SearchApiWebSearch read organic_results only, so with search_params such as {"engine": "perplexity"} the request succeeded but the component returned no documents and no links. SearchApi's answer engines (google_ai_mode, perplexity, chatgpt, gemini, bing_copilot) share one response shape: the answer as markdown plus typed text_blocks, the cited sources as reference_links, and for Google AI Mode the web_results it drew from.

When a response carries that shape, the component now returns the synthesized answer as the first Document (content from the markdown field, rendered from text_blocks when the API omits it), one Document per reference link, then the web results. links holds the reference links followed by the web result links. Classic engines are unchanged.

Tests cover the parser with and without markdown, with web results, with an empty answer, and one run through the component with a mocked httpx.get.

Related Issues

None open. SearchApiWebSearch reads organic_results only, so with search_params={"engine": "perplexity"} (or google_ai_mode, chatgpt, gemini, bing_copilot) it returns no documents and no links although the request succeeds.

Proposed Changes

When a response carries the shared AI answer shape (text_blocks, markdown, reference_links), the component now returns the synthesized answer as the first Document (content is the markdown field, rendered from text_blocks when the API omits it), one Document per cited reference link, and for Google AI Mode the web_results it drew from. links holds the reference links, then the web result links. Documents for the classic engines are unchanged. The class docstring documents the behaviour.

How did you test it?

Unit tests for the parser (markdown present, markdown absent, web results, empty answer) and one run through run with a mocked httpx.get. hatch run fmt-check, hatch run test:unit and hatch run test:types equivalents pass locally (ruff, pytest with 21 unit tests, mypy). The integration test still requires SEARCHAPI_API_KEY.

Notes for the reviewer

_text_blocks_to_markdown and its two helpers are module-level functions rather than methods so the rendering stays easy to unit test and under the ruff complexity limits.

Checklist

• I have read the contributors guidelines and the code of conduct
• I have updated the related issue with new insights and changes (no issue to update)
• I added unit tests and updated the docstrings
• I've used one of the conventional commit types for my PR title: feat