Extract and correctly attribute the source URLs and metadata Perplexity used to ground a Sonar chat completion answer.

Documented steps

1. After a normal (non-streaming) chat completions call, read the top-level citations field: an array of plain source URL strings, in referenced order (https://docs.perplexity.ai/api-reference/sonar-post).
2. Read the richer top-level search_results field (array of search result objects) for per-source metadata — this is separate from citations.
3. If return_related_questions was set true, also read the top-level related_questions array (follow-up query strings) generated from the same search results.
4. If return_images was set true, read the top-level images array — otherwise this field is absent/null.
5. In streaming mode, don't poll for citations/search_results until the final chunk(s) have arrived — earlier chunks only carry delta.content (https://docs.perplexity.ai/docs/sonar/features).
6. Never rely on the model inlining literal URLs inside message.content as ground truth — treat citations/search_results as authoritative, per Perplexity's own structured-output guidance to prefer these fields over in-text/JSON links (https://docs.perplexity.ai/docs/sonar/features).