# The retrieval model

Understand what is shared, what stays native, and where evidence comes from.

Canonical: https://searchsdk.dev/docs/concepts
Markdown: https://searchsdk.dev/docs/concepts.md

<a id="normalized-and-native"></a>

## Normalized and native

Normalized search offers a portable vocabulary for a small shared operation. Provider-native APIs preserve a provider's own arguments, job models, streaming events, and returned data. Choose a surface based on the operation you need, rather than assuming that a similar name implies identical behavior.

<a id="four-kinds-of-material"></a>

## Four kinds of material

| Material | Interpretation |
| --- | --- |
| Search result | A ranked source link with provider search text. |
| Extracted content | Content retrieved for a known page, with its own extraction outcome. |
| Generated answer | A synthesis that requires separate source and claim handling. |
| Corpus passage | A span from an ingested document; document and revision identity matter. |

WonderSearch.ai and WonderSearch.com refer to one corpus candidate. Corpus passage IDs are not web source URLs and should never be fabricated into normalized web results.

<a id="provider-differences"></a>

## Provider differences remain visible

- Native scores are preserved when available. They are not a common relevance scale.
- A result ceiling does not promise an exact count or multiple pages.
- Provider-native hostname semantics still apply after request validation.
- An absent usage value is unknown, rather than free.