flutter_ai_sdk 1.9.0

A unified wrapper for AI APIs (OpenAI, Anthropic, Google AI, Ollama, Mistral, xAI, DeepSeek, OpenRouter) with streaming, tool calling, context management, and multimodal support.

Changelog

All notable changes to Flutter AI SDK are documented here.

1.9.0 - 2026-07-27

Batch processing

• FlutterAI.submitBatch/getBatchStatus/getBatchResults/ waitForBatchCompletion: new facade methods for asynchronous batch processing at ~50% lower cost, backed by a new BatchProvider interface (kept separate from BaseProvider — only Anthropic and OpenAI have a batch API). Throws AIFeatureNotSupportedError where unsupported.
• Anthropic: POST /v1/messages/batches, polls processing_status, parses the JSONL results file.
• OpenAI: uploads a JSONL input file (purpose=batch), creates the batch job, polls status, downloads and parses the JSONL output file.
• New BatchRequest/BatchJob/BatchResult models and a pollBatchJob helper (exponential backoff, capped) shared by both providers.
• Batch results are not guaranteed to come back in submission order — always match by BatchRequest.customId/BatchResult.customId.
• New AIHttpClient.postMultipart for file uploads.

1.8.0 - 2026-07-27

Embeddings

• FlutterAI.embed/embedBatch: new facade methods generating embedding vectors, backed by a new EmbeddingProvider interface (deliberately separate from BaseProvider — not every chat provider has an embeddings API). Throws AIFeatureNotSupportedError on providers that don't implement it (Anthropic, xAI, DeepSeek, OpenRouter).
• Implemented on OpenAI (text-embedding-3-small), Google AI (gemini-embedding-001, via batchEmbedContents), Ollama (embeddinggemma, local, no API key) and Mistral (mistral-embed, reusing the shared OpenAI-compatible wire format).
• New EmbeddingRequest/EmbeddingResponse models and a cosineSimilarity utility for comparing vectors.

1.7.0 - 2026-07-27

Thinking / reasoning output

• ThinkingContent: new content type surfacing the model's intermediate reasoning, plus AIResponse.thinking/hasThinking and Message.hasThinking convenience accessors.
• AIConfig.thinking (ThinkingConfig, optional ThinkingEffort) opts in per-provider:
  • Anthropic: adaptive extended thinking (thinking: {type: "adaptive"}), effort mapped to output_config.effort.
  • Google AI: generationConfig.thinkingConfig (includeThoughts, thinkingLevel).
  • DeepSeek: thinking: {type: "enabled"}; reasoning_content parsed into ThinkingContent on any OpenAI-compatible provider that returns it, not just DeepSeek.
  • Other providers ignore the configuration.
• Streaming: new StreamChunk.thinkingDelta/isThinkingDelta, kept separate from regular text deltas so accumulation logic isn't polluted.
• Thinking output is informational only — it isn't replayed back to the model on later turns.

1.6.0 - 2026-07-27

More providers

• New providers: Mistral AI, xAI (Grok), DeepSeek and OpenRouter — all expose an OpenAI-compatible chat completions API, so they're implemented as thin subclasses of a new OpenAICompatibleProvider base that reuses the existing OpenAI wire-format mapper unchanged.
• AIResponse.provider on these new providers is correctly tagged (the shared mapper's parseResponse now accepts the calling provider instead of always defaulting to AIProvider.openai).
• DeepSeek has no /v1 segment in its base URL (https://api.deepseek.com), unlike the other three.

1.5.0 - 2026-07-11

Conversation persistence

• Memory: wired up the (previously dormant, unused since v1.0.0) Memory/InMemoryMemory/LimitedMemory classes to ContextManager and FlutterAI — attachMemory/detachMemory for automatic background saves, saveConversation/loadConversation for one-shot save/restore.
• JsonFileMemory: new file-based implementation (one JSON file per conversation); resolves to a stub throwing UnsupportedError on the web, where dart:io is unavailable. No new dependency in the main package.
• ConversationSummary: lightweight metadata via Memory.listSummaries for building a conversation-list UI without loading full histories.
• Split context/memory.dart into a context/persistence/ module (one class per file), matching the rest of the codebase.
• Fixed Conversation/Message JSON round-tripping: updatedAt was dropped on Conversation.fromJson; audio and document content silently turned into empty text on deserialization; base64 images lost their detail level and were mis-parsed as URLs.

1.4.0 - 2026-07-08

Prompt caching & universal document input

• Prompt caching: new AIConfig.promptCaching (PromptCaching, 5 min or 1 h TTL) — explicit cache_control on Anthropic; cache hit counters parsed on every provider (Usage.cachedTokens, new Usage.cacheWriteTokens).
• Documents: DocumentContent now works on every cloud provider — URL sources added on Anthropic, base64 file blocks added on OpenAI (URL passed as a text reference), Google AI unchanged. Support table in the README.

1.3.0 - 2026-07-08

Structured outputs & token counting

• Structured outputs: ResponseFormat.json(schema: ...) now uses each provider's native guaranteed-schema mechanism — OpenAI json_schema (with opt-in strict mode), Anthropic output_config.format, Gemini responseJsonSchema, Ollama schema format.
• Token counting: new countTokens on providers and FlutterAI.countTokens({message}) on the facade — exact server-side counts on Anthropic (/messages/count_tokens) and Google AI (:countTokens); local estimation elsewhere.

1.2.0 - 2026-07-07

Architecture overhaul & new features

• Architecture: full restructuring into one-class-per-file modules — config/, models/content/ (sealed hierarchy as part files), models/tools/; one folder per provider with a dedicated wire-format mapper; shared streaming loop in BaseProvider (template method); new ProviderRegistry factory supporting custom provider registration. The public API is unchanged.
• Tool Runner: automatic agentic tool-calling loop (ToolRunner, ExecutableTool) — parallel tool execution, error feedback to the model, iteration budget, observability callbacks.
• Ollama provider: run local models (Llama, Qwen, Gemma...) with streaming (NDJSON), tools, JSON mode and vision; no API key required.
• Anthropic: consecutive same-role messages are merged, as required by the API's role alternation (fixes parallel tool results).
• Dependencies upgraded for Flutter 3.44.

1.1.0 - 2026-07-06

Model refresh

• Default models updated to current generations: gpt-5.5, claude-opus-4-8, gemini-3.5-flash.
• Model context limits updated (GPT-5.x, Claude 4/5, Gemini 3.x).
• Anthropic provider: never sends temperature and top_p together (rejected by Claude 4+); maps the refusal and model_context_window_exceeded stop reasons.
• Providers can receive an injected HTTP client; FlutterAI accepts a custom provider. Unit tests for all providers and a CI workflow added.
• Dependencies upgraded (mime 2.x, rxdart 0.28, flutter_lints 6).

1.0.0 - 2025-11-30

Initial release

• Unified API for OpenAI, Anthropic and Google AI.
• Streaming with chunk events, context management and memory.
• Multimodal content (text, images, audio, documents).
• Function calling for all providers, typed error handling, token estimation.