ai-sdk-ollama
Version:
Vercel AI SDK Provider for Ollama using official ollama-js library
921 lines (651 loc) • 42.1 kB
Markdown
# Changelog
## 4.2.0
### Minor Changes
- a48c0e1: Add MCP Apps examples demonstrating the host flow for ui:// resources
- Node example: connect with mcpAppClientCapabilities, split tools by visibility, generateText with model-visible tools, read and fingerprint ui:// HTML resource, simulate iframe proxy allow/deny
- Browser example: Vite dev server hosts MCP Apps sandbox, proxies read-resource and call-tool requests, renders interactive dashboard via experimental_MCPAppRenderer
- Shared stdio MCP server with dashboard resource, model+app tool, and app-only refresh tool
## 4.1.0
### Minor Changes
- 514304a: Support two AI SDK v7 provider-spec features that were previously unimplemented:
- `includeRawChunks`: `doStream` now emits `{ type: 'raw', rawValue }` parts carrying each unmodified Ollama chunk when the option is set (off by default).
- `usage.raw`: `doGenerate`/`doStream` now expose Ollama's own counters and timings (`prompt_eval_count`, `eval_count`, `total_duration`, `load_duration`, `prompt_eval_duration`, `eval_duration`) on `usage.raw`.
The streaming `finish` part now also carries the same `providerMetadata` that `doGenerate` already returned.
Internally, the two files that had grown past 1000 lines were split into focused modules with no behaviour change. `models/chat-language-model.ts` keeps the `LanguageModelV4` implementation and reliability orchestration, with request mapping in `models/chat-request.ts`, result assembly in `models/chat-result.ts`, and the streaming state machine in `models/chat-stream.ts`. `utils/object-generation-reliability.ts` keeps recovery orchestration, with JSON text repair in `utils/json-text-repair.ts` and schema fallback/coercion in `utils/json-schema-coercion.ts`. `enhancedRepairText`, `cascadeRepairText` and the `RepairTextFunction` type are still exported from the package root; only the internal module paths changed.
## 4.0.0
### Major Changes
- cbf2dd6: Add support for Vercel AI SDK v7 (beta).
- Bump peer dependency to `ai@^7`, and dependencies to `@ai-sdk/provider@4` and `@ai-sdk/provider-utils@5`.
- Migrate all models (chat, embedding, image, reranking) to the `LanguageModelV4` / `EmbeddingModelV4` / `ImageModelV4` / `RerankingModelV4` provider specs, declaring `specificationVersion: 'v4'`.
- Read the new `SharedV4FileData` tagged file-data union for image inputs, and the consolidated `file` tool-result output type (was `image-data` / `image-url` / `file-data`). Fixes multimodal image inputs under v7.
- Update the `tool()` helper usages (web search / web fetch) for the v7 `ToolExecutionOptions` and 3-generic `Tool<INPUT, OUTPUT, CONTEXT>` signature.
- Support the new per-call `reasoning` effort option (`LanguageModelV4CallOptions.reasoning`). The provider maps `'none'`/`'minimal'`/`'low'`/`'medium'`/`'high'`/`'xhigh'`/`'provider-default'` onto Ollama's `think` parameter, overriding the model-level `think` setting, so you can set reasoning effort per request.
BREAKING: This release requires `ai@^7`. AI SDK v7 renamed the re-exported agent callback types `ToolLoopAgentOnFinishCallback` and `ToolLoopAgentOnStepFinishCallback` to `GenerateTextOnFinishCallback` and `GenerateTextOnStepFinishCallback`.
## 4.0.0-beta.0
### Major Changes
- cbf2dd6: Add support for Vercel AI SDK v7 (beta).
- Bump peer dependency to `ai@^7`, and dependencies to `@ai-sdk/provider@4` and `@ai-sdk/provider-utils@5`.
- Migrate all models (chat, embedding, image, reranking) to the `LanguageModelV4` / `EmbeddingModelV4` / `ImageModelV4` / `RerankingModelV4` provider specs, declaring `specificationVersion: 'v4'`.
- Read the new `SharedV4FileData` tagged file-data union for image inputs, and the consolidated `file` tool-result output type (was `image-data` / `image-url` / `file-data`). Fixes multimodal image inputs under v7.
- Update the `tool()` helper usages (web search / web fetch) for the v7 `ToolExecutionOptions` and 3-generic `Tool<INPUT, OUTPUT, CONTEXT>` signature.
- Support the new per-call `reasoning` effort option (`LanguageModelV4CallOptions.reasoning`). The provider maps `'none'`/`'minimal'`/`'low'`/`'medium'`/`'high'`/`'xhigh'`/`'provider-default'` onto Ollama's `think` parameter, overriding the model-level `think` setting, so you can set reasoning effort per request.
BREAKING: This release requires `ai@^7`. AI SDK v7 renamed the re-exported agent callback types `ToolLoopAgentOnFinishCallback` and `ToolLoopAgentOnStepFinishCallback` to `GenerateTextOnFinishCallback` and `GenerateTextOnStepFinishCallback`.
## 3.8.8
### Patch Changes
- b4dde41: Return AI SDK `tool-calls` finish reasons when Ollama responses emit tool calls.
## 3.8.7
### Patch Changes
- 40cf495: chore: update dependencies + migrate to vite 8
Minor/patch dependency refresh via npm-check-updates (--target minor, 3-day cooldown) — no major bumps. Forced vite ^8 and pinned turbo via pnpm override.
## 3.8.4
### Patch Changes
- 6f81c5d: Updated ai-sdk to v6.0.177
## 3.8.3
### Patch Changes
- 6019088: Updated AI SDK to 6.0.154
## 3.8.2
### Patch Changes
- 92ca938: Fix TypeScript 6.0 build failure caused by deprecated baseUrl option in tsup DTS generation
## 3.8.1
### Patch Changes
- aeafdd6: - Update `generateText` to detect and use the stable `output` option instead of the deprecated `experimental_output` when `enableToolsWithStructuredOutput` is enabled, including a two-phase tools + structured output path and an integration test to guard this behavior.
## 3.8.0
### Minor Changes
- ed617f1: - **Fix:** Make `OllamaProviderOptions`, `OllamaChatProviderOptions`, and `OllamaEmbeddingProviderOptions` compatible with the AI SDK's `providerOptions` (`Record<string, JSONObject>`). They are now defined as type aliases from Zod schemas (`z.infer`), so passing e.g. `providerOptions: { ollama: { structuredOutputs: true } }` into `streamText`, `generateText`, or `ToolLoopAgent` type-checks correctly (fixes #548).
- **New:** Export provider option Zod schemas: `ollamaProviderOptionsSchema`, `ollamaChatProviderOptionsSchema`, `ollamaEmbeddingProviderOptionsSchema`, and `ollamaRerankingProviderOptionsSchema` for validation or parsing.
- **New:** Image generation support via `ollama.imageModel(modelId)` and `OllamaImageModel`, using the AI SDK's `generateImage()` and experimental Ollama image models (e.g. `x/z-image-turbo`, `x/flux2-klein`). Supports `providerOptions.ollama` (e.g. `steps`, `negative_prompt`).
- **Change:** Reranking model now uses `parseProviderOptions` from `@ai-sdk/provider-utils` for `providerOptions.ollama`; `OllamaRerankingProviderOptions` is now a type inferred from the exported schema.
## 3.7.1
### Patch Changes
- 70def97: Support media/image content in tool result messages. When a tool returns `output.type === 'content'` with `image-data`, `image-url`, or `file-data` (image) parts, the provider now sends them in the tool message's `images` array to Ollama. Fixes #527.
## 3.7.0
### Minor Changes
- 229a396: Use jsonrepair for tool-call argument parsing
- **Tool-call JSON repair**: `parseToolArguments` now uses [jsonrepair](https://github.com/josdejong/jsonrepair) when `JSON.parse` fails, so malformed tool-argument strings (trailing commas, unquoted keys, single quotes) are repaired instead of returning `{}`. Same logic is used when converting messages in `convertToOllamaChatMessages`.
- **Quoted/double-encoded JSON**: When the model returns a quoted JSON string (e.g. `"{\"query\":\"weather\"}"`), the inner JSON is now parsed so tool arguments are not lost.
- **Exports**: `parseToolArguments` and `resolveToolCallingOptions` are now exported from the main and browser entry points for advanced use.
- **Example**: New `tool-json-repair-example.ts` in examples/node demonstrates tool-argument repair; `json-repair-example.ts` and README updated to reference it.
## 3.6.0
### Minor Changes
- 2cd78cc: **Structured output / object generation**
- **Cascade JSON repair**: Object generation repair now uses a cascade—[jsonrepair](https://github.com/josdejong/jsonrepair) first, then the built-in Ollama-specific repair (`enhancedRepairText`) for edge cases (Python `True`/`None`, URLs with `//`, smart quotes, etc.). Repair remains on by default; use `enableTextRepair: false` or a custom `repairText` to override.
- **Exports**: `cascadeRepairText` and `enhancedRepairText` are exported for advanced use and custom repair pipelines.
- **Reliability**: Type validation after repair so non-JSON or string-wrapped output is rejected when the schema expects an object/array; fallback merge no longer spreads primitives into the result.
- **Docs & examples**: READMEs updated with cascade repair description and links; new `examples/node/src/test-cascade-repair.ts` example for repair behavior (run with `--llm` for LLM object generation).
## 3.5.0
### Minor Changes
- 13d0d64: Fix streaming of reasoning (think) content so that `fullStream` correctly emits `reasoning-start`, `reasoning-delta`, and `reasoning-end` parts with a stable ID when using models with `think: true` (e.g. DeepSeek-R1, Qwen 3). Ensures the AI SDK can aggregate reasoning content and that multiple reasoning chunks from the API are represented as a single reasoning span. Message conversion now includes reasoning parts when converting assistant messages to Ollama format.
## 3.4.0
### Minor Changes
- 8ddf295: Updated to ai-sdk to 6.0.70
## 3.3.0
### Minor Changes
- e8a7dfc: Added comprehensive object generation reliability features to improve the success rate and consistency of structured output generation with Ollama models.
### New Features
- **Enhanced JSON Repair**: Automatic repair of malformed JSON responses including:
- Extraction from markdown code blocks
- Removal of comments and trailing commas
- Fixing smart quotes, single quotes, and unquoted keys
- Handling Python constants (True/False/None)
- Closing incomplete objects and arrays
- URL preservation in strings
- **Retry Logic**: Configurable retry attempts (default: 3) for failed object generations
- **Schema Recovery**: Automatic schema validation and recovery when validation fails
- **Type Mismatch Fixing**: Automatic conversion of type mismatches (e.g., string to number) based on schema requirements
- **Fallback Values**: Optional fallback value generation when all retry attempts fail
- **New Provider Options**:
- `reliableObjectGeneration?: boolean` - Enable/disable reliability features (default: true)
- `objectGenerationOptions?: ObjectGenerationOptions` - Fine-grained control over reliability behavior
### Usage
The reliability features are enabled by default when using `generateObject` or `streamObject`. You can customize behavior via the new `objectGenerationOptions`:
```typescript
const result = await generateObject({
model: ollama('llama3.2'),
schema: mySchema,
prompt: 'Generate user data',
objectGenerationOptions: {
maxRetries: 5,
fixTypeMismatches: true,
enableTextRepair: true,
},
});
```
### Breaking Changes
None. This is a backward-compatible addition.
## 3.2.0
### Minor Changes
- 1553227: Update examples, tests, and documentation to use `generateText` with `Output.object()` instead of deprecated `generateObject`.
- Updated all example files to use the new API pattern
- Updated integration tests to use `generateText` with `Output.object()`
- Updated documentation and README files to reflect the new API
- Updated code comments to reference the new API
This change aligns with AI SDK v6 recommendations. The deprecated `generateObject` function still works but is no longer shown in examples or documentation.
## 3.1.1
### Patch Changes
- 616f652: Fix tool call handling in final chunk
Some models send tool_calls in the final chunk (done: true) instead of during the stream. This change checks the final chunk for tool calls and enqueues them so they aren't missed, preventing tool-calling models from failing silently.
Thanks to [@petrgrishin](https://github.com/petrgrishin) for [PR #391](https://github.com/jagreehal/ai-sdk-ollama/pull/391) - the first contribution to the repo! 🎉
## 3.1.0
### Minor Changes
- bb2dbf3: Updated ai-sdk to v6.0.26
## 3.0.1
### Patch Changes
- fa45199: Fix code formatting issues across 9 source files to ensure consistent code style and pass quality checks.
## 3.0.0
### Major Changes
- 22d327f: Migrate to AI SDK v6 with full compatibility and alignment with official SDK.
## Breaking Changes
- **AI SDK v6 Required**: Now requires `ai@^6.0.0` (previously `ai@^5.0.0`)
- **Removed Custom Utilities**: Deleted custom implementations of utilities now available in official AI SDK:
- `createAsyncIterableStream`, `createStitchableStream`, `createResolvablePromise`, `fixJson`, `parsePartialJsonAsync` removed
- Use official AI SDK exports from `'ai'` instead: `parsePartialJson`, `simulateReadableStream`, `smoothStream`
- **Removed Custom Middleware**: Deleted custom middleware implementations (`wrapLanguageModel`, `defaultSettingsMiddleware`, `extractReasoningMiddleware`, `simulateStreamingMiddleware`)
- All middleware now re-exported from official AI SDK
- **Removed Custom Agent**: Deleted custom `ToolLoopAgent` implementation
- Use official `ToolLoopAgent` from `'ai'` package
- `toolCalled` helper removed, use `hasToolCall` from `'ai'` instead
- **Structured Output API**: `experimental_output` promoted to stable `output` in AI SDK v6
- **Usage Properties**: Token usage now uses `inputTokens`/`outputTokens` instead of `promptTokens`/`completionTokens`
## New Features
- **Full AI SDK v6 Compatibility**: All features align with official AI SDK v6 specification
- **Re-exported Utilities**: Stream utilities, middleware, and agents from official SDK for consistency
- **Improved Type Safety**: Full TypeScript support with LanguageModelV3 specification
- **Reranking Support**: Native reranking API support (AI SDK v6 feature)
- **MCP Support**: Full MCP integration support (OAuth, resources, prompts, elicitation)
## Migration Guide
1. **Update Dependencies**:
```bash
npm install ai-sdk-ollama ai@^6.0.0
```
2. **Update Imports**:
```typescript
// Before
import { ToolLoopAgent, toolCalled } from 'ai-sdk-ollama';
import { parsePartialJson } from 'ai-sdk-ollama';
// After
import { ollama } from 'ai-sdk-ollama';
import { ToolLoopAgent, hasToolCall, parsePartialJson } from 'ai';
```
3. **Update Structured Output**:
```typescript
// Before
experimental_output: Output.object({ schema });
// After
output: Output.object({ schema });
```
4. **Update Stop Conditions**:
```typescript
// Before
stopWhen: [toolCalled('done')];
// After
stopWhen: [hasToolCall('done')];
```
5. **Update Usage Properties**:
```typescript
// Before
result.usage.promptTokens;
result.usage.completionTokens;
// After
result.usage.inputTokens;
result.usage.outputTokens;
result.totalUsage.inputTokens; // For agents
```
## Improvements
- **Reduced Bundle Size**: Removed duplicate implementations, now re-exporting from official SDK
- **Better Maintainability**: Aligned with official SDK to prevent drift
- **Enhanced Type Safety**: Full LanguageModelV3 specification support
- **Consistent APIs**: All utilities follow official AI SDK patterns
## 2.2.0
### Minor Changes
- 72a26e3: ### Fixed
- **Real-time streaming for flow-based UIs**: Fixed issue where `streamText`'s `fullStream` was waiting for completion before emitting events, causing flow interfaces to only receive control events (start, finish) without text or tool call events. The enhanced `fullStream` now streams all events (text-delta, tool-call, tool-result) in real-time as they occur. Resolves [#344](https://github.com/jagreehal/ai-sdk-ollama/issues/344).
### Added
- **`stopWhen` support**: Added support for the `stopWhen` property in both `streamText` and `generateText` functions, allowing users to customize multi-turn tool calling behavior. When not provided and tools are enabled, defaults to `stepCountIs(5)` for multi-turn tool calling.
### Improved
- **AI SDK compatibility**: Enhanced both `streamText` and `generateText` to automatically support all AI SDK properties using `Parameters<typeof _streamText>[0]` type extraction, ensuring 100% forward compatibility with future AI SDK changes without manual updates.
## 2.1.0
### Minor Changes
- 285e4e4: Enhanced `streamText` wrapper to support `fullStream` with synthesis
- Added synthesis support for `fullStream` in addition to `textStream`
- When tool-calling models (like `gpt-oss:120b`) invoke tools without generating text first, the enhanced `fullStream` now automatically synthesizes a response based on tool results
- Emits proper `TextStreamPart` events (`text-start`, `text-delta`, `text-end`) for flow-based UIs
- Fixes issue where flow interfaces only received control events (`start`, `finish`) without any text content when models called tools first
## 2.0.1
### Patch Changes
- 826eb83: Add API key configuration support for cloud Ollama services
- Added `apiKey` parameter to `createOllama` options
- API key is automatically set as `Authorization: Bearer {apiKey}` header
- Existing Authorization headers take precedence over apiKey
- Added header normalization to handle Headers instances, arrays, and plain objects
- Updated README with API key configuration examples for different runtimes (Node.js, Bun, Deno, serverless)
## 2.0.0
### Major Changes
- a23b4a5: ## Breaking Change: Rename `reasoning` to `think`
The `reasoning` parameter in `OllamaChatSettings` has been renamed to `think` to align with Ollama's native API parameter name. This change ensures consistency with the official Ollama API and improves type safety by using `Pick<ChatRequest, 'keep_alive' | 'format' | 'tools' | 'think'>`.
### Migration Guide
**Before:**
```typescript
const model = ollama('gpt-oss:20b-cloud', { reasoning: true });
```
**After:**
```typescript
const model = ollama('gpt-oss:20b-cloud', { think: true });
```
### What Changed
- Removed `reasoning?: boolean` from `OllamaChatSettings`
- Added `think` parameter via `Pick<ChatRequest, 'keep_alive' | 'format' | 'tools' | 'think'>`
- Updated all internal references from `this.settings.reasoning` to `this.settings.think`
- Updated examples and tests to use the new `think` parameter
The functionality remains the same - only the parameter name has changed to match Ollama's API.
## 1.1.0
### Minor Changes
- 201b13b: Add `keep_alive` parameter support and improve type safety
### Added
- **`keep_alive` parameter**: Control how long models stay loaded in memory after requests
- Accepts duration strings (e.g., `"10m"`, `"24h"`), numbers in seconds, negative numbers for indefinite, or `0` to unload immediately
- Works across all chat operations (generate, stream, tool calling, object generation)
### Improved
- **Type safety**: Now uses `Pick<ChatRequest, 'keep_alive' | 'format' | 'tools'>` from the official ollama-js package
- **Type consistency**: `OllamaProviderSettings` extends `Pick<Config, 'headers' | 'fetch'>` and `OllamaEmbeddingSettings` extends `Pick<EmbedRequest, 'dimensions'>`
- **Type exports**: Re-export more types from ollama-js for better developer experience (`ChatRequest`, `EmbedRequest`, `Config`, `ToolCall`, `Tool`, `Message`, `ChatResponse`, `EmbedResponse`)
## 1.0.2
### Patch Changes
- 444c16e: Update AI SDK dependencies to version 5.0.97 and @ai-sdk/react to 2.0.97 in examples. Also update rimraf to 6.1.2 across the project.
## 1.0.1
### Patch Changes
- 2110218: Update dependencies:
- Update `@ai-sdk/provider-utils` to `^3.0.15`
- Update `ai` peer dependency to `^5.0.86`
- Update `vitest` to `^4.0.6`
## 1.0.0
### Major Changes
- 30d2450: ## Fixed Tool Calling Message Conversion
Fixed critical issues with tool calling message conversion that were preventing proper multi-turn conversations:
### Changes
- **Tool result messages**: Now use proper `role: 'tool'` with `tool_name` field instead of `role: 'user'` with `[Tool Result]:` prefix
- **Assistant messages**: Properly include `tool_calls` array for tool execution
- **Finish reason handling**: Returns `'tool-calls'` when tools execute to enable conversation continuation
- **Reliable tool calling**: Disabled by default (`?? false`) for better AI SDK compatibility
- **Test updates**: Updated all test expectations to match new message format
### Impact
- ✅ Standard AI SDK `generateText` and `streamText` now work perfectly with ai-sdk-ollama provider
- ✅ Full compatibility with AI SDK ecosystem and multi-turn tool calling
- ✅ Enhanced functions still provide synthesis and reliability features when needed
- ✅ Users can choose between standard (compatible) or enhanced (reliable) approaches
This ensures both standard AI SDK patterns and enhanced ai-sdk-ollama functions work seamlessly for tool calling scenarios.
## 0.13.0
### Minor Changes
- 7bb9fdd: update deps
## 0.12.0
### Minor Changes
- 34c3e3b: # Enhanced generateText with Automatic Response Synthesis
## What's New
- **Automatic Response Synthesis**: `generateText` now automatically detects when tools execute but return empty responses and synthesizes a comprehensive response using the tool results
- **Prototype Preservation**: Enhanced responses preserve all original AI SDK methods and getters using proper prototype inheritance
- **Experimental Output Support**: New opt-in `enableToolsWithStructuredOutput` feature allows combining tool calling with `experimental_output` (structured output)
- **Type Safety**: Full TypeScript support with proper generic type inference for `experimental_output` schemas
## Breaking Changes
None - this is a backward-compatible enhancement.
## Migration Guide
No migration required. The enhanced behavior is enabled by default and preserves all existing functionality.
### New Features
#### Automatic Synthesis (Default)
```typescript
import { generateText, ollama } from 'ai-sdk-ollama';
// Tools execute but return empty response? No problem!
const result = await generateText({
model: ollama('llama3.2'),
prompt: 'Calculate 25 * 1.08',
tools: { math: mathTool },
});
// Result now includes synthesized text explaining the calculation
console.log(result.text); // "The calculation 25 * 1.08 equals 27..."
```
#### Experimental Output with Tools (Opt-in)
```typescript
import { z } from 'zod';
const result = await generateText({
model: ollama('llama3.2'),
prompt: 'Get weather and format as JSON',
tools: { weather: weatherTool },
toolChoice: 'required',
experimental_output: z.object({
location: z.string(),
temperature: z.number(),
condition: z.string(),
}),
enhancedOptions: {
enableToolsWithStructuredOutput: true, // Opt-in feature
},
});
// Combines tool execution with structured output
console.log(result.experimental_output); // Properly typed schema
```
## Technical Details
- Uses `Object.create()` and `Object.getOwnPropertyDescriptors()` to preserve prototype methods
- Synthesis attempts up to 2 times with configurable prompts
- Maintains full compatibility with AI SDK's type system
- Enhanced responses include combined token usage from both tool execution and synthesis phases
## 0.11.0
### Minor Changes
- fb666c1: Enhanced JSON repair for reliable object generation
- **New Feature**: Added `enhancedRepairText` function that automatically fixes 14+ types of common JSON issues from LLM outputs
- **Improved Reliability**: Enhanced `objectGenerationOptions` with comprehensive JSON repair capabilities including:
- Markdown code block extraction
- Comment removal
- Smart quote fixing
- Unquoted key handling
- Trailing comma removal
- Incomplete object/array completion
- Python constant conversion (True/False/None)
- JSONP wrapper removal
- Single quote to double quote conversion
- URL and escaped quote handling
- Ellipsis pattern resolution
- **New Example**: Added `json-repair-example.ts` demonstrating enhanced repair capabilities
- **Enhanced Configuration**: `enableTextRepair` now defaults to `true` for better out-of-the-box reliability
- **Comprehensive Testing**: Added extensive test suite covering all repair scenarios
- **Backward Compatible**: All existing functionality remains unchanged
## 0.10.1
### Patch Changes
- 7108a12: Fix synthesis conflict between `messages` and `prompt` parameters in `streamText` and `generateText`. The synthesis logic now correctly detects whether the original call used `messages` or `prompt` and constructs the follow-up synthesis call accordingly, preventing "prompt field is not supported when messages is specified" errors when using the `messages` + `system` pattern with tool calling.
## 0.10.0
### Minor Changes
- c6bb667: ## ✨ Browser Example: React + AI Elements Migration
### 🚀 Major Changes
**Browser Example Overhaul:**
- **Migrated from vanilla JS to React**: Complete rewrite using React 19 and modern hooks
- **Added AI Elements integration**: Now uses `@ai-sdk/react` with `useChat` hook and AI Elements components
- **Implemented shadcn/ui components**: Modern, accessible UI components with Tailwind CSS
- **Enhanced streaming architecture**: Uses `toUIMessageStreamResponse()` for proper UI message handling
- **Added comprehensive AI Elements**: 20+ AI-specific components (Message, Response, Conversation, PromptInput, etc.)
**New Features:**
- Real-time model loading and selection from Ollama API
- Dynamic connection status with visual indicators
- Model size formatting and fallback options
- Enhanced error handling and loading states
- Responsive design with modern card-based layout
**Technical Improvements:**
- TypeScript-first implementation with full type safety
- Vite API plugin for seamless Ollama integration
- Proper message streaming with UI message format
- Component-based architecture for better maintainability
### 📦 Dependencies Updated
**AI SDK:**
- `ai`: `^5.0.56` → `^5.0.57`
- `@ai-sdk/react`: `^2.0.57` (new)
**React:**
- `react`: `^19.1.1` (new)
- `react-dom`: `^19.1.1` (new)
- `@types/react`: `^19.1.14` → `^19.1.15`
**Development:**
- `@types/node`: `^24.5.2` → `^24.6.0`
- `@typescript-eslint/*`: `^8.44.1` → `^8.45.0`
- `typescript-eslint`: `^8.44.1` → `^8.45.0`
### 🗂️ File Changes
**Added:**
- `examples/browser/main.tsx` - React entry point
- `examples/browser/src/App.tsx` - Main application component
- `examples/browser/src/components/ai-elements/` - 20 AI Elements components
- `examples/browser/vite-api-plugin.ts` - Vite plugin for Ollama API
- `examples/browser/components/ui/card.tsx` - shadcn/ui card component
**Removed:**
- `examples/browser/main.ts` - Old vanilla JS entry point
**Updated:**
- `examples/browser/package.json` - React dependencies and AI Elements
- `examples/browser/README.md` - Complete rewrite with new architecture
- `examples/browser/index.html` - Updated for React
- `examples/browser/vite.config.js` - Added API plugin integration
## 0.9.0
### Minor Changes
- 35f19de: Add web search and web fetch tools for Ollama integration
- Add `webSearch` tool for performing web searches using Ollama's web search capabilities
- Add `webFetch` tool for fetching web content and URLs
- Support for both browser and Node.js environments
- Comprehensive integration tests and examples
- Updated documentation with usage examples and prerequisites
## 0.8.1
### Patch Changes
- e57ddf2: ## Enhanced Function Renaming & Documentation Improvements
### Function Renaming
- Renamed `generateTextOllama` to `generateText` (enhanced version from ai-sdk-ollama)
- Renamed `streamTextOllama` to `streamText` (enhanced version from ai-sdk-ollama)
- Maintains backward compatibility while providing clearer API
### Documentation Improvements
- **README.md**: Complete restructure with better user flow
- Added Quick Start section with immediate installation and basic example
- Moved value proposition ("Why Choose") section earlier
- Added dedicated "Enhanced Tool Calling" section highlighting main differentiator
- Reorganized examples under "More Examples" for better progression
- Removed redundant content and improved clarity
- **packages/ai-sdk-ollama/README.md**: Applied same improvements
- Consistent structure with main README
- Better user journey from basic to advanced features
- Updated table of contents to match new structure
### Key Benefits
- **Better Developer Experience**: Clearer function names and improved documentation flow
- **Enhanced Tool Calling**: Highlighted the main selling point with dedicated section
- **User-Friendly**: Users can now get started in 30 seconds and understand value immediately
- **Consistent**: Both READMEs now have the same improved structure and flow
## 0.8.0
### Minor Changes
- 7ce6ed0: Enhanced tool calling with reliable wrapper functions
## What's New
- **New Enhanced Wrapper Functions**: Added `generateTextOllama()` and `streamTextOllama()` for guaranteed tool calling reliability
- **Automatic Response Synthesis**: Enhanced functions automatically complete responses when tools are executed but return empty text
- **Configurable Reliability Options**: Control synthesis behavior with `enhancedOptions` parameter
- **Improved Documentation**: Comprehensive examples and comparison tables for standard vs enhanced functions
## Key Features
- **Reliable Tool Calling**: Standard `generateText()` may return empty responses after tool execution. Enhanced wrappers guarantee complete, useful responses every time
- **Backward Compatible**: All existing code continues to work unchanged
- **Production Ready**: Designed for critical applications that can't handle unpredictable empty responses
- **Cross Provider Compatible**: Enhanced functions work with any AI SDK provider
## Breaking Changes
None - this is a purely additive enhancement.
## Migration
No migration required. Existing code works unchanged. To get enhanced reliability:
```typescript
// Before (may return empty text after tool calls)
const { text } = await generateText({
model: ollama('llama3.2'),
tools,
prompt,
});
// After (guaranteed complete responses)
const { text } = await generateTextOllama({
model: ollama('llama3.2'),
tools,
prompt,
});
```
## 0.7.0
### Minor Changes
- 8f0a292: Comprehensive reliability improvements and new Ollama-specific functions
## New Features
### Ollama-Specific AI Functions
- **generateTextOllama**: Enhanced text generation with reliability features
- **generateObjectOllama**: Structured object generation with schema validation
- **streamTextOllama**: Real-time text streaming with tool calling support
- **streamObjectOllama**: Progressive object streaming with reliability features
### Reliability Features
- **Tool Calling Reliability**: Enhanced tool calling with retry logic and parameter normalization
- **Object Generation Reliability**: Schema validation, type mismatch fixing, and fallback generation
- **Streaming Reliability**: Better stop conditions and response synthesis
- **Error Recovery**: Automatic retry mechanisms and graceful error handling
## Examples and Documentation
### New Example Files
- **Comprehensive Demo**: `generate-all-ollama-demo.ts` - showcases all Ollama-specific functions
- **Streaming Demos**: `stream-text-ollama-demo.ts` and `stream-object-ollama-demo.ts`
- **Debug Tools**: `debug-streaming-issue.ts` and `debug-gpt-oss-tools.ts`
- **Testing Examples**: Various test files for different use cases
### Enhanced Examples
- **Browser Example**: Fixed to use `createOllama()` for proper provider configuration
- **Node Examples**: Updated with better error handling and TypeScript compliance
- **Tool Calling**: Comprehensive examples with weather and search tools
## Technical Improvements
### TypeScript Fixes
- Fixed variable naming conflicts in all example files
- Resolved async/await issues with tool calls
- Fixed Zod schema definitions for record types
- Improved type safety across all examples
### API Enhancements
- Better error messages and debugging information
- Enhanced configuration options for reliability features
- Improved streaming performance and reliability
- Better integration with Ollama's native capabilities
## Breaking Changes
None - all changes are backward compatible
## Migration Guide
Existing code continues to work unchanged. New Ollama-specific functions are available as additional options for enhanced reliability.
## 0.6.2
### Patch Changes
- 0e2f392: Fix TypeScript compilation errors in examples
- Fixed variable naming conflicts in stream-vs-generate-test.ts, debug-streaming-issue.ts, generate-all-ollama-demo.ts, stream-object-ollama-demo.ts, and stream-text-ollama-demo.ts
- Fixed undefined variable 'ollamaRaw' in existing-client-example.ts
- Fixed browser example to use createOllama() instead of passing baseURL to ollama() function
- Fixed async tool calls access in streaming examples
- Fixed Zod schema definitions for record types
- All examples now compile and run successfully
## 0.6.1
### Patch Changes
- 80fbf5b: Update to ollama-0.5.18
- Updated Ollama provider to support version 0.5.18
- Enhanced embedding model functionality
- Improved provider implementation with better error handling
- Updated examples to demonstrate new features
## 0.6.0
### Minor Changes
- a7e2377: Allow passing an existing Ollama client to `createOllama` and expose raw client/methods for direct model operations. Closes #67.
## 0.5.5
### Patch Changes
- 12d4f4a: ai ^5.0.35 → ^5.0.37
## 0.5.4
### Patch Changes
- 6e39959: chore: bump ai to ^5.0.30
## 0.5.3
### Patch Changes
- e58649c: chore: bump ai to ^5.0.29
## 0.5.2
### Patch Changes
- 96ec6dc: Updated to ai v5.0.23
## 0.5.1
### Patch Changes
- edb4d47: Updated ai package version to 5.0.15
## 0.5.0
### Minor Changes
- Add browser support with automatic environment detection
- Add browser-specific provider using ollama/browser package
- Implement dual package exports for Node.js and browser environments
- Add comprehensive browser example with Vite and interactive UI
- Update build configuration to generate separate browser and Node.js bundles
- Add browser compatibility tests and examples
- e02f8af: Feature: Browser support and streaming improvements. Closes https://github.com/jagreehal/ai-sdk-ollama/issues/14
- feat(browser): Automatic browser support via `ollama/browser` with dedicated `index.browser` export. Works seamlessly with bundlers and `ai` v5 in browser contexts.
- fix(streaming): Emit trailing `text-delta` on the final `done` chunk to avoid empty streams for models that only flush content at the end. Note: streams may include one additional text chunk now.
- tests: Add `gpt-oss:20b` integration coverage and make prompts/token limits more robust; update unit tests to reflect final text emission on `done`.
- docs/examples: Switch Node examples to per-file `tsx` execution and update READMEs; clarify how to run browser and node examples.
- chore(repo): Monorepo migration (no user-facing API changes), Dependabot config for package folder, and CI refinements.
No breaking changes to the public API.
## 0.4.0
### Minor Changes
- **Reasoning Support**: Added support for reasoning (chain-of-thought) output
- 🧠 **Reasoning Content**: Models that support reasoning can now output their thinking process
- 📝 **Content Types**: Support for `LanguageModelV2Reasoning` content type in both non-streaming and streaming responses
- 🔄 **Streaming Support**: Full streaming support with `reasoning-start`, `reasoning-delta`, and `reasoning-end` events
- ⚙️ **Configurable**: Enable reasoning with `{ reasoning: true }` setting
- 🧪 **Comprehensive Testing**: Added unit tests for reasoning functionality
- 📚 **Documentation**: Updated README and examples with reasoning usage
- 🎯 **Backward Compatible**: Reasoning is disabled by default, existing code continues to work
### Technical Improvements
- Added `reasoning` setting to `OllamaChatSettings` interface
- Enhanced `doGenerate` method to handle `thinking` field from Ollama responses
- Enhanced `doStream` method to emit reasoning stream parts
- Added reasoning support to content conversion logic
- Updated type definitions to include reasoning content types
## 0.3.0
### Minor Changes
- **Auto-Structured Outputs**: Enhanced structured outputs with intelligent auto-detection
- 🎯 **Smart Auto-Detection**: Automatically enables structured outputs when JSON schema is provided
- 🔧 **Backward Compatibility**: Explicit `structuredOutputs: true/false` settings are still respected
- ⚠️ **User-Friendly Warnings**: Clear warnings when auto-enabling structured outputs
- 📚 **Enhanced Documentation**: Updated examples and README with auto-detection guidance
- 🧪 **Comprehensive Testing**: Added integration tests for auto-detection scenarios
- 🛠️ **Improved Developer Experience**: No need to manually set `structuredOutputs: true` for object generation
### Technical Improvements
- Enhanced `shouldEnableStructuredOutputs()` method for intelligent auto-detection
- Improved schema validation and error handling
- Updated README with auto-detection examples and best practices
- Added comprehensive integration tests for edge cases
- Streamlined configuration for common use cases
## 0.2.0
### Minor Changes
- bf0905a: Fix streaming examples and improve TypeScript type checking
- Fix "Stream error: text part not found" by using textStream instead of fullStream for basic streaming
- Fix TypeScript errors in all examples (error handling, index access, undefined checks)
- Remove rootDir restriction in tsconfig.json to enable type checking for examples
- Fix tool call parameter handling and error messages
- Remove deprecated model capabilities and suggestions utilities
- Improve error handling with proper type checking throughout examples
- Update streaming examples to work with AI SDK v5 API changes
All notable changes to this project will be documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [0.1.0] - 2025-08-06
### Added
- 🎉 Initial release of AI SDK Ollama Provider
- ✅ Full support for Vercel AI SDK v5 (`LanguageModelV2` and `EmbeddingModelV2`)
- 🤖 **Chat Language Model** with streaming support
- Text generation with **dual parameter support** (AI SDK standard + native Ollama)
- **Cross provider compatibility** via AI SDK parameters (temperature, maxOutputTokens, etc.)
- **Advanced Ollama features** via native options (mirostat, num_ctx, etc.)
- **Parameter precedence system** - Ollama options override AI SDK parameters
- Structured output support (JSON objects)
- Tool calling capabilities
- Image input support (with compatible models)
- 🔍 **Embedding Model** for text embeddings
- Batch embedding support (up to 2048 texts)
- Support for all Ollama embedding models
- 🧠 **Model Intelligence System**
- Comprehensive model capability database
- Smart model suggestions based on requirements
- Automatic feature detection and validation
- Helpful error messages with actionable recommendations
- 🛠️ **Provider Features**
- Official `ollama` package integration with **direct option pass-through**
- **Future proof**: All current and future Ollama parameters supported automatically
- Custom base URL configuration
- Custom headers support
- Custom fetch implementation
- Comprehensive error handling with custom OllamaError class
- 📦 **Modern Package**
- TypeScript with full type safety
- ES modules with CommonJS compatibility
- Node.js 22+ support
- Clean, organized codebase structure
- 🧪 **Quality Assurance**
- Tests (unit + integration)
- Full TypeScript coverage
- Linting with ESLint + Prettier
- Automated testing with Vitest
- 📚 **Examples & Documentation**
- 8 comprehensive examples covering all features
- Basic chat, streaming, tool calling, embeddings
- Dual parameter demonstrations
- Model capabilities and validation examples
- Comprehensive README with AI SDK v5+ compatibility
- 🖼️ **Image Processing Support**: Complete implementation of AI SDK v5 image handling with Ollama
- Support for image URLs, base64 encoded images, and multiple images
- Full integration with vision models like LLaVA and bakllava
- Streaming support with images
- Mixed content types (text + image + text)
### Technical Details
- Built with AI SDK v5 (`@ai-sdk/provider: ^2.0.0`)
- Uses official Ollama client (`ollama: ^0.5.16`)
- Requires Node.js >=22
- Fully typed with TypeScript 5.9+
- ES module first with CJS fallback
### Supported Models
- **Chat Models**: llama3.2, mistral, phi4-mini, qwen2.5, codellama, and all Ollama chat models
- **Vision Models**: llava, bakllava, llama3.2-vision, minicpm-v
- **Embedding Models**: nomic-embed-text, all-minilm, mxbai-embed-large, and all Ollama embedding models
[0.1.0]: https://github.com/jagreehal/ai-sdk-ollama/releases/tag/v0.1.0