Search + Extract
Ground an LLM in the live web
Replace a model's frozen training data with pages that actually answer the question.
A model's training data is frozen and its guesses are confident. Search returns the pages that actually answer the question, already stripped to readable prose, so the context you paste into the prompt is evidence rather than recall.
- One call returns ranked results and their readable body text, so there is no second fetch stage to build and maintain.
- Navigation, cookie banners and footers are stripped before the text reaches your prompt, which is context you would otherwise pay tokens for.
- Every result carries its score broken into six signals, so a retrieval step can drop weak sources instead of passing them to the model.
Ask for enriched results and one call returns the ranking and the body text together, instead of a search followed by twenty fetches you have to write yourself.