Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion package.json
Original file line number Diff line number Diff line change
Expand Up @@ -54,13 +54,21 @@
"github": {
"release": true
},
"hooks": {
"after:bump": "bun scripts/stamp-changelogs.ts ${version}"
},
"plugins": {
"@release-it/conventional-changelog": {
"preset": "angular"
},
"@release-it-plugins/workspaces": {
"workspaces": [
"packages/*"
"packages/android",
"packages/apple-llm",
"packages/core-ai",
"packages/dev-tools",
"packages/llama",
"packages/utils"
]
}
}
Expand Down
22 changes: 22 additions & 0 deletions packages/android/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
# Changelog

Changes to `@react-native-ai/android`.

## Unreleased

This is the first release. It is an AI SDK provider for Gemini Nano built on ML Kit GenAI. ([#367](https://github.com/callstackincubator/ai/pull/367))

### Features

- Generation, streaming and image prompts with Gemini Nano.
- Tool calling, on Gemini Nano 4.
- Transcription and embeddings.
- `getModelInfo()`, which reports whether the model is available, downloadable or downloading. ([#370](https://github.com/callstackincubator/ai/pull/370))
- Options that ML Kit doesn't support return an `unsupported` warning instead of failing the request.

### Migrating from `@react-native-ai/adk`

This package replaces `@react-native-ai/adk`, which is no longer maintained. ([#368](https://github.com/callstackincubator/ai/pull/368))

- Replace `adk()` with `android()`.
- For cloud Gemini, use `@react-native-firebase/ai`, because Google's Gen AI SDK refuses API keys on Android.
3 changes: 2 additions & 1 deletion packages/android/package.json
Original file line number Diff line number Diff line change
Expand Up @@ -47,7 +47,8 @@
},
"homepage": "https://github.com/callstackincubator/ai#readme",
"publishConfig": {
"registry": "https://registry.npmjs.org/"
"registry": "https://registry.npmjs.org/",
"access": "public"
},
"dependencies": {
"@ai-sdk/provider": "^4.0.33",
Expand Down
82 changes: 82 additions & 0 deletions packages/apple-llm/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,82 @@
# Changelog

Changes to `@react-native-ai/apple`. For 0.12.0 and earlier, see the [GitHub releases](https://github.com/callstackincubator/ai/releases).

## Unreleased

### Breaking changes

- Requires AI SDK 7 (`ai@7`). The provider now implements the v4 provider specification. ([#233](https://github.com/callstackincubator/ai/pull/233))
- Removed `apple.isAvailable()`, `apple.transcriptionModel().isAvailable()` and `AppleTranscription.isAvailable()`. Use `getModelInfo()` instead. It rejects with `UNSUPPORTED_OS` on versions before iOS 26. ([#370](https://github.com/callstackincubator/ai/pull/370))
- `AppleEmbeddings.getInfo()` is now `AppleEmbeddings.getModelInfo()`, and the `EmbeddingInfo` type is now `AppleEmbeddingModelInfo`. ([#333](https://github.com/callstackincubator/ai/pull/333))
- `AppleFoundationModels.generateText` now resolves to `{ content, usage }` instead of an array of parts. ([#251](https://github.com/callstackincubator/ai/pull/251))

### Deprecations

- `apple.textEmbeddingModel()` is deprecated in favor of `apple.embeddingModel()`. ([#251](https://github.com/callstackincubator/ai/pull/251))

### Features

- iOS 27 and WWDC26 Apple Intelligence APIs ([#218](https://github.com/callstackincubator/ai/pull/218)):
- Private Cloud Compute, selected with `providerOptions.apple.model` or `createAppleProvider({ model })`.
- `apple.getModelInfo({ locale, model })`, which reports availability, supported languages, context size and capabilities.
- Image prompts. `image/*` file parts are sent to Foundation Models as attachments.
- Image Playground generation through `apple.imageModel()`.
- The Vision `OCRTool` and `BarcodeReaderTool` built-in tools, through `providerOptions.apple.builtInTools`.
- Builds with Xcode 27 and still builds with Xcode 26.
- The `reasoningLevel` provider option on iOS 27. The AI SDK `reasoning` call option maps to it. ([#232](https://github.com/callstackincubator/ai/pull/232), [#233](https://github.com/callstackincubator/ai/pull/233))
- Stable error codes on `AppleLLMError.code`, including `CONTEXT_WINDOW_EXCEEDED` and `RATE_LIMITED`. Foundation Models failures get their own codes. ([#212](https://github.com/callstackincubator/ai/pull/212), [#213](https://github.com/callstackincubator/ai/pull/213), [#232](https://github.com/callstackincubator/ai/pull/232), [#320](https://github.com/callstackincubator/ai/pull/320))
- Token counting with `apple().countTokens(text)` and `AppleFoundationModels.countTokens`. ([#212](https://github.com/callstackincubator/ai/pull/212), [#251](https://github.com/callstackincubator/ai/pull/251))
- Token usage reporting. iOS 27 reports usage natively, and iOS 26.4 counts it with the on-device tokenizer. A length limit is reported as the `length` finish reason. ([#251](https://github.com/callstackincubator/ai/pull/251), [#258](https://github.com/callstackincubator/ai/pull/258))
- Tool call ids, reasoning and tool parts in results and streams, with assistant and tool messages replayed as a structured transcript. ([#240](https://github.com/callstackincubator/ai/pull/240))
- Tool calls stream on iOS 26, and `streamText` works with `Output.object`. ([#251](https://github.com/callstackincubator/ai/pull/251))
- `required` and named `toolChoice`, and the `maxToolCalls` provider option, on iOS 27. ([#265](https://github.com/callstackincubator/ai/pull/265), [#266](https://github.com/callstackincubator/ai/pull/266))
- The `guardrails` option on `createAppleProvider`. Guardrail violations and refusals finish with `content-filter`. ([#228](https://github.com/callstackincubator/ai/pull/228), [#321](https://github.com/callstackincubator/ai/pull/321))
- `getModelInfo()` reports why a model is unavailable. Private Cloud Compute quota is reported too. ([#249](https://github.com/callstackincubator/ai/pull/249), [#329](https://github.com/callstackincubator/ai/pull/329))
- An Expo config plugin that enables Private Cloud Compute: `["@react-native-ai/apple", { "privateCloudCompute": true }]`. ([#279](https://github.com/callstackincubator/ai/pull/279))
- References in generation schemas. ([#314](https://github.com/callstackincubator/ai/pull/314))
- The `seed` call setting. ([#322](https://github.com/callstackincubator/ai/pull/322))
- `prepare()` prewarms the language model. ([#323](https://github.com/callstackincubator/ai/pull/323))
- The content tagging use case. ([#324](https://github.com/callstackincubator/ai/pull/324))
- The Spotlight search built-in tool. ([#328](https://github.com/callstackincubator/ai/pull/328))
- Transcription now includes:
- word timings ([#325](https://github.com/callstackincubator/ai/pull/325))
- live microphone transcription on iOS 27 ([#330](https://github.com/callstackincubator/ai/pull/330))
- streaming transcription of recorded audio ([#337](https://github.com/callstackincubator/ai/pull/337))
- Speech speed is applied as the utterance rate. ([#317](https://github.com/callstackincubator/ai/pull/317))
- `apple.embeddingModel().getModelInfo()`. ([#333](https://github.com/callstackincubator/ai/pull/333))
- `updateTools()` on the provider for replacing registered tools at runtime. ([#196](https://github.com/callstackincubator/ai/pull/196))

### Fixes

- Generation schemas:
- keep every numeric bound ([#252](https://github.com/callstackincubator/ai/pull/252))
- constrain numeric enums and handle fractional values ([#253](https://github.com/callstackincubator/ai/pull/253), [#294](https://github.com/callstackincubator/ai/pull/294))
- give nested schemas unique names ([#277](https://github.com/callstackincubator/ai/pull/277))
- keep properties in declaration order ([#313](https://github.com/callstackincubator/ai/pull/313))
- Tools:
- failures are reported as tool call errors ([#254](https://github.com/callstackincubator/ai/pull/254))
- an unknown tool gets a hint on how to register it ([#283](https://github.com/callstackincubator/ai/pull/283))
- a tool that is no longer registered no longer hangs the native call ([#240](https://github.com/callstackincubator/ai/pull/240))
- Call settings:
- reasoning settings now reach native code ([#240](https://github.com/callstackincubator/ai/pull/240))
- reasoning is dropped for models that can't reason ([#281](https://github.com/callstackincubator/ai/pull/281))
- temperature works without `topP` or `topK` ([#310](https://github.com/callstackincubator/ai/pull/310))
- call settings that Foundation Models ignores now return warnings ([#282](https://github.com/callstackincubator/ai/pull/282))
- `generateText` can be aborted. ([#280](https://github.com/callstackincubator/ai/pull/280))
- Availability is checked for the requested model instead of always the on-device model. ([#232](https://github.com/callstackincubator/ai/pull/232))
- iOS 26 rate limits are reported as `RATE_LIMITED`. ([#262](https://github.com/callstackincubator/ai/pull/262))
- Private Cloud Compute is reported as unavailable when the entitlement is missing. ([#278](https://github.com/callstackincubator/ai/pull/278))
- Image attachments are labeled, including images without a filename, so vision tools can find them. ([#264](https://github.com/callstackincubator/ai/pull/264), [#284](https://github.com/callstackincubator/ai/pull/284))
- Image generation is reported as unsupported on iOS 27. ([#311](https://github.com/callstackincubator/ai/pull/311))
- Transcription:
- reports the language as a language code ([#285](https://github.com/callstackincubator/ai/pull/285))
- resolves languages to supported locales ([#319](https://github.com/callstackincubator/ai/pull/319))
- sends only the audio view's bytes ([#256](https://github.com/callstackincubator/ai/pull/256))
- The current locale is returned as a BCP 47 code without keywords. ([#334](https://github.com/callstackincubator/ai/pull/334))
- Speech:
- uses the language of the call ([#316](https://github.com/callstackincubator/ai/pull/316))
- asks for Personal Voice access only when requested ([#318](https://github.com/callstackincubator/ai/pull/318))
- sends voice quality as a string ([#255](https://github.com/callstackincubator/ai/pull/255))
- settles for whitespace-only text ([#231](https://github.com/callstackincubator/ai/pull/231))
- Embedding text with no tokens returns a zero vector. ([#257](https://github.com/callstackincubator/ai/pull/257))
26 changes: 26 additions & 0 deletions packages/core-ai/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,26 @@
# Changelog

Changes to `@react-native-ai/core-ai`.

## Unreleased

This is the first release. It is an AI SDK provider for Apple Core AI model bundles exported with [`apple/coreai-models`](https://github.com/apple/coreai-models). It requires iOS 27, runs on physical devices only, and has `@react-native-ai/apple` as a peer dependency.

### Features

- Language models ([#219](https://github.com/callstackincubator/ai/pull/219), [#245](https://github.com/callstackincubator/ai/pull/245), [#326](https://github.com/callstackincubator/ai/pull/326), [#338](https://github.com/callstackincubator/ai/pull/338)) support:
- generation and streaming
- persistent sessions
- structured output
- the `reasoning` option
- function tools
- image prompts with vision language bundles
- Image generation with FLUX.2 and Sana Sprint, including FLUX.2 klein 4B on iPhone. ([#219](https://github.com/callstackincubator/ai/pull/219), [#295](https://github.com/callstackincubator/ai/pull/295))
- Video generation with Wan 2.1. ([#309](https://github.com/callstackincubator/ai/pull/309))
- Speech transcription with coreai-models speech bundles. ([#299](https://github.com/callstackincubator/ai/pull/299))
- Text and image embeddings with CLIP exports. ([#302](https://github.com/callstackincubator/ai/pull/302))
- Image upscaling, classification and depth estimation. ([#301](https://github.com/callstackincubator/ai/pull/301))
- Segmentation and object detection, including tracking objects through videos. ([#219](https://github.com/callstackincubator/ai/pull/219), [#303](https://github.com/callstackincubator/ai/pull/303))
- Raw `.aimodel` function calls with tensors and image inputs, through the `CoreAI` native module. ([#300](https://github.com/callstackincubator/ai/pull/300), [#315](https://github.com/callstackincubator/ai/pull/315))
- Model `source` is one string. It takes a path, a `file://` URL, or `bundle://` for models shipped in the app. ([#345](https://github.com/callstackincubator/ai/pull/345))
- An Expo config plugin. It sets the iOS 27 deployment target, links only the `tasks` you list, and has an `increasedMemoryLimit` option. ([#245](https://github.com/callstackincubator/ai/pull/245), [#331](https://github.com/callstackincubator/ai/pull/331))
3 changes: 2 additions & 1 deletion packages/core-ai/package.json
Original file line number Diff line number Diff line change
Expand Up @@ -50,7 +50,8 @@
},
"homepage": "https://github.com/callstackincubator/ai#readme",
"publishConfig": {
"registry": "https://registry.npmjs.org/"
"registry": "https://registry.npmjs.org/",
"access": "public"
},
"dependencies": {
"@ai-sdk/provider": "^4.0.33",
Expand Down
19 changes: 19 additions & 0 deletions packages/dev-tools/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
# Changelog

Changes to `@react-native-ai/dev-tools`. For 0.12.0 and earlier, see the [GitHub releases](https://github.com/callstackincubator/ai/releases).

## Unreleased

### Breaking changes

- AI SDK 7 removed `tracer` from per-call telemetry options, so `experimental_telemetry: { tracer }` no longer sends spans. Register the dev tools tracer once instead ([#233](https://github.com/callstackincubator/ai/pull/233)):

```ts
registerTelemetry(
new OpenTelemetry({ tracer: getAiSdkTracer({ serviceName: 'my-app' }) })
)
```

### Fixes

- `dist/` is included in the npm package, so the Rozenite plugin loads. ([#209](https://github.com/callstackincubator/ai/pull/209))
22 changes: 22 additions & 0 deletions packages/json-ui/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
# Changelog

Changes to `@react-native-ai/json-ui`.

## Unreleased

### Breaking changes

- Requires AI SDK 7 (`ai@7`). ([#233](https://github.com/callstackincubator/ai/pull/233))

### Fixes

- JSON UI tools work with on-device models ([#275](https://github.com/callstackincubator/ai/pull/275)):
- Props are accepted as object literals with unquoted keys or single quotes.
- Invalid props return an error the model can retry instead of ending the generation.
- Type names match regardless of case, and unknown types list the supported ones.
- Headings and labels have their own default styles.
- The system prompt names `getUILayout` instead of a tool that doesn't exist.

## 1.0.0-alpha.1

- First release. It renders a UI that a model builds progressively through tool calls. ([#196](https://github.com/callstackincubator/ai/pull/196))
42 changes: 42 additions & 0 deletions packages/llama/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,42 @@
# Changelog

Changes to `@react-native-ai/llama`. For 0.12.0 and earlier, see the [GitHub releases](https://github.com/callstackincubator/ai/releases).

## Unreleased

### Breaking changes

- Requires AI SDK 7 (`ai@7`) and `llama.rn` 0.13.0-rc.8 or later. ([#233](https://github.com/callstackincubator/ai/pull/233))

### Deprecations

- `textEmbeddingModel()` and `rerankModel()` are deprecated in favor of `embeddingModel()` and `rerankingModel()`. ([#251](https://github.com/callstackincubator/ai/pull/251))

### Features

- Reasoning, structured output and multimodal input for AI SDK 7 ([#239](https://github.com/callstackincubator/ai/pull/239)):
- The `reasoning` call option maps to llama.rn thinking and its token budget.
- Images and audio stay in the same message as their text.
- Token counting with `llama.languageModel(...).countTokens(text)`. ([#251](https://github.com/callstackincubator/ai/pull/251))
- The `projectorImageMaxTokens` option, which caps the tokens an image can use. ([#270](https://github.com/callstackincubator/ai/pull/270))

### Fixes

- Tools are sent with their argument schema, and a named `toolChoice` is supported. ([#239](https://github.com/callstackincubator/ai/pull/239))
- `toolChoice: 'none'` keeps tools out of the prompt. ([#270](https://github.com/callstackincubator/ai/pull/270))
- Aborting `generateText` or a stream stops native generation and rejects. ([#239](https://github.com/callstackincubator/ai/pull/239), [#270](https://github.com/callstackincubator/ai/pull/270))
- Fixed these crashes and freezes ([#270](https://github.com/callstackincubator/ai/pull/270)):
- two concurrent `prepare()` calls
- the second step of a tool call on hybrid models
- aborting a stream
- structured output with thinking turned on
- full-resolution photos
- `maxOutputTokens` no longer carries over from one call to the next. ([#270](https://github.com/callstackincubator/ai/pull/270))
- Structured output returns its JSON. ([#270](https://github.com/callstackincubator/ai/pull/270))
- Speech is returned as 16-bit WAV instead of silence. ([#270](https://github.com/callstackincubator/ai/pull/270))
- `embedMany` works with more than 8 values. ([#270](https://github.com/callstackincubator/ai/pull/270))
- Reasoning is returned before the answer. ([#270](https://github.com/callstackincubator/ai/pull/270))
- Think tokens split correctly when streaming. ([#214](https://github.com/callstackincubator/ai/pull/214), [#216](https://github.com/callstackincubator/ai/pull/216))
- Provider options reach native code. ([#216](https://github.com/callstackincubator/ai/pull/216))
- A download that fails midway is no longer reported as downloaded. ([#270](https://github.com/callstackincubator/ai/pull/270))
- Unknown token usage is reported as `undefined` instead of `0`. ([#251](https://github.com/callstackincubator/ai/pull/251))
14 changes: 14 additions & 0 deletions packages/utils/CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,14 @@
# Changelog

Changes to `@react-native-ai/utils`.

## Unreleased

This is the first release. It holds model-agnostic helpers for AI SDK language models.

### Features

- `wrapLanguageModelWithHistory`, which keeps long conversations within a model's context ([#218](https://github.com/callstackincubator/ai/pull/218), [#263](https://github.com/callstackincubator/ai/pull/263), [#296](https://github.com/callstackincubator/ai/pull/296)):
- It summarizes older messages, using the model's `countTokens` when the model has it.
- It keeps a rolling window of messages.
- It drops completed tool calls from earlier turns.
3 changes: 2 additions & 1 deletion packages/utils/package.json
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,8 @@
"url": "https://github.com/callstackincubator/ai/issues"
},
"publishConfig": {
"registry": "https://registry.npmjs.org/"
"registry": "https://registry.npmjs.org/",
"access": "public"
},
"author": "Mike Grabowski <grabbou@gmail.com>",
"scripts": {
Expand Down
43 changes: 43 additions & 0 deletions scripts/stamp-changelogs.ts
Original file line number Diff line number Diff line change
@@ -0,0 +1,43 @@
// Runs from release-it's after:bump hook. Renames each released package's
// "## Unreleased" section to the new version, so the changelog committed with
// the release says which version shipped each change.
import { existsSync, readFileSync, writeFileSync } from 'node:fs'
import { join } from 'node:path'

const version = process.argv[2]
if (!version) {
console.error('Usage: bun scripts/stamp-changelogs.ts <version>')
process.exit(1)
}

const root = join(import.meta.dir, '..')
const manifest = JSON.parse(readFileSync(join(root, 'package.json'), 'utf8'))
const workspaces: string[] =
manifest['release-it'].plugins['@release-it-plugins/workspaces'].workspaces

const date = new Date().toISOString().slice(0, 10)
const heading = `## ${version} (${date})`

for (const workspace of workspaces) {
const file = join(root, workspace, 'CHANGELOG.md')
if (!existsSync(file)) {
console.error(`${workspace} has no CHANGELOG.md`)
process.exit(1)
}

const changelog = readFileSync(file, 'utf8')
const stamped = changelog.includes('\n## Unreleased\n')
? changelog.replace('\n## Unreleased\n', `\n${heading}\n`)
: changelog.replace(
/\n## /,
`\n${heading}\n\nNo changes. Released to keep package versions in sync.\n\n## `
)

if (stamped === changelog) {
console.error(`Could not stamp ${workspace}/CHANGELOG.md`)
process.exit(1)
}

writeFileSync(file, stamped)
console.log(`${workspace}/CHANGELOG.md: ${heading}`)
}
Loading