Tags

Add AI-generated documentation

Change function calling to use a more complex decision making process. This should both reduce token consumption & improve function selection accuracy. Context trimming only affects function selection, full context is still always used when forming responses.

Refactored code to support multiple platform with generalized patterns. Added experimental support for Discord.

Optimizations for token usage and context handling in various cases like image analysis.

Fixed thread/context handling for various use cases with the new, refactored code. Google Search function removed for now because it was getting in the way too often in its unfinished state.

Migrated to OpenAI Python SDK v1.2 & added analyze_images function

Fix problems arising from `the JSON object must be str, bytes or bytearray, not OpenAIObject` errors

Code refactored so that function call second stage replies can be streamed. Channel summaries converted to use this streaming logic. First replies without function calling still lack streaming support.

Image generation logic reworked massively, now using https://gitlab.psychedelic.fi/neuralvisor/middleware/gpu-resources-parallelization for load balancing request to Stable Diffusion web UI

Refactoring continues to streamline the project. Image generation function call now fixed to use only supported SDXL resolutions.

Major refactoring, code now leverages OpenAI function calling for various tasks and it's now much easier to integrate new features.

- Added time-based buffering of streaming responses
- Fixed bugs that prevented captioner and image generation from working due to switching to a non-privileged user in deployment
- Switched CI/CD pipeline to use shell runner, sped up pylint by using --jobs=8. Deployments are now lightning fast!
- Refactor system message & code analysis logic

Added logging module & support for auto-switching to gpt-3.5-turbo-16k if the context limit of gpt-4 is reached.

Main improvement is that context_manager now delegates post handlers to their own tasks, allowing for parallel processing of user messages.