

First-stage retrieval (fast, crude)cheap score

1Subscription pricing and plans0.88
2Cancel your subscription0.82
3Subscription terms of service0.79
4Pause your subscription temporarily0.60
5How to update your payment method0.55
6End your membership and stop billing0.41

First-stage ranks by a cheap score, so “Subscription pricing” tops the list on keyword overlap alone. Press Rerank to score each result against the query with a cross-encoder.