How should this corpus be chunked, embedded and searched?
Paste or drop a sample of the documents your RAG or search will cover, and a few queries your users really ask. Your browser chunks it the way you set it - fixed window, recursive, heading sections or sentences, at your size and overlap - and shows what that breaks: chunks that end mid-sentence, code blocks and tables split in two, headings cut off from their text, fragments, duplicates, repeated footers, mixed languages. It compares all four strategies, and runs BM25 over the chunks for every query, so you see which queries keyword search already finds and which need embeddings. Free, before you sign in, and nothing is uploaded. Then the plan lane picks the chunking and the embedding model, and the retrieval lane designs keyword, dense or hybrid search - and the page checks both against its own numbers.
The companies, products and codes in the examples are invented. The first three are plans (the first also carries a saved retrieval run, one click after the plan), the last two are retrieval designs; each comes with a saved model reply, so you can see the whole page without signing in or spending a credit.
What this does, and what it does not
The check is a reading of the text, not an embedding: no model runs in it. Each document is split into blocks with exact character offsets, so every chunk is a slice of your text and every break can be located. Token counts are an estimate close to OpenAI's cl100k tokenizer (roughly one token per four characters of English), not a model's own tokenizer. BM25 here is the textbook Okapi formula (k1 1.2, b 0.75) with light stemming and accent folding - the keyword leg of a hybrid search, not your search engine's exact ranking. Whether embeddings find a paraphrase cannot be measured without the embedding model, so the page says which queries keyword search misses and leaves that judgement to the paid lanes and to your own evaluation.
The paid lanes see your sample (the middle is cut past 30,000 characters, and the page says so) and the page's measurements. Every chunk id they cite is checked, a plan is re-run on your sample, and every retrieval route is checked against the page's BM25. A paste stays in this browser unless you run a lane. Derived from two agent skills by Seth Hobson (wshobson/agents, MIT): @wshobson/embedding-strategies and @wshobson/hybrid-search-implementation. Sources are listed in llms.txt.