analysis
Search playground
Run semantic retrieval against your crawl, read the results, and use Page Lens to inspect matches.
What it is
The search playground lets you query your crawl by meaning — like asking which pages actually answer a question. Your query is embedded on-device with the same model used for the crawl, then ranked against stored page and chunk vectors. Nothing is sent to the cloud for search, and no AI writes answers for you.
When it's available
Switch to Search from the view switcher in the title bar after embeddings finish. If a crawl has no embeddings yet, search will report that and return no results.
How ranking works
Each page ranks by its best-matching passage — the specific chunk that answers your query — not just page-level similarity. This mirrors how retrieval systems and LLMs with a vector index see your site.
Reading result cards
Strong matches stand out; weaker ones fade.
| Field | Meaning |
|---|---|
| Title + path | Page identity, like a search result |
| Passage | The chunk that made the page rank |
| Match % | How strongly the page answers your query |
| Chunk position | Where on the page the match lives (e.g. 3/8) |
Grid and table views
Toggle between a thumbnail grid and a table from the dock toolbar next to the search bar. Table view adds sortable columns including semantic group and optional per-fanout scores. The query palette stores recent searches for quick reuse.
Page Lens
Click a result card to open Page Lens. The left rail shows query variants, a chunk map, and similar pages on the site. The right side shows the full page content with the best-matching chunk highlighted and runner-up passages marked. Press back or Esc to return to results.
Switch query phrasing
When fan-outs are available, pick a variant pill to re-score highlights on the page against a different phrasing of your query. This lets you see how match regions move as the wording changes.
Attach to the agent
Drag a result card into the agent chat, or use @ to attach a page. The agent receives the page as crawl context for follow-up questions.
Compare workspace
Attach a competitor workspace to search both crawls in one merged result list, scored on the same scale. Results are labeled by site. Requires a paid plan and embeddings on both crawls.
Query fan-outs
On paid plans, Silkra generates alternate phrasings of your query and scores results against each one. Fan-out columns appear in table view so you can compare which wording surfaces different pages.