Retrieval race
One query fires at three retrieval strategies over the same corpus — keyword, vector, and hybrid — and three ranked lists populate side by side. The catch: the answer chunk is phrased in synonyms, so keyword search can't match it. Watch exactly where each strategy ranks the right answer, including when keyword search buries it below the visible results entirely. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Read the query and note the answer chunk uses different words than the query ("recover access" vs "reset password").
- Run the race — keyword, vector, and hybrid retrieval each rank the same corpus.
- See the answer's rank in each strategy — found near the top, or buried below the visible top-k where the model never sees it.
What this clears up (the fundamentals)
- Retrieval is usually the bottleneck, not the model — when a RAG system "hallucinates" or says "I don't know," the right chunk was often never in the context. The failure is upstream.
- Keyword search misses synonyms — it matches literal terms, so a chunk that answers the question in different words scores zero and sinks below the top-k.
- Vector search matches meaning — embeddings put "recover account access" near "reset password" even with no shared words, so it surfaces the answer keyword search lost.
- Hybrid fusion is the safe default — reciprocal-rank fusion blends both rankings, keeping keyword precision while recovering the answers vector alone would find — and vice versa.
Where it's used
A RAG-engineering literacy tool — the fastest way to see why "just add RAG" isn't enough and why retrieval quality decides answer quality. It's a miatz build-lab concept playable; learn to build it yourself.
FAQ
Why does keyword search miss the answer?
Because it ranks by literal term overlap. A chunk that answers "how do I reset my password" by saying "recover your account access" shares no query words, so keyword search scores it near zero and ranks it far down the list.
What is hybrid (RRF) retrieval?
Reciprocal-rank fusion combines two rankings by summing 1/(k + rank) across them. A chunk that ranks well in either keyword or vector search bubbles up, so hybrid keeps keyword's precision and vector's recall.
Why does the rank matter so much?
Because only the top-k chunks are put into the model's context. If the answer ranks below the cutoff, the model literally never reads it — no amount of prompting fixes a retrieval miss.
Is anything uploaded?
No. The corpus, scoring and ranking all run in your browser — nothing is transmitted or stored. Turn off your Wi-Fi and it still works.
Limits
A teaching model with a tiny corpus and pre-computed vector scores — real retrieval depends on your embeddings, index, and re-ranking. The lesson (measure the answer's rank across strategies) transfers directly.
Related
Part of the miatz playgrounds. Explore the rest from the playground catalog.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.