PR #7750 - pydantic/pydantic-ai

Depends on #7434
Part of #3126

Adds citations from Bedrock and xAI in the TextPart.citations shape from #7434, plus recordings and docs for OpenRouter (including Perplexity models), whose citations come from #7434's OpenAI Chat mapping.

How it works, at a glance

Provider	What's cited	How to turn it on	Anchor
Bedrock, Anthropic models	TXT and PDF documents, in the prompt or returned by tools	`Citations()`	Whole text part
OpenRouter	Web pages, including Perplexity's	Turn on web search	None: OpenRouter's offsets are 0/0
xAI	Web, X and collection citations	`Citations()`	The `[[1]](https://...)` marker the model writes

When Citations() asks for citations. The same rule as #7434, because these providers fail the same way:

Situation	Bedrock (Anthropic models)	xAI
Text output	Asks	Asks
Structured output	Doesn't ask: markup in fields 3 of 3 raw, `outputConfig` gets a 400	Doesn't ask: tool output failed 6 of 6 live (the model answers in prose instead of calling the output tool; 5 of 5 passed without), and markers reached `NativeOutput` fields 2 of 6
A function tool forced with `tool_choice`	Raises `UserError` when the request has a document: markup in the arguments 3 of 3, 0 of 3 without	Asks: forced tool arguments stayed clean 3 of 3 live
Function tools the model chooses	Asks; the docs warn (markup 2 of 9 on Haiku 4.5)	Asks: tools are called as usual (6 of 6) and their arguments stayed clean (0 of 4)

xai_include_inline_citations still takes precedence over citations when set.

Decisions

Bedrock citations only on models that support them. A live probe (2026-10-20) found Claude returns citations, while Nova, Llama, Mistral Pixtral and DeepSeek reject the whole request (This model doesn't support citations). A bedrock_supports_citations profile flag is set for Anthropic models, and the setting is ignored elsewhere.
Bedrock leaves out documents it can't cite. Plain-text documents go as text rather than bytes, so with citations on a text document that isn't UTF-8 raises a UserError (in the docs' provider table). Other formats (docx, csv, md, html) and S3 documents go without citations, because Bedrock rejects S3 with citations on (S3 as a document source is not allowed when citations are enabled).
Bedrock document titles are positional (Document 1), the name Pydantic AI gives each document. The docs say this.
A Bedrock document with citations keeps its place in the message, like a PDF, so a cache point after it still works.
xAI collection citations are document sources; their chunk and collection IDs stay in provider_details['citations']. xAI is the one provider where Citations() changes the answer text, because of the markers.
xAI streaming attaches citations only when it's unambiguous: to the last text part, only when exactly one output has citations and its text equals that part. Otherwise they're left off rather than guessed.

Limits

Structured output on Bedrock and xAI gets no citations from Citations() (above).
A streamed xAI answer followed by a tool call gets no citations; the docs say so.
xAI: only X citations are recorded. Web and collection citations are mapped from the SDK types; collection citations couldn't be run live. Anchors are checked against the text's bounds only.
Bedrock web locations are mapped from the documented web member, but aren't recorded.

Verification

Recorded tests: Bedrock text, PDF, and a document returned by a tool; OpenRouter and OpenRouter with Perplexity, streamed and not; xAI streamed and not. Since main's #10845, Bedrock Claude without max_tokens streams, so the Converse shape is covered by unit tests and the stream parity test.
Recordings starting with emoji and accented text check that OpenAI Responses, OpenAI Chat and xAI anchors select the right text. A live probe (2026-10-22) showed Anthropic and Bedrock count document positions in Python characters.
Live acceptance on 2026-10-22 fixed three Bedrock problems: S3 documents with citations got a 400, the prompt cache missed on the second turn (now 8,754 cached tokens, against 0), and structured output fields got markup.
Live check on 2026-10-23 of every output mode with Citations() on and off, on every provider, and of function tools with documents. It found the xAI structured-output failure and the Bedrock forced-tool markup, both fixed above. On every other provider structured output behaves the same with and without Citations().

Checklist

Any AI generated code has been reviewed line-by-line by the human PR author, who stands by it.
No breaking changes in accordance with the version policy.
Any permitted compatibility impact has a label, warning, migration, release note, and exact API-check waiver.
PR title is fit for the release changelog.