Rig v0.27.0 now available!

[joshua-mo-143]

I'm happy to announce that Rig v0.27.0 is now available!

This is the last release of the year, and it has a size to match. To all the contributors: a huge thank you for your hard work. Your persistence and willingness to help has really not only made the maintainers' lives easier - your changes have also made it into the hands of many serious Rust teams making use of LLMs today, as well as many AI teams who may be considering Rust.

This is an absolutely huge patch with lots of changes, so let's dive right in!

New Features

Concurrent tool execution for non-stream agent prompts

You can now execute tools concurrently for non-streamed agent responses!

Simply use the .with_tool_concurrency() function:

let res = agent.prompt("Hello world!").multi_turn(3).with_tool_concurrency(2).await.unwrap();

LSH for in-memory vector store

You can now use Locality Sensitive Hashing with the in-memory vector store! This should significantly speed up search times.

Big thanks to @Dev79844 for contributing this.

Removal of worker crate dependency and feature

The worker crate (which we originally depended on for WASM compatibility) has now been removed thanks to the original work done by @FayCarsons which was then extended to try and remove this.

In place of this, we now have the wasm feature - try it out if you're compiling to WASM!

Other features

• You can now pass the tool call ID to prompt hooks! This should help anyone trying to make an ACP plugin for Zed.
• We now support reqwest-middleware out of the box! Simply enable the required feature (reqwest-middleware).
• Streamed reasoning deltas are now internally expressed as their own field rather than under Reasoning. This should reduce ambiguity when streaming items.
• OpenAI Chat Completions now supports strict mode
• You can now specify the level of reasoning in the Groq integration using GroqAdditionalParameters, serializing it to a JSON value then adding it to your agent or LLM's additional parameters. Note that this means the normal value is now unset by default - you will need to set it if you don't already.

Bug fixes

• Fixed an issue with prompt hooks panicking with no chat history
• Groq now accepts tool calls
• Fixed an issue with some model fields incorrectly being assigned the serde flatten macro attribute
• Fixed a JSON schema issue with OpenAI due to them not supporting the full feature set for the JSON schema
• Fixed an issue for OpenRouter where completion requests were being incorrectly flattened in the wrong fields
• Fixed an issue for Mistral where tool call results were not properly being turned back into tool result items
• Fixed an issue where API keys were not being properly censored

Miscellaneous

• Added a new generic multipart body abstraction which should now make it much easier to work with transcription multipart requests
• Fixed a typo in docs example for AgentBuilder
• Added a Clippy lint to deny todo!(), unimplemented!() and dbg!()
• Flattened the Display impl for tool calls in the case that the inner error also contains the same starting text as the tool call error (ie, a recursive tool call error).
• Added Cairnify to the list of companies using Rig.

Breaking

• If you are using Groq you will need to ensure that you set the reasoning format if you need a separate reasoning field
• If you use the worker feature you will now need to switch to the wasm feature
• If you are using streamed reasoning deltas, depending on the provider you may need to check that you need to switch to reasoning delta items rather than full reasoning items

---

Related related PRs: #1150 #1148

Thanks for reading! We build rig, rig builds us.