Version 1.2.0

1. ResponsesAPI

• Unified function/endpoint combining chat simplicity with tool use and state management.
• Out-of-the-box tool support - It natively supports first-party tools like web_search, file_search, and computer_use, enabling you to invoke these capabilities without additional orchestration.
• Built-in multi-turn conversation chaining-  Use the previous_response_id parameter to link requests into a chain of turns, and the instructions parameter to inject or override system/developer messages on a per-call basis.
• Multimodal input and output - Beyond text, the API accepts images and audio in the same request, letting you build fully multimodal, tool-augmented experiences in a single call.

2. Core and OpenAI Enhancements

• JSON mode handling improvements and fallback json-repair implementation (port of json-repair by @mangiucugna)
• New models: o3, o4-mini, gpt-4.1, and gpt-4.5 series
• Web search support (gpt-4o-search-preview)
• Chat completion parameters expanded (store, reasoning_effort, service_tier, parallel_tool_calls, metadata)
• Streaming and non-streaming IO conversion adapters developed and enhanced
• Token counting updated (jtokkit v1.1.0)
• Usage analytics improved

3. Anthropic Platform Enhancements

• Thinking and streaming settings integration
• Claude 3.7 Sonnet (Vanilla and via Bedrock)
• Citations handling, text blocks encoding improvements
• Caching support
• Enhanced token-limit error handling and mapped Anthropic to OpenAI exceptions
• A ton of new examples (also for Vision and PDF processing)

4. Google Gemini Integration

• New Google Gemini module and models introduced (Gemini 2.5 / 2.0 Pro and Flash)
• Gemini JSON schema handling improved, including OpenAI wrapper integration
• System message caching, domain content management, and usage tracking adjustments
• Btw. Google Vertex now also supports JSON schema mode

5. Perplexity Sonar Integration

• New Perplexity Sonar module and models introduced (sonar-deep-research, reasoning-pro, sonar-pro, etc.)
• Sonar JSON and regex response support, and citations formatting/handling
• OpenAI chat completion wrappers

6. Other Providers: Deepseek, Groq, Grok, FireworksAI, and Novita

• Groq JSON handling unified and adjusted, with deepseek-r1-distill-llama-70b integration
• JSON schema handling for Grok models
• FireworksAI improvements (document inlining), Deepseek model integrations
• Message conversions, filtering thinking tokens, reasoning effort examples
• Llama 4 family
• New Deepseek models (deepseek-r1, DeepSeek-R1 distill) across providers (FireworksAI, Groq, Together AI), plus other models such as Phi-3-vision-128k-instruct, Deepseek-v2-lite-chat, and Llama-3.3-70b
• New chat completion provider: Novita — Welcome to the family!

7. General Project Setup and CI/CD

• Build setup adjustments (build.sbt registrations, env helpers)
• GitHub CI - upload-artifact version bump (to v4)
• Example datasets added (e.g., norway_wiki dump), imports optimized
• README extended with more examples

Version 1.1.2

• Amazon Bedrock support: Anthropic models, payload encoding, and AWS stream decoders

• O1 models support: system/developer messages, JSON mode, etc.

• New non-OpenAI models: Deepseek v3, Gemini 2.0 Flash (w. thinking), and Llama3

• Adapters: chat completion intercept (for logging and benchmarking)

• Other: new examples, Google Vertex upgrade, custom parseJson (for createChatCompletionWithJSON)

Version 1.1.1

Sure, here's the proofread version:

• New Models:

  • Claude 3.5 Sonnet / Haiku, Llama 3.2, grok-beta, gpt-4o-2024-11-20, etc.

• New Providers:

  • Grok and Deepseek (with examples)

• Enhanced Anthropic Integration:

  • Message/block caching
  • System prompt generalized as a "message"
  • PDF block processing

• Better Support for JSON:

  • createChatCompletionWithJSON

• Adapters:

  • Failover models support for chat completion (createChatCompletionWithFailover)
  • Core adapters/wrappers abstracted and relocated to ws-client

• Fixes:

  • Scala 2.13 JSON schema serialization

• Refactoring:

  • ChatCompletionBodyMaker (removed WSClient dependency)
  • Removed org.joda.time.DateTime

Version 1.1.0

• Support for O1 Models: Introduced special handling for settings and conversion of system messages.
• New Endpoints: Added endpoints for runs, run steps, vector stores, vector store files, and vector store file batches.
• Structured Output: Enhanced output with JSON schema and introduced experimental reflective inference from a case class.
• New Providers: Added support for Google Vertex AI, TogetherAI, Cerebras, and Mistral.
• Fixes: Resolved issues related to fine-tune jobs format, handling of log-probs with special characters, etc.
• Examples/Scenarios: A plethora of new examples!

Lastly, a huge thanks 🙏 to our contributors: @bburdiliak, @branislav-burdiliak, @vega113, and @SunPj!

Version 1.0.0

API Updates

• New functions/endpoints for assistants, threads, messages, and files (thanks @branislav-burdiliak) 🎉
• Fine-tuning API updated (checkpoints, explicit dimensions, etc.)
• Chat completion with images (GPT vision), support for logprobs
• Audio endpoint updates - transcription, translation, speech
• New message hierarchy
• Support for tool calling
• New token count subproject (thanks @branislav-burdiliak , @piotrkuczko)

New Models and Providers

• Improved support for Azure OpenAI services
• Support for Azure AI models
• Support for Groq, Fireworks, OctoAI, Ollama, and FastChat (with examples) 🎉
• New Anthropic client project with OpenAI-compatible chat completion adapter (thanks @branislav-burdiliak) 🎉
• NonOpenAIModelId const object holding the most popular non-openai models introduced (e.g. Llama3-70b, Mixtral-8x22B)

Factories and Adapters

• Full, core, and chat completion factories refactored to support easier and more flexible streaming extension
• Service wrappers refactored and generalized to adapters
• Route adapter allowing the use of several providers (e.g., Groq and Anthropic) alongside OpenAI models for chat-completions 🎉
• Other adapters that have been added include round robin/random order load balancing, logging, retry, chat-to-completion, and settings/messages adapter

Bug Fixes and Improvements

• Fixed null bytes handling in TopLogprobInfo JSON format
• WS request - handling slash at the end of URLs
• Made "data:" prefix handling in WSStreamRequestHelper more robust
• MultipartWritable - added extension-implied content type for Azure file upload
• Fixed chat completion's response_format_type

Examples and Tests

• New example project demonstrating usage of most of the features, providers, and adapters (more than 50)
• Tests for token counts and JSON (de)serialization

Breaking Changes/Migrations: ⚠️

• Instead of the deprecated MessageSpec, use typed {System, User, Assistant, Tool}Message
• Instead of createChatFunCompletion with FunctionSpec(s), migrate to createChatToolCompletion using ToolSpec(s)
• Use a new factory OpenAIChatCompletionServiceFactory with a custom URL and auth headers when only chat-completion is supported by a provider
• Migrate streaming service creation from OpenAIServiceStreamedFactory to OpenAIStreamedServiceFactory or OpenAIServiceFactory.withStreaming() (required import: OpenAIStreamedServiceImplicits._)
• Note that OpenAIServiceStreamedExtra.listFineTuneEventsStreamed has been removed
• Migrate OpenAIMultiServiceAdapter.ofRoundRobin and ofRandomOrder to adapters.roundRobin and adapters.randomOrder (where adapter is, e.g., OpenAIServiceAdapters.forFullService)
• Migrate OpenAIRetryServiceAdapter to adapters.retry
• CreateImageSettings -> CreateImageEditSettings

Version 0.5.0

• Fine-tuning functions/endpoints updated to match the latest API:
  • Root URL changed to /fine_tuning/jobs.
  • Fine tune job holder (FineTuneJob) adjusted - e.g. finished_at added, training_file vs training_files (⚠️ Breaking Changes).
  • OpenAIService.createFineTune- settings (CreateFineTuneSettings) adapted - all attributes except n_epochs, model, and suffix dropped: e.g. batch_size, learning_rate_multiplier, prompt_loss_weight (⚠️ Breaking Changes).
  • OpenAIService.listFineTunes - new optional params: after and limit
  • OpenAIService.listFineTuneEvents - new optional params: after and limit
• Azure service factory functions fixed: forAzureWithApiKey and forAzureWithAccessToken.
• New models added: gpt-3.5-turbo-instruct (with 0914 snapshot), davinci-002, and babbage-002.
• Deprecations: OpenAIService.createEdit and OpenAIServiceStreamedExtra.listFineTuneEventsStreamed (not supported anymore).
• Links to the official OpenAI documentation updated.

Version 0.4.1

• Retries:
  • RetryHelpers trait and retry adapter implementing non-blocking retries introduced (thanks @phelps-sg).
  • RetryHelpers - fixing exponential delay, logging, fun(underlying) call, removing unused implicits.
  • New exceptions introduced (OpenAIScalaUnauthorizedException, OpenAIScalaRateLimitException, etc) with error handling/catching plus "registering" those that should be automatically retried.
  • ⚠️ Old/deprecated OpenAIRetryServiceAdapter removed in favor of a new adapter in the client module (migrate if necessary).
• Support for "OpenAI-API-compatible" providers such as FastChat(Vicuna, Alpaca, LLaMA, Koala, fastchat-t5-3b-v1.0, mpt-7b-chat, etc), Azure, or any other similar service with a custom URL. Explicit factory methods for Azure (with API Key or Access Token) provided.
• Scalafmt and Scalafix configured and applied to the entire source base.
• Fixing test dependencies (e.g., akka-actor-testkit-typed, scala-java8-compat) and replacing mockito-scala-scalatest with scalatestplus mockito.
• Configuring GitHub CI with crossScalaVersions, newer Java, etc. (thanks @phelps-sg)

Version 0.4.0

• Function call support for the chat completion with a json output:
  • Function createChatFunCompletion introduced with functions defined by a new data holder FunctionSpec
  • To stay compatible with the standard createChatCompletion function the existing MessageSpec has not been changed
  • Important ⚠️: to harmonize (redundant) response classes, ChatMessage was replaced by the identical MessageSpec class
• New gpt-3.5-turbo and gpt-4 models for the snapshot 0613 added to ModelId. Old models deprecated.
• OpenAIService.close method declared with parentheses (thanks @phelps-sg).
• Scala 2.12 bumped to 2.12.18 and Scala 2.13 to 2.13.11.
• sbt version bumped to 1.9.0 (thanks @phelps-sg)
• Command renamed to EndPoint and Tag renamed to Param.
• OpenAIMultiServiceAdapter - ofRotationType deprecated in favor of a new name ofRoundRobinType
• OpenAIMultiServiceAdapter - ofRandomAccessType deprecated in favor of ofRandomOrderType.

Version 0.3.3

• Scala 3 compatibility refactorings - JsonFormats, Scala-Guice lib upgrade, Guice modules, Enums migrations
• OpenAIService.close function for the underlying ws client exposed
• OpenAI service wrapper with two adapters introduced: 1. Multi-service (load distribution) - rotation or random order, 2. Retry service - custom number of attempts, wait on failure, and logging

Version 0.3.2

• Removing deprecated/unused command/endpoint (engines) and settings (ListFineTuneEventsSettings)
• New exception OpenAIScalaTokenCountExceededException introduced - thrown for completion or chat completion when the number of tokens needed is higher than allowed
• Migrating Command and Tag enums to sealed case objects to simplify Scala 3 compilation
• Resolving the scala-java8-compat eviction problem for Scala 2.13 and Scala 3