This frame introduces an `aggregated_by` field to describe the type of text included
in the frame and allows unspoken groupings of text to be pushed through the pipeline
and treated similar to TTSTextFrames.
Adding support for setting whether or not the text in the TextFrame
should be added to the LLM context (by the LLM assistant aggregator).
Defaults to `True`.
Now:
- For TTS word-by-word output and `TTSSpeakFrames`: `TTSTextFrame`s' have `includes_inter_frame_spaces=False`.
- For all other TTS output: `TTSTextFrame` pass through the received text frames' `includes_inter_frame_spaces` value. So far, this value has always been `True`: LLMs send text chunks already containing all necessary spaces.
- `LLMTextFrame`s set `includes_inter_frame_spaces=False` at init time, per the aforementioned assumption.
Before, when requesting a user image from a function call we had to wait for a
random time before we could indicate the function call was done. This was to
given time to the aggregator to process the image before marking the function
call as completed.
To avoid this, we now wait for the requested image to be received by the LLM
assistant agrgegator (using an asyncio event). Then, we can successfully mark
the function call as completed.
Immediate is the "default", i.e. has the more obvious name (e.g. `ManuallySwitchServiceFrame` v `ManuallySwitchServiceControlFrame`), since that's *probably* what users will want to reach for. Also, the immediate frames are more likely to behave like what we had before the last few commits, where the service switch would always "jump the queue" by having an immediate effect once it hit the `ServiceSwitcher` in the pipeline, jumping ahead of frames in front of it destined for the service.
Removing `VisionImageRawFrame` lets us simplify LLM services' logic, getting us closer to the idealized architecture where all they care about is handling context frames.
This change is in service of getting us closer to ready to deprecate usage of `OpenAILLMContext` and subclasses in favor of the universal `LLMContext`, at least for the traditional text-to-text LLMs.
Why remove `VisionImageRawFrame` rather than deprecate? It's "internal"—only created by `VisionImageFrameAggregator`—and never intended to be used directly by users (it would be difficult to use directly anyway).
Move the logic that was once in `VisionImageFrameAggregator` directly into the examples. Reasoning:
- If `UserImageRequester` is defined in the examples, it makes sense for `UserImageProcessor` to be too, as it’s the flip side of the same coin, so to speak
- The logic is now pretty trivial
- This kind of one-shot, history-less image-describing pipeline shouldn't be common at all; it's ok for it to live in examples rather than as a dedicated class
- In the short term, this enables us to create `LLMContext`s for services that support it and `OpenAILLMContext`s for services that don't yet (AWS)
This commit also adds missing translation from OpenAI-format image context messages to AWS format. Note that this isn't a wasted effort in the face of the upcoming migration to universal `LLMContext`—this work will be reused as it has to be implemented there too.
This lets a text frame bypass TTS while still being included in the LLM
context. Useful for cases like structured text that isn’t meant to be spoken but
should still contribute to context.
1. `ToolsSchema` has been supported in `LLMSetToolsFrame` for a while but wasn't properly reflected in these type hints
2. The new universal `LLMContext` expects tools to be either a `ToolsSchema` or `NOT_GIVEN`.
The same functionality can be achieved using either:
- `LLMMessagesUpdateFrame` with the desired messages, with `run_llm` set to `True`
- `OpenAILLMContextFrame` with a new context initialized with the desired messages
System frames are now queued. Before, system frames could be generated from any
task and would not guarantee any order which was causing undesired
behavior. Also, it was possible to get into some rare recursion issues because
of the way system frames were executed (they were executed in-place, meaning
calling `push_frame()` would finish after the system frame traversed all the
pipeline). This makes system frames more deterministic.
