Schemas

Message content

URL to fetch content from

Pre-supplied document content (used as fallback if fetch fails). In sources_mode=context, only the first 5000 characters are injected into the prompt.

Document title

User question

Canonical domain filter for search (without www or URL scheme) — if omitted, no site: constraint is applied

Stream the response as SSE

Conversation history — strictly alternating user/assistant pairs (user first), max 10 pairs (20 messages)

Response length mode

Answer language

Whether the response should use Markdown formatting

Whether to generate up to 5 related search queries (search-engine style)

Rewrite search query before web search

Caller-supplied sources — bypasses Bloom search when provided. query_rewrite is ignored when this field is set.

'search' fetches+reranks sources like web results (default); 'context' injects full content directly without reranking

Request / trace identifier.

Reranked text chunk.

Relevance score between 0 and 1.

Original snippet from the search provider.

Present only when extra_snippets=true.

Present only when full_content is set.

Publication date (ISO 8601).

Request / trace identifier.

Defaults to fr-FR.

Defaults to 10.

Defaults to 0.

Defaults to false.

If provided, fetches pages for each SERP result and returns the full page content on each link. markdown returns the extracted body as Markdown; html returns the raw page HTML. When combined with extra_snippets, results follow the reranked order.

Maximum number of scored chunks retained per URL in the output. Defaults to 3.

Minimum reranker score (0–1) for a chunk to be retained. Defaults to 0.10.

Restrict results to these domains. POST only. Up to 10 entries (extras are dropped). Mutually exclusive with exclude_domains. Equivalent to site: operators in q.

Exclude these domains from results. POST only. Up to 10 entries (extras are dropped). Equivalent to -site: operators in q.