Conversations API, file uploads, Batch API

Breaking Changes:

1. Default models updated to current generation

   • Chat & Responses default model changed from gpt-4.1 to gpt-5.4
   • Embeddings default changed from text-embedding-ada-002 to text-embedding-3-small. Vectors are not comparable across models: pin model = "text-embedding-ada-002" or re-embed existing data if you rely on the old default
   • completion() default changed from the retired text-davinci-003 to gpt-3.5-turbo-instruct

2. completion() no longer sets default temperature or max_tokens

   • Only explicitly provided parameters are sent to the API

3. ResponsesChatSession now forwards all session options with every request

   • Previously only model, instructions, and tools were sent; options like reasoning and temperature were silently dropped
   • Unknown keys stored in session options will now be rejected by the API

New Features:

1. Added Conversations API support for server-side conversation state

   • create_conversation(), conversation(), update_conversation(), delete_conversation()
   • conversation_items(), add_conversation_items(), conversation_item(), delete_conversation_item()
   • ResponsesChatSession accepts a conversation option, used in place of previous_response_id chaining

2. Added file uploads and other multipart endpoints

   • upload_file() and file_content() for uploading & downloading files
   • audio_transcription() for speech to text
   • image_edit() for image editing with one or more input images
   • Encoding handled by the new openai.multipart module

3. Added Batch API support

   • create_batch(), batch(), cancel_batch(), batches()
   • Combined with upload_file() and file_content() for the full request file → poll → results workflow

4. Added support for the developer role in messages for both Chat Completions and Responses APIs

5. Added compatibility clients for more providers

   • openai.compat.anthropic, openai.compat.deepseek, openai.compat.mistral, openai.compat.moonshot, openai.compat.mimo

6. ChatSession now returns the raw API response as a second return value and stores it on chat.last_response

Improvements:

1. Responses API parsing no longer fails on unrecognized output item types

   • Responses containing reasoning, web_search_call, and other non-message items now parse correctly

2. Updated examples and documentation to current models, removed temperature from introductory examples

Modern tool calling support

New Features:

1. Added modern tool-calling support to ChatSession

   • Supports tools, tool_choice, and parallel_tool_calls
   • Tool results can be sent back with role = "tool" and tool_call_id

2. Added per-request parameter overrides for session methods

   • chat:send(...) and chat:generate_response(...) now accept request-specific overrides
   • ResponsesChatSession:send(...) now supports the same pattern

3. Added new tool-calling examples

   • New chat completions example in examples/chat_completions/tools.lua
   • Updated Responses example to show a full tool-call round trip

Improvements:

1. Improved ResponsesChatSession parity with ChatSession

   • Better support for tool-call round trips using function_call_output
   • Session usage and examples now align more closely across both APIs

2. Improved streaming tool-call handling

   • Streaming tool-call deltas are aggregated into complete assistant messages

3. Expanded documentation and test coverage

   • README now documents tool calling and per-request overrides
   • Added tests for tool calls, streaming aggregation, and session overrides

OpenRouter compatibility client

New Features:

1. Added OpenRouter compatibility client
   • require("openai.compat.openrouter") for OpenRouter API
   • See examples/openrouter.lua for usage

Responses API, Gemini compatibility, simplified streaming

Breaking Changes:

1. ChatSession no longer exported from main module

   • Use require("openai.chat_completions").ChatSession instead

2. Default model changed from gpt-3.5-turbo to gpt-4.1

3. chat() streaming callback now receives (delta, raw) instead of (chunk)

   • Use create_chat_completion() for the previous raw chunk behavior

New Features:

1. Added Responses API support

   • create_response(params): Create a new response
   • response(response_id): Get a response by ID
   • delete_response(response_id): Delete a response
   • cancel_response(response_id): Cancel a response

2. Added ResponsesChatSession for conversational state with Responses API

3. Added Gemini compatibility client

   • require("openai.compat.gemini") for Google Gemini API

4. Added create_chat_completion() method for raw streaming behavior

Improvements:

1. SSE parsing moved to dedicated openai.sse module

2. Removed lpeg dependency in favor of simpler line-based SSE parsing

v1.4.3

• Better support for API compatible endpoints like ollm
• Don't send values where defaults are available (fixes issue with o1-preview failing due to temperature parameter)5

Improved support for images & assistants

New Features:

1. Added support for image generation endpoint

   • New method: image_generation(params)

2. Added new endpoints for Assistants API:

   • assistants(): Get list of assistants
   • threads(): Get list of threads
   • thread_messages(thread_id): Get messages for a specific thread
   • delete_thread(thread_id): Delete a specific thread

3. Added new endpoints for managing files:

   • files(): Get list of files
   • file(file_id): Get details of a specific file
   • delete_file(file_id): Delete a specific file

Improvements:

1. Updated content format to support image_url input for GPT Vision

   • Added support for image_url type in content format

2. Made authorization optional

   • Allows usage with local models that don't require API key

v1.2.0 - Support for chat with functions

OpenAI allows sending a list of function declarations that the LLM can decide to call based on the prompt. The function calling interface must be used with chat completions and the gpt-4-0613 or gpt-3.5-turbo-0613 models or later.

This update supports providing a function schema to the initialization of the chat session object.

The function result is then detected when parsing the response from a chat message.

See https://github.com/leafo/lua-openai/blob/main/examples/example5.lua or the README for a complete example.

v1.1.0 - Improved error message output for bad-status errors

Returned error messages will include the message from the server if available.

Full Changelog: v1.0.0...v1.1.0

v1.0.0

Initial release

luarocks install lua-openai

Full Changelog: https://github.com/leafo/lua-openai/commits/v1.0.0