Pydantic AI has added structured citations to model response text in its latest release. Applications can preserve provider citations through conversation history and a Vercel adapter, rather than handling only links embedded in answers. The release also fixes input-guardrail behaviour and concurrent workspace handling. Citation metadata identifies a supplied source without independently checking whether it supports the claim.
Applications using Pydantic AI can handle a model’s citations separately from the words of its answer. The language-model application framework released this capability on October 24 in v2.61.0. A TextPart response now holds citation fields: the source and its associated location in the answer. The original wording stays intact, and additional raw metadata from the provider is stored outside the common citation structure.[1]
Bedrock, OpenRouter and xAI citations receive a common representation. Pydantic AI returns citation metadata to its originating provider during replay. The Vercel AI adapter also keeps those fields intact. Provider formats differ: some locate a span within the response, while others refer to a whole block or to a model-written link. Mapping them into a common structure does not independently verify the source’s support for the answer.[1]
The update also changes how applications handle exceptional cases. InputGuardrail covers the user prompt in its redaction and blocks previously permissive failure paths. Prefect keeps each environment associated with the correct run when workspaces execute concurrently. A request for thinking=False uses the minimum supported reasoning level where a model requires reasoning. Cohere tool results containing files generate an explicit error for an unsupported operation, replacing silent omission of the file.[1]
