google_generative_ai

• Update the README.md to indicate that this package is deprecated; using the Firebase SDK is the preferred path forward.

We do not expect format exceptions to occur unless the server API is
changing. Change from a plain FormatException to a specific exception
which indicates the two possible resolution strategies - upgrade if not
already on the latest SDK, or file an issue.

• Throw a more relevant exception for server error HTTP responses.

• Add support for model side Code Execution. Enable code execution by
  configuring Tools with a codeExecution argument.
• Use a default role 'model' when a chat response comes back with no role.

• Add support for model side Code Execution. Enable code execution by configuring Tools with a codeExecution argument.
• Use a default role 'model' when a chat response comes back with no role.

Fixes #184

The vertex backend uses a similar data model, but names the field
citations instead of citationSources. Read either field, whichever
exists, and parse it otherwise consistently.

• Allow the Vertex format for citation metadata - read either citationSources or citations keys, whichever exists. Fixes a FormatException when parsing vertex results with citations.

Maintains the internal details of using the client, but offers a general
escape hatch for any unary request by exposing Task.

Add a makeRequest utility method in an extension, and update the
existing unary requests to use it.

• Internal changes to enable reuse in the Vertex SDK. No user visible changes.

• Add support for GenerationConfig.responseSchema for constraining JSON mime
  type output formats.
• Require Dart 3.1.

• Add support for GenerationConfig.responseSchema for constraining JSON mime type output formats.
• Require Dart 3.1.

• Concatenate multiple TextPart into the text String in case the model
  replies with more than one part.
• Fix handling of format argument to Schema.number and Schema.integer.
• Export UsageMetadata.
• Include the full GenerateContentRequest (previously omitted
  safetySettings, generationConfig, tools, toolConfig, and
  systemInstruction) in countTokens requests. This aligns the token count
  with the token count the backend will see in practice for a
  generateContent request.
• Add a text getter on Candidate to make it easer to retrieve the text from
  candidates other than the first in a response.

• Concatenate multiple TextPart into the text String in case the model replies with more than one part.
• Fix handling of format argument to Schema.number and Schema.integer.
• Export UsageMetadata.
• Include the full GenerateContentRequest (previously omitted safetySettings, generationConfig, tools, toolConfig, and systemInstruction) in countTokens requests. This aligns the token count with the token count the backend will see in practice for a generateContent request.
• Add a text getter on Candidate to make it easer to retrieve the text from candidates other than the first in a response.

• Add support for parsing Vertex AI specific fields in CountTokensResponse.
• Add named constructors on Schema for each value type.
• Add GenerationConfig.responseMimeType which supports setting
  'application/json' to force the model to reply with JSON parseable output.
• Add outputDimensionality argument support for embedContent and
  batchEmbedContent.
• Add Content.functionResponses utility to reply to multiple function calls in
  parallel.
• Breaking The Part class is no longer sealed. Exhaustive switches over
  a Part instance will need to add a wildcard case.

• Add support for parsing Vertex AI specific fields in CountTokensResponse.
• Add named constructors on Schema for each value type.
• Add GenerationConfig.responseMimeType which supports setting 'application/json' to force the model to reply with JSON parseable output.
• Add outputDimensionality argument support for embedContent and batchEmbedContent.
• Add Content.functionResponses utility to reply to multiple function calls in parallel.
• Breaking The Part class is no longer sealed. Exhaustive switches over a Part instance will need to add a wildcard case.

Add usageMetadata to GenerateContentResponse (#143)

Add `UsageMetadata` and relevante parsing. Add a `usageMetadata` field
on `GenerateContentResponse`.

Add usage of the new field to the advanced text sample.

Refactor the `GenerateContentResponse` parse method to handled each
field individually at the top level. There is a behavior change for an
error case, but it is not visible through the message formats that are
returned from the backend in practice.

Prepare to publish.

• Add support for parsing the usageMetadata field in GenerateContentResponse messages.

• Use API version v1beta by default.
• Add note to README warning about leaking API keys.

• Use API version v1beta by default.
• Add note to README warning about leaking API keys.

• Add support on content generating methods for overriding "tools" passed when
  the generative model was instantiated.
• Add support for forcing the model to use or not use function calls to generate
  content.

• Add support on content generating methods for overriding "tools" passed when the generative model was instantiated.
• Add support for forcing the model to use or not use function calls to generate content.

• Allow specifying an API version in a requestOptions argument when constructing a model.
• Add support for referring to uploaded files in request contents.
• Add support for passing tools with functions the model may call while generating responses.
• Add support for passing a system instruction when creating the model.
• Breaking Added new subclasses FilePart, FunctionCall, and FunctionResponse of the sealed class Part.

• Update the package version that is sent with the HTTP client name.
• Throw more actionable error objects than FormatException for errors. Errors were previously only correctly parsed in generateContent calls.
• Add support for tuned models.
• Add support for batchEmbedContents calls.

• Remove usage of new SDK features - support older SDKs 3.0 and above.

• Fix an issue parsing generateContent() responses that do not include content (this can occur for some finishReasons).
• Fix an issue parsing generateContent() responses that include citation sources with unpopulated fields
• Add link to ai.google.dev docs.

• Breaking HarmCategory.unknown renamed to unspecified. Removed unused unknown values in the HarmProbability and FinishReason enums.
• Add additional API documentation.
• Update the getting started instructions in the readme.

• Initial release.