• Fixes various issues in the MistralFunctionCallingAdapter wrapper engine for Mistral-Large and Mistral-Small function calling models.
• Fixes an issue in PromptPipeline.explain() where manual examples would not be explained.
• Fixes an issue in PromptPipeline.ensure_bound_function_calls() where passing an ID translator would mutate the ID of the underlying messages

New Features

• Hugging Face: Models loaded through the HuggingEngine now use chat templates for conversational prompting and tool usage if available by default. This should make it much easier to get started with a Hugging Face model in Kani.
• Added the ability to supply a custom tokenizer to the OpenAIEngine (e.g., for using OpenAI-compatible APIs)\

Fixes/Improvements

• Fixed a missing dependency in the llama extra
• The HuggingEngine will now automatically set device_map="auto" if the accelerate library is installed

• Fixes an issue where PromptPipeline.ensure_bound_function_calls() could still let unbound function calls through in cases of particularly long prompts with prefixing system prompts

• Added max_function_rounds to Kani.full_round, Kani.full_round_str, and Kani.full_round_stream:

  The maximum number of function calling rounds to perform in this round. If this number is reached, the model is allowed to generate a final response without any functions defined.
  Default unlimited (continues until model's response does not contain a function call).

• Added __repr__ to engines
• Fixed an issue where Kani could underestimate the token usage for certain OpenAI models using parallel function calling

• Add Kani.add_completion_to_history (useful for token counting, see #29)
• Add an example of an AIFunction definition to PromptPipeline.explain() when a function-related step is included
• Add id_translator arg to PromptPipeline.ensure_bound_function_calls()
• Ensure that OpenAIEngine and HuggingEngine streams yield a completion including prompt and completion token usage
• Various Mistral-7B Instruct v0.3 prompt fixes

• OpenAI: Added support for GPT-4o

New Features

Streaming

kani now supports streaming to print tokens from the engine as they are received! Streaming is designed to be a drop-in superset of the chat_round and full_round methods, allowing you to gradually refactor your code without ever leaving it in a broken state.

To request a stream from the engine, use Kani.chat_round_stream() or Kani.full_round_stream(). These methods will return a StreamManager, which you can use in different ways to consume the stream.

The simplest way to consume the stream is to iterate over it with async for, which will yield a stream of str.

# CHAT ROUND:
stream = ai.chat_round_stream("What is the airspeed velocity of an unladen swallow?")
async for token in stream:
    print(token, end="")
msg = await stream.message()

# FULL ROUND:
async for stream in ai.full_round_stream("What is the airspeed velocity of an unladen swallow?"):
    async for token in stream:
        print(token, end="")
    msg = await stream.message()

After a stream finishes, its contents will be available as a ChatMessage. You can retrieve the final message or BaseCompletion with:

msg = await stream.message()
completion = await stream.completion()

The final ChatMessage may contain non-yielded tokens (e.g. a request for a function call). If the final message or completion is requested before the stream is iterated over, the stream manager will consume the entire stream.

[!TIP]
For compatibility and ease of refactoring, awaiting the stream itself will also return the message, i.e.:

msg = await ai.chat_round_stream("What is the airspeed velocity of an unladen swallow?")

(note the await that is not present in the above examples). This allows you to refactor your code by changing chat_round to chat_round_stream without other changes.

- msg = await ai.chat_round("What is the airspeed velocity of an unladen swallow?")
+ msg = await ai.chat_round_stream("What is the airspeed velocity of an unladen swallow?")

Issue: #30

New Models

kani now has bundled support for the following new models:

Hosted

• Claude 3 (including function calling)

Open Source

• Llama 3 (all sizes)
• Command R and Command R+ (including function calling)
• Mistral-7B and Mixtral-8x7B
• Gemma (all sizes)

Although these models have built-in support, kani supports every chat model available on Hugging Face through transformers or llama.cpp using the new Prompt Pipelines feature (see below)!

Issue: #34

llama.cpp

To use GGUF-quantized versions of models, kani now supports the LlamaCppEngine, which uses the llama-cpp-python library to interface with the llama.cpp library. Any model with a GGUF version is compatible with this engine!

Prompt Pipelines

A prompt pipeline creates a reproducible pipeline for translating a list of ChatMessage into an engine-specific format using fluent-style chaining.

To build a pipeline, create an instance of PromptPipeline() and add steps by calling the step methods documented below. Most pipelines will end with a call to one of the terminals, which translates the intermediate form into the desired output format.

Pipelines come with a built-in explain() method to print a detailed explanation of the pipeline and multiple examples (selected based on the pipeline steps).

Here’s an example using the PromptPipeline to build a LLaMA 2 chat-style prompt:

from kani import PromptPipeline, ChatRole

LLAMA2_PIPELINE = (
    PromptPipeline()

    # System messages should be wrapped with this tag. We'll translate them to USER
    # messages since a system and user message go together in a single [INST] pair.
    .wrap(role=ChatRole.SYSTEM, prefix="<<SYS>>\n", suffix="\n<</SYS>>\n")
    .translate_role(role=ChatRole.SYSTEM, to=ChatRole.USER)

    # If we see two consecutive USER messages, merge them together into one with a
    # newline in between.
    .merge_consecutive(role=ChatRole.USER, sep="\n")
    # Similarly for ASSISTANT, but with a space (kani automatically strips whitespace from the ends of
    # generations).
    .merge_consecutive(role=ChatRole.ASSISTANT, sep=" ")

    # Finally, wrap USER and ASSISTANT messages in the instruction tokens. If our
    # message list ends with an ASSISTANT message, don't add the EOS token
    # (we want the model to continue the generation).
    .conversation_fmt(
        user_prefix="<s>[INST] ",
        user_suffix=" [/INST]",
        assistant_prefix=" ",
        assistant_suffix=" </s>",
        assistant_suffix_if_last="",
    )
)

# We can see what this pipeline does by calling explain()...
LLAMA2_PIPELINE.explain()

# And use it in our engine to build a string prompt for the LLM.
prompt = LLAMA2_PIPELINE(ai.get_prompt())

Integration with HuggingEngine and LlamaCppEngine

Previously, to use a model with a different prompt format than the ones bundled with the library, one had to create a subclass of the HuggingEngine to implement the prompting scheme. With the release of Prompt Pipelines, you can now supply a PromptPipeline in addition to the model ID to use the HuggingEngine directly!

For example, the LlamaEngine (huggingface) is now equivalent to the following:

engine = HuggingEngine(
  "meta-llama/Llama-2-7b-chat-hf",
  prompt_pipeline=LLAMA2_PIPELINE
)

The engine will use the passed pipeline to automatically infer a model's token usage, making it easier than ever to implement new models.

Issue: #32

Improvements

• The OpenAIEngine now uses the official openai-python package. (#31)
  • This means that aiohttp is no longer a direct dependency, and the HTTPClient has been deprecated. For API-based models, we recommend using the httpx library.
• Added arguments to the chat_in_terminal helper to control maximum width, echo user inputs, show function call arguments and results, and other interactive utilities (#33)
• The HuggingEngine can now automatically determine a model's context length.
• Added a warning message if an @ai_function is missing a docstring. (#37)
• Added WrapperEngine to make writing wrapper extensions easier.

Breaking Changes

• All kani models (e.g. ChatMessage) are no longer immutable. This means that you can edit the chat history directly, and token counting will still work correctly.
• As the ctransformers library does not appear to be maintained, we have removed the CTransformersEngine and replaced it with the LlamaCppEngine.
• The arguments to chat_in_terminal (except the first) are now keyword-only.
• The arguments to HuggingEngine (except model_id, max_context_size, and prompt_pipeline) are now keyword-only.
• Generation arguments for OpenAI models now take dictionaries rather than kani.engines.openai.models.* models. (If you aren't sure if you're affected by this, you probably aren't.)

Bug Fixes

• Fixed an issue with Claude 3 and parallel function calling.

It should be a painless upgrade from kani v0.x to kani v1.0! We tried our best to ensure that we didn't break any existing code. If you encounter any issues, please reach out on our Discord.

• Added support for Llama 3
• Added WrapperEngine to make writing wrapper extensions easier
• Refactored internal Command R prompt building for easier runtime extension
• Updated documentation

New Features

Streaming

kani now supports streaming to print tokens from the engine as they are received! Streaming is designed to be a drop-in superset of the chat_round and full_round methods, allowing you to gradually refactor your code without ever leaving it in a broken state.

To request a stream from the engine, use Kani.chat_round_stream() or Kani.full_round_stream(). These methods will return a StreamManager, which you can use in different ways to consume the stream.

The simplest way to consume the stream is to iterate over it with async for, which will yield a stream of str.

# CHAT ROUND:
stream = ai.chat_round_stream("What is the airspeed velocity of an unladen swallow?")
async for token in stream:
    print(token, end="")
msg = await stream.message()

# FULL ROUND:
async for stream in ai.full_round_stream("What is the airspeed velocity of an unladen swallow?"):
    async for token in stream:
        print(token, end="")
    msg = await stream.message()

After a stream finishes, its contents will be available as a ChatMessage. You can retrieve the final message or BaseCompletion with:

msg = await stream.message()
completion = await stream.completion()

The final ChatMessage may contain non-yielded tokens (e.g. a request for a function call). If the final message or completion is requested before the stream is iterated over, the stream manager will consume the entire stream.

[!TIP]
For compatibility and ease of refactoring, awaiting the stream itself will also return the message, i.e.:

msg = await ai.chat_round_stream("What is the airspeed velocity of an unladen swallow?")

(note the await that is not present in the above examples). This allows you to refactor your code by changing chat_round to chat_round_stream without other changes.

- msg = await ai.chat_round("What is the airspeed velocity of an unladen swallow?")
+ msg = await ai.chat_round_stream("What is the airspeed velocity of an unladen swallow?")

Issue: #30

New Models

kani now has bundled support for the following new models:

Hosted

• Claude 3 (including function calling)

Open Source

• Command R and Command R+ (including function calling)
• Mistral-7B and Mixtral-8x7B
• Gemma (all sizes)

Although these models have built-in support, kani supports every chat model available on Hugging Face through transformers or llama.cpp using the new Prompt Pipelines feature (see below)!

Issue: #34

llama.cpp

To use GGUF-quantized versions of models, kani now supports the LlamaCppEngine, which uses the llama-cpp-python library to interface with the llama.cpp library. Any model with a GGUF version is compatible with this engine!

Prompt Pipelines

A prompt pipeline creates a reproducible pipeline for translating a list of ChatMessage into an engine-specific format using fluent-style chaining.

To build a pipeline, create an instance of PromptPipeline() and add steps by calling the step methods documented below. Most pipelines will end with a call to one of the terminals, which translates the intermediate form into the desired output format.

Pipelines come with a built-in explain() method to print a detailed explanation of the pipeline and multiple examples (selected based on the pipeline steps).

Here’s an example using the PromptPipeline to build a LLaMA 2 chat-style prompt:

from kani import PromptPipeline, ChatRole

LLAMA2_PIPELINE = (
    PromptPipeline()

    # System messages should be wrapped with this tag. We'll translate them to USER
    # messages since a system and user message go together in a single [INST] pair.
    .wrap(role=ChatRole.SYSTEM, prefix="<<SYS>>\n", suffix="\n<</SYS>>\n")
    .translate_role(role=ChatRole.SYSTEM, to=ChatRole.USER)

    # If we see two consecutive USER messages, merge them together into one with a
    # newline in between.
    .merge_consecutive(role=ChatRole.USER, sep="\n")
    # Similarly for ASSISTANT, but with a space (kani automatically strips whitespace from the ends of
    # generations).
    .merge_consecutive(role=ChatRole.ASSISTANT, sep=" ")

    # Finally, wrap USER and ASSISTANT messages in the instruction tokens. If our
    # message list ends with an ASSISTANT message, don't add the EOS token
    # (we want the model to continue the generation).
    .conversation_fmt(
        user_prefix="<s>[INST] ",
        user_suffix=" [/INST]",
        assistant_prefix=" ",
        assistant_suffix=" </s>",
        assistant_suffix_if_last="",
    )
)

# We can see what this pipeline does by calling explain()...
LLAMA2_PIPELINE.explain()

# And use it in our engine to build a string prompt for the LLM.
prompt = LLAMA2_PIPELINE(ai.get_prompt())

Integration with HuggingEngine and LlamaCppEngine

Previously, to use a model with a different prompt format than the ones bundled with the library, one had to create a subclass of the HuggingEngine to implement the prompting scheme. With the release of Prompt Pipelines, you can now supply a PromptPipeline in addition to the model ID to use the HuggingEngine directly!

For example, the LlamaEngine (huggingface) is now equivalent to the following:

engine = HuggingEngine(
  "meta-llama/Llama-2-7b-chat-hf",
  prompt_pipeline=LLAMA2_PIPELINE
)

Issue: #32

Improvements

• The OpenAIEngine now uses the official openai-python package. (#31)
  • This means that aiohttp is no longer a direct dependency, and the HTTPClient has been deprecated. For API-based models, we recommend using the httpx library.
• Added arguments to the chat_in_terminal helper to control maximum width, echo user inputs, show function call arguments and results, and other interactive utilities (#33)
• The HuggingEngine can now automatically determine a model's context length.
• Added a warning message if an @ai_function is missing a docstring. (#37)

Breaking Changes

• All kani models (e.g. ChatMessage) are no longer immutable. This means that you can edit the chat history directly, and token counting will still work correctly.
• As the ctransformers library does not appear to be maintained, we have removed the CTransformersEngine and replaced it with the LlamaCppEngine.
• The arguments to chat_in_terminal (except the first) are now keyword-only.
• The arguments to HuggingEngine (except model_id, max_context_size, and prompt_pipeline) are now keyword-only.
• Generation arguments for OpenAI models now take dictionaries rather than kani.engines.openai.models.* models. (If you aren't sure if you're affected by this, you probably aren't.)

It should be a painless upgrade from kani v0.x to kani v1.0! We tried our best to ensure that we didn't break any existing code. If you encounter any issues, please reach out on our Discord.

Most likely the last release before v1.0! This update mostly contains improvements to chat_in_terminal to improve usability in interactive environments like Jupyter Notebook.

Possible Breaking Change

All arguments to chat_in_terminal except the Kani instance must now be keyword arguments; positional arguments are no longer accepted.

For example, chat_in_terminal(ai, 1, "!stop") must now be written chat_in_terminal(ai, rounds=1, stopword="!stop").

Improvements

• You may now specify None as the user query in chat_round and full_round. This will request a new ASSISTANT message without adding a USER message to the chat history (e.g. to continue an unfinished generation).

Added the following keyword args to chat_in_terminal to improve usability in interactive environments like Jupyter Notebook:

• echo: Whether to echo the user's input to stdout after they send a message (e.g. to save in interactive notebook outputs; default false)
• ai_first: Whether the user should send the first message (default) or the model should generate a completion before prompting the user for a message.
• width: The maximum width of the printed outputs (default unlimited).
• show_function_args: Whether to print the arguments the model is calling functions with for each call (default false).
• show_function_returns: Whether to print the results of each function call (default false).
• verbose: Equivalent to setting echo, show_function_args, and show_function_returns to True.

• OpenAI: Added support for Jan 25 models without specifying max_context_length explicitly
• OpenAI: Fixed an issue where the token count for parallel function calls would only consider the first function call

• OpenAI: Fixes an issue where a tool call could have an unbound tool call ID when using always_included_messages near the maximum context length

New Features

• Added support for the Claude API through the AnthropicEngine
  • Currently, this is only for chat messages - we don't yet have access to the new function calling API. We plan to add Claude function calling to Kani as soon as we get access!
• Renamed ToolCallError to a more general PromptError
  • Technically a minor breaking change, though a search of GitHub shows that no one has used ToolCallError yet

Fixes

• Fixed an issue where parallel tool calls could not be validated (thanks @arturoleon!)

• Fixes an issue where emoji in a chat history might cause issues when saving/loading the kani state
• (OpenAI) Fixes an issue where the content field might get omitted in certain requests, causing an API error

• Internal changes to the OpenAIEngine to make extending it easier
• No consumer-facing changes

As of Nov 6, 2023, OpenAI added the ability for a single assistant message to request calling multiple functions in
parallel, and wrapped all function calls in a ToolCall wrapper. In order to add support for this in kani while
maintaining backwards compatibility with OSS function calling models, a ChatMessage now actually maintains the
following internal representation:

ChatMessage.function_call is actually an alias for ChatMessage.tool_calls[0].function. If there is more
than one tool call in the message, when trying to access this property, kani will raise an exception.

To translate kani's FUNCTION message types to OpenAI's TOOL message types, the OpenAIEngine now performs a translation based on binding free tool call IDs to following FUNCTION messages deterministically.

Breaking Changes

To the kani end user, there should be no change to how functions are defined and called. One breaking change was necessary:

• Kani.do_function_call and Kani.handle_function_call_exception now take an additional tool_call_id parameter, which may break overriding functions. The documentation has been updated to encourage overriders to handle *args, **kwargs to prevent this happening again.

New Features

kani can now handle making multiple function calls in parallel if the model requests it. Rather than returning an ASSISTANT message with a single function_call, an engine can now return a list of tool_calls. kani will resolve these tool calls in parallel using asyncio, and add their results to the chat history in the order of the list provided.

Returning a single function_call will continue to work for backwards compatibility.

• OpenAI: The OpenAIClient (internal class used by OpenAIEngine) now expects OpenAIChatMessages as input rather than kani.ChatMessage in order to better type-validate API requests
• OpenAI: Updated token estimation to better reflect current token counts returned by the API

New Feature: Message Parts API

The Message Parts API is intended to provide a foundation for future multimodal LLMs and other engines that require engine-specific input without compromising kani's model-agnostic design. This is accomplished by allowing ChatMessage.content to be a list of MessagePart objects, in addition to a string.

This change is fully backwards-compatible and will not affect existing code.

When writing code with compatibility in mind, the ChatMessage class exposes ChatMessage.text (always a string or None) and ChatMessage.parts (always a list of message parts), which we recommend using instead of ChatMessage.content. These properties are dynamically generated based on the underlying content, and it is safe to mix messages with different content types in a single Kani.

Generally, message part classes are defined by an engine, and consumed by the developer. Message parts can be used in any role’s message - for example, you might use a message part in an assistant message to separate out a chain of thought from a user reply, or in a user message to supply an image to a multimodal model.

For more information, see the Message Parts documentation.

Up next: we're adding support for multimodal vision-language models like LLaVA and GPT-Vision through a kani extension!

Improvements

• LLaMA 2: Improved the prompting in non-strict mode to group consecutive user/system messages into a single [INST] wrapper. See the tests for how kani translates consecutive message types into the LLaMA prompt.
• Other documentation and minor improvements

BREAKING CHANGES

• Kani.full_round now emits every message generated during the round, not just assistant messages
  • This means that you will need to handle FUNCTION messages, and potentially SYSTEM messages from a function exception handler.
  • Kani.full_round_str's default behaviour is unchanged.
• Kani.full_round_str now takes in a message_formatter rather than a function_call_formatter
  • By default, this handler only returns the contents of ASSISTANT messages.
• Kani.do_function_call now returns a FunctionCallResult rather than a bool
  • To migrate any overriding functions, you should change the following:
  • Rather than calling Kani.add_to_history in the override, save the ChatMessage to a variable
  • Update the return value from a boolean to FunctionCallResult(is_model_turn=<old return value>, message=<message from above>)
• Kani.handle_function_call_exception now returns a ExceptionHandleResult rather than a bool
  • To migrate any overriding functions, you should change the following:
  • Rather than calling Kani.add_to_history in the override, save the ChatMessage to a variable
  • Update the return value from a boolean to ExceptionHandleResult(should_retry=<old return value>, message=<message from above>)

Improvements

• Added kani.utils.message_formatters
• Added kani.ExceptionHandleResult and kani.FunctionCallResult
• Documentation improvements

Fixes

• Fixed an issue where ChatMessage.copy_with could cause unset values to appear in JSON serializations

Improvements

• Updated dependencies to allow more recent versions
• The documentation now shows fully-qualified class names in reference sections
• Added .copy_with method to ChatMessage and FunctionCall to make updating chat history easier
• Various documentation updates