Simulates fixed-size chunking (~150 words). Each passage is scored in isolation, with no title and no surrounding text, exactly the object that competes in the vector index.
Heuristic scan for the failure modes in the article: unnamed subject, dangling pronouns, vague time anchors, buried answers, and low specificity. An orphaned chunk is one that never says what it is about.
What it checks
No model reads your page, it reads passages of it, scored alone. A passage that carries its meaning through pronouns or references to earlier paragraphs ("the division", "this approach", "last year") loses its subject when cut, and its vector drifts out of range of the query it should answer. This tool splits your content into passages and flags the orphaned ones, chunks that never name their own subject.
It runs entirely in your browser, using transparent heuristics rather than a real embedding model.
How to use it
- Paste your page or section.
- See how it splits into passages, each scored alone.
- Fix the orphaned chunks it flags.
Frequently asked questions
How do AI engines chunk a page?
Retrievers split your HTML into uniform passages of roughly 100 to 300 words (often about 128 tokens), then score each chunk in isolation against the query. Your page is represented by its single best chunk, so every passage has to name its own subject.
What is an orphaned chunk?
A passage that only makes sense with the context above it, a dangling pronoun (it, this), a vague date (last year), or an answer buried below the heading that promised it. Orphaned chunks score poorly because they read as being about nothing.
Is the Chunk Retrievability Analyzer free to use?
Yes. The Chunk Retrievability Analyzer is completely free, with no sign-up, no usage limits and no watermark on the output.
Is my data private?
Yes. The Chunk Retrievability Analyzer runs entirely in your browser. Nothing you paste or enter is uploaded, stored or sent to any server.
Related tools and reading
Keep going with the playbooks and tools behind this one.