We have made a number of refinements to the way ellmer converts
JSON to R data structures. These are breaking changes, although we don’t
expect them to affect much code in the wild. Most importantly, tools are
now invoked with their inputs coerced to standard R data structures
(#461); opt-out by setting convert = FALSE in
tool().
Additionally ellmer now converts NULL to NA
for type_boolean(), type_integer(),
type_number(), and type_string() (#445), and
does a better job with arrays when required = FALSE
(#384).
chat_ functions no longer have a turn
argument. If you need to set the turns, you can now use
Chat$set_turns() (#427). Additionally,
Chat$tokens() has been renamed to
Chat$get_tokens() and returns a data frame of tokens,
correctly aligned to the individual turn. The print method now uses this
to show how many input/output tokens were used by each turn
(#354).
Two new interfaces help you do multiple chats with a single function call:
batch_chat() and
batch_chat_structured() allow you to submit multiple chats
to OpenAI and Anthropic’s batched interfaces. These only guarantee a
response within 24 hours, but are 50% of the price of regular requests
(#143).
parallel_chat() and
parallel_chat_structured() work with any provider and allow
you to submit multiple chats in parallel (#143). This doesn’t give you
any cost savings, but it’s can be much, much faster.
This new family of functions is experimental because I’m not 100% sure that the shape of the user interface is correct, particularly as it pertains to handling errors.
google_upload() lets you upload files to Google
Gemini or Vertex AI (#310). This allows you to work with videos, PDFs,
and other large files with Gemini.
models_google_gemini(),
models_anthropic(), models_openai(),
models_aws_bedrock(), models_ollama() and
models_vllm(), list available models for Google Gemini,
Anthropic, OpenAI, AWS Bedrock, Ollama, and VLLM respectively. Different
providers return different metadata so they are only guaranteed to
return a data frame with at least an id column (#296).
Where possible (currently for Gemini, Anthropic, and OpenAI) we include
known token prices (per million tokens).
interpolate() and friends are now vectorised so you
can generate multiple prompts for (e.g.) a data frame of inputs. They
also now return a specially classed object with a custom print method
(#445). New interpolate_package() makes it easier to
interpolate from prompts stored in the inst/prompts
directory inside a package (#164).
chat_anthropic(), chat_azure(),
chat_openai(), and chat_gemini() now take a
params argument, that coupled with the
params() helper, makes it easy to specify common model
parameters (like seed and temperature) across
providers. Support for other providers will grow as you request it
(#280).
ellmer now tracks the cost of input and output tokens. The cost
is displayed when you print a Chat object, in
tokens_usage(), and with Chat$get_cost(). You
can also request costs in parallel_chat_structured(). We do
our best to accurately compute the cost, but you should treat it as an
estimate rather than the exact price. Unfortunately LLM providers
currently make it very difficult to figure out exactly how much your
queries cost (#203).
We have support for three new providers:
chat_huggingface() for models hosted at https://huggingface.co
(#359, @s-spavound).chat_mistral() for models hosted at https://mistral.ai
(#319).chat_portkey() and models_portkey() for
models hosted at https://portkey.ai (#363, @maciekbanas).We also renamed (with deprecation) a few functions to make the naming scheme more consistent (#382, @gadenbuie):
chat_azure_openai() replaces
chat_azure().chat_aws_bedrock() replaces
chat_bedrock().chat_anthropic() replaces
chat_anthropic().chat_google_gemini() replaces
chat_gemini().We have updated the default model for a couple of providers:
chat_anthropic() uses Sonnet 3.7 (which it also now
displays) (#336).chat_openai() uses GPT-4.1 (#512)New Chat$get_provider() lets you access the
underlying provider object (#202).
Chat$chat_async() and
Chat$stream_async() gain a tool_mode argument
to decide between "sequential" and
"concurrent" tool calling. This is an advanced feature that
primarily affects asynchronous tools (#488, @gadenbuie).
Chat$stream() and Chat$stream_async()
gain support for streaming the additional content types generated during
a tool call with a new stream argument. When
stream = "content" is set, the streaming response yields
Content objects, including the
ContentToolRequest and ContentToolResult
objects used to request and return tool calls (#400, @gadenbuie).
New Chat$on_tool_request() and
$on_tool_result() methods allow you to register callbacks
to run on a tool request or tool result. These callbacks can be used to
implement custom logging or other actions when tools are called, without
modifying the tool function (#493, @gadenbuie).
Chat$chat(echo = "output") replaces the
now-deprecated echo = "text" option. When using
echo = "output", additional output, such as tool requests
and results, are shown as they occur. When echo = "none",
tool call failures are emitted as warnings (#366, @gadenbuie).
ContentToolResult objects can now be returned
directly from the tool() function and now includes
additional information (#398 #399, @gadenbuie):
extra: A list of additional data associated with the
tool result that is not shown to the chatbot.request: The ContentToolRequest that
triggered the tool call. ContentToolResult no longer has an
id property, instead the tool call ID can be retrieved from
request@id.They also include the error condition in the error
property when a tool call fails (#421, @gadenbuie).
ContentToolRequest gains a tool
property that includes the tool() definition when a request
is matched to a tool by ellmer (#423, @gadenbuie).
tool() gains an .annotations argument
that can be created with the tool_annotations() helper.
Tool annotations are described in the Model Context
Protocol and can be used to describe the tool to clients. (#402,
@gadenbuie)
New tool_reject() function can be used to reject a
tool request with an explanation for the rejection reason.
tool_reject() can be called within a tool function or in a
Chat$on_tool_request() callback. In the latter case,
rejecting a tool call will ensure that the tool function is not
evaluated (#490, #493, @gadenbuie).
All requests now set a custom User-Agent that identifies that the requests come from ellmer (#341). The default timeout has been increased to 5 minutes (#451, #321).
chat_anthropic() now supports the thinking content
type (#396), and content_image_url() (#347). It gains a
beta_header argument to opt-in to beta features (#339). It
(along with chat_bedrock()) no longer chokes after
receiving an output that consists only of whitespace (#376). Finally,
chat_anthropic(max_tokens =) is now deprecated in favour of
chat_anthropic(params = ) (#280).
chat_google_gemini() and
chat_google_vertex() gain more ways to authenticate. They
can use GEMINI_API_KEY if set (@t-kalinowski, #513), authenticate with
Google default application credentials (including service accounts, etc)
(#317, @atheriel)
and use viewer-based credentials when running on Posit Connect (#320,
@atheriel).
Authentication with default application credentials requires the
{gargle} package. They now also can now handle responses that include
citation metadata (#358).
chat_ollama() now works with tool()
definitions with optional arguments or empty properties (#342, #348,
@gadenbuie), and
now accepts api_key and consults the
OLLAMA_API_KEY environment variable. This is not needed for
local usage, but enables bearer-token authentication when Ollama is
running behind a reverse proxy (#501, @gadenbuie).
chat_openai(seed =) is now deprecated in favour of
chat_openai(params = ) (#280).
create_tool_def() can now use any Chat instance
(#118, @pedrobtz).
live_browser() now requires {shinychat} v0.2.0 or
later which provides access to the app that powers
live_browser() via shinychat::chat_app(), as
well as a Shiny module for easily including a chat interface for an
ellmer Chat object in your Shiny apps (#397, @gadenbuie). It now
initializes the UI with the messages from the chat turns, rather than
replaying the turns server-side (#381).
Provider gains name and
model fields (#406). These are now reported when you print
a chat object and are used in token_usage().
option(ellmer_verbosity) is no longer supported;
instead use the standard httr2 verbosity functions, such as
httr2::with_verbosity(); these now support streaming
data.
chat_cortex() has been renamed
chat_cortex_analyst() to better disambiguate it from
chat_snowflake() (which also uses “Cortex”) (#275,
@atheriel).
All providers now wait for up to 60s to get the complete
response. You can increase this with, e.g.,
option(ellmer_timeout_s = 120) (#213, #300).
chat_azure(), chat_databricks(),
chat_snowflake(), and chat_cortex_analyst()
now detect viewer-based credentials when running on Posit Connect (#285,
@atheriel).
chat_deepseek() provides support for DeepSeek models
(#242).
chat_openrouter() provides support for models hosted
by OpenRouter (#212).
chat_snowflake() allows chatting with models hosted
through Snowflake’s Cortex
LLM REST API (#258, @atheriel).
content_pdf_file() and
content_pdf_url() allow you to upload PDFs to supported
models. Models that currently support PDFs are Google Gemini and Claude
Anthropic. With help from @walkerke and @andrie (#265).
Chat$get_model() returns the model name
(#299).
chat_azure() has greatly improved support for Azure
Entra ID. API keys are now optional and we can pick up on ambient
credentials from Azure service principals or attempt to use interactive
Entra ID authentication when possible. The broken-by-design
token argument has been deprecated (it could not handle
refreshing tokens properly), but a new credentials argument
can be used for custom Entra ID support when needed instead (for
instance, if you’re trying to use tokens generated by the
AzureAuth package) (#248, #263, #273, #257, @atheriel).
chat_azure() now reports better error messages when
the underlying HTTP requests fail (#269, @atheriel). It now also defaults to
api_version = "2024-10-21" which includes data for
structured data extraction (#271).
chat_bedrock() now handles temporary IAM credentials
better (#261, @atheriel) and chat_bedrock()
gains api_args argument (@billsanto, #295).
chat_databricks() now handles the
DATABRICKS_HOST environment variable correctly whether it
includes an HTTPS prefix or not (#252, @atheriel). It also respects the
SPARK_CONNECT_USER_AGENT environment variable when making
requests (#254, @atheriel).
chat_gemini() now defaults to using the
gemini-2.0-flash model.
print(Chat) no longer wraps long lines, making it
easier to read code and bulleted lists (#246).
New chat_vllm() to chat with models served by vLLM
(#140).
The default chat_openai() model is now
GPT-4o.
New Chat$set_turns() to set turns.
Chat$turns() is now Chat$get_turns().
Chat$system_prompt() is replaced with
Chat$set_system_prompt() and
Chat$get_system_prompt().
Async and streaming async chat are now event-driven and use
later::later_fd() to wait efficiently on curl socket
activity (#157).
New chat_bedrock() to chat with AWS bedrock models
(#50).
New chat$extract_data() uses the structured data API
where available (and tool calling otherwise) to extract data structured
according to a known type specification. You can create specs with
functions type_boolean(), type_integer(),
type_number(), type_string(),
type_enum(), type_array(), and
type_object() (#31).
The general ToolArg() has been replaced by the more
specific type_*() functions. ToolDef() has
been renamed to tool.
content_image_url() will now create inline images
when given a data url (#110).
Streaming ollama results works once again (#117).
Streaming OpenAI results now capture more results, including
logprobs (#115).
New interpolate() and prompt_file()
make it easier to create prompts that are a mix of static text and
dynamic values.
You can find how many tokens you’ve used in the current session
by calling token_usage().
chat_browser() and chat_console() are
now live_browser() and
live_console().
The echo can now be one of three values: “none”,
“text”, or “all”. If “all”, you’ll now see both user and assistant
turns, and all content types will be printed, not just text. When
running in the global environment, echo defaults to “text”,
and when running inside a function it defaults to “none”.
You can now log low-level JSON request/response info by setting
options(ellmer_verbosity = 2).
chat$register_tool() now takes an object created by
Tool(). This makes it a little easier to reuse tool
definitions (#32).
new_chat_openai() is now
chat_openai().
Claude and Gemini are now supported via
chat_claude() and chat_gemini().
The Snowflake Cortex Analyst is now supported via
chat_cortex() (#56).
Databricks is now supported via chat_databricks()
(#152).