ZIm/crates/language_models/src/provider/open_ai_compatible.rs
Cale Sennett c2f0df9b8e Add capabilities to OpenAI-compatible model settings (#36370)
### TL;DR
* Adds `capabilities` configuration for OpenAI-compatible models
* Relates to
https://github.com/zed-industries/zed/issues/36215#issuecomment-3193920491

### Summary
This PR introduces support for configuring model capabilities for
OpenAI-compatible language models. The implementation addresses the
issue that not all OpenAI-compatible APIs support the same features -
for example, Cerebras' API explicitly does not support
`parallel_tool_calls` as documented in their [OpenAI compatibility
guide](https://inference-docs.cerebras.ai/resources/openai#currently-unsupported-openai-features).

### Changes

1. **Model Capabilities Structure**:
- Added `ModelCapabilityToggles` struct for UI representation with
boolean toggle states
- Implemented proper parsing of capability toggles into
`ModelCapabilities`

2. **UI Updates**:
- Modified the "Add LLM Provider" modal to include checkboxes for each
capability
- Each OpenAI-compatible model can now be configured with its specific
capabilities through the UI

3. **Configuration File Structure**:
- Updated the settings schema to support a `capabilities` object for
each `openai_compatible` model
- Each capability (`tools`, `images`, `parallel_tool_calls`,
`prompt_cache_key`) can be individually specified per model

### Example Configuration

```json
{
  "openai_compatible": {
    "Cerebras": {
      "api_url": "https://api.cerebras.ai/v1",
      "available_models": [
        {
          "name": "gpt-oss-120b",
          "max_tokens": 131000,
          "capabilities": {
            "tools": true,
            "images": false,
            "parallel_tool_calls": false,
            "prompt_cache_key": false
          }
        }
      ]
    }
  }
}
```

### Tests Added

- Added tests to verify default capability values are correctly applied
- Added tests to verify that deselected toggles are properly parsed as
`false`
- Added tests to verify that mixed capability selections work correctly

Thanks to @osyvokon for the desired `capabilities` configuration
structure!


Release Notes:

- OpenAI-compatible models now have configurable capabilities (#36370;
thanks @calesennett)

---------

Co-authored-by: Oleksiy Syvokon <oleksiy@zed.dev>
2025-08-18 12:35:08 +03:00

550 lines
18 KiB
Rust

use anyhow::{Context as _, Result, anyhow};
use credentials_provider::CredentialsProvider;
use convert_case::{Case, Casing};
use futures::{FutureExt, StreamExt, future::BoxFuture};
use gpui::{AnyView, App, AsyncApp, Context, Entity, Subscription, Task, Window};
use http_client::HttpClient;
use language_model::{
AuthenticateError, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,
LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,
LanguageModelToolChoice, RateLimiter,
};
use menu;
use open_ai::{ResponseStreamEvent, stream_completion};
use schemars::JsonSchema;
use serde::{Deserialize, Serialize};
use settings::{Settings, SettingsStore};
use std::sync::Arc;
use ui::{ElevationIndex, Tooltip, prelude::*};
use ui_input::SingleLineInput;
use util::ResultExt;
use crate::AllLanguageModelSettings;
use crate::provider::open_ai::{OpenAiEventMapper, into_open_ai};
#[derive(Default, Clone, Debug, PartialEq)]
pub struct OpenAiCompatibleSettings {
pub api_url: String,
pub available_models: Vec<AvailableModel>,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
pub struct AvailableModel {
pub name: String,
pub display_name: Option<String>,
pub max_tokens: u64,
pub max_output_tokens: Option<u64>,
pub max_completion_tokens: Option<u64>,
#[serde(default)]
pub capabilities: ModelCapabilities,
}
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
pub struct ModelCapabilities {
pub tools: bool,
pub images: bool,
pub parallel_tool_calls: bool,
pub prompt_cache_key: bool,
}
impl Default for ModelCapabilities {
fn default() -> Self {
Self {
tools: true,
images: false,
parallel_tool_calls: false,
prompt_cache_key: false,
}
}
}
pub struct OpenAiCompatibleLanguageModelProvider {
id: LanguageModelProviderId,
name: LanguageModelProviderName,
http_client: Arc<dyn HttpClient>,
state: gpui::Entity<State>,
}
pub struct State {
id: Arc<str>,
env_var_name: Arc<str>,
api_key: Option<String>,
api_key_from_env: bool,
settings: OpenAiCompatibleSettings,
_subscription: Subscription,
}
impl State {
fn is_authenticated(&self) -> bool {
self.api_key.is_some()
}
fn reset_api_key(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
let credentials_provider = <dyn CredentialsProvider>::global(cx);
let api_url = self.settings.api_url.clone();
cx.spawn(async move |this, cx| {
credentials_provider
.delete_credentials(&api_url, &cx)
.await
.log_err();
this.update(cx, |this, cx| {
this.api_key = None;
this.api_key_from_env = false;
cx.notify();
})
})
}
fn set_api_key(&mut self, api_key: String, cx: &mut Context<Self>) -> Task<Result<()>> {
let credentials_provider = <dyn CredentialsProvider>::global(cx);
let api_url = self.settings.api_url.clone();
cx.spawn(async move |this, cx| {
credentials_provider
.write_credentials(&api_url, "Bearer", api_key.as_bytes(), &cx)
.await
.log_err();
this.update(cx, |this, cx| {
this.api_key = Some(api_key);
cx.notify();
})
})
}
fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
if self.is_authenticated() {
return Task::ready(Ok(()));
}
let credentials_provider = <dyn CredentialsProvider>::global(cx);
let env_var_name = self.env_var_name.clone();
let api_url = self.settings.api_url.clone();
cx.spawn(async move |this, cx| {
let (api_key, from_env) = if let Ok(api_key) = std::env::var(env_var_name.as_ref()) {
(api_key, true)
} else {
let (_, api_key) = credentials_provider
.read_credentials(&api_url, &cx)
.await?
.ok_or(AuthenticateError::CredentialsNotFound)?;
(
String::from_utf8(api_key).context("invalid {PROVIDER_NAME} API key")?,
false,
)
};
this.update(cx, |this, cx| {
this.api_key = Some(api_key);
this.api_key_from_env = from_env;
cx.notify();
})?;
Ok(())
})
}
}
impl OpenAiCompatibleLanguageModelProvider {
pub fn new(id: Arc<str>, http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
fn resolve_settings<'a>(id: &'a str, cx: &'a App) -> Option<&'a OpenAiCompatibleSettings> {
AllLanguageModelSettings::get_global(cx)
.openai_compatible
.get(id)
}
let state = cx.new(|cx| State {
id: id.clone(),
env_var_name: format!("{}_API_KEY", id).to_case(Case::Constant).into(),
settings: resolve_settings(&id, cx).cloned().unwrap_or_default(),
api_key: None,
api_key_from_env: false,
_subscription: cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
let Some(settings) = resolve_settings(&this.id, cx) else {
return;
};
if &this.settings != settings {
this.settings = settings.clone();
cx.notify();
}
}),
});
Self {
id: id.clone().into(),
name: id.into(),
http_client,
state,
}
}
fn create_language_model(&self, model: AvailableModel) -> Arc<dyn LanguageModel> {
Arc::new(OpenAiCompatibleLanguageModel {
id: LanguageModelId::from(model.name.clone()),
provider_id: self.id.clone(),
provider_name: self.name.clone(),
model,
state: self.state.clone(),
http_client: self.http_client.clone(),
request_limiter: RateLimiter::new(4),
})
}
}
impl LanguageModelProviderState for OpenAiCompatibleLanguageModelProvider {
type ObservableEntity = State;
fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
Some(self.state.clone())
}
}
impl LanguageModelProvider for OpenAiCompatibleLanguageModelProvider {
fn id(&self) -> LanguageModelProviderId {
self.id.clone()
}
fn name(&self) -> LanguageModelProviderName {
self.name.clone()
}
fn icon(&self) -> IconName {
IconName::AiOpenAiCompat
}
fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
self.state
.read(cx)
.settings
.available_models
.first()
.map(|model| self.create_language_model(model.clone()))
}
fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
None
}
fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
self.state
.read(cx)
.settings
.available_models
.iter()
.map(|model| self.create_language_model(model.clone()))
.collect()
}
fn is_authenticated(&self, cx: &App) -> bool {
self.state.read(cx).is_authenticated()
}
fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
self.state.update(cx, |state, cx| state.authenticate(cx))
}
fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView {
cx.new(|cx| ConfigurationView::new(self.state.clone(), window, cx))
.into()
}
fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
self.state.update(cx, |state, cx| state.reset_api_key(cx))
}
}
pub struct OpenAiCompatibleLanguageModel {
id: LanguageModelId,
provider_id: LanguageModelProviderId,
provider_name: LanguageModelProviderName,
model: AvailableModel,
state: gpui::Entity<State>,
http_client: Arc<dyn HttpClient>,
request_limiter: RateLimiter,
}
impl OpenAiCompatibleLanguageModel {
fn stream_completion(
&self,
request: open_ai::Request,
cx: &AsyncApp,
) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>>>
{
let http_client = self.http_client.clone();
let Ok((api_key, api_url)) = cx.read_entity(&self.state, |state, _| {
(state.api_key.clone(), state.settings.api_url.clone())
}) else {
return futures::future::ready(Err(anyhow!("App state dropped"))).boxed();
};
let provider = self.provider_name.clone();
let future = self.request_limiter.stream(async move {
let Some(api_key) = api_key else {
return Err(LanguageModelCompletionError::NoApiKey { provider });
};
let request = stream_completion(http_client.as_ref(), &api_url, &api_key, request);
let response = request.await?;
Ok(response)
});
async move { Ok(future.await?.boxed()) }.boxed()
}
}
impl LanguageModel for OpenAiCompatibleLanguageModel {
fn id(&self) -> LanguageModelId {
self.id.clone()
}
fn name(&self) -> LanguageModelName {
LanguageModelName::from(
self.model
.display_name
.clone()
.unwrap_or_else(|| self.model.name.clone()),
)
}
fn provider_id(&self) -> LanguageModelProviderId {
self.provider_id.clone()
}
fn provider_name(&self) -> LanguageModelProviderName {
self.provider_name.clone()
}
fn supports_tools(&self) -> bool {
self.model.capabilities.tools
}
fn supports_images(&self) -> bool {
self.model.capabilities.images
}
fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
match choice {
LanguageModelToolChoice::Auto => self.model.capabilities.tools,
LanguageModelToolChoice::Any => self.model.capabilities.tools,
LanguageModelToolChoice::None => true,
}
}
fn telemetry_id(&self) -> String {
format!("openai/{}", self.model.name)
}
fn max_token_count(&self) -> u64 {
self.model.max_tokens
}
fn max_output_tokens(&self) -> Option<u64> {
self.model.max_output_tokens
}
fn count_tokens(
&self,
request: LanguageModelRequest,
cx: &App,
) -> BoxFuture<'static, Result<u64>> {
let max_token_count = self.max_token_count();
cx.background_spawn(async move {
let messages = super::open_ai::collect_tiktoken_messages(request);
let model = if max_token_count >= 100_000 {
// If the max tokens is 100k or more, it is likely the o200k_base tokenizer from gpt4o
"gpt-4o"
} else {
// Otherwise fallback to gpt-4, since only cl100k_base and o200k_base are
// supported with this tiktoken method
"gpt-4"
};
tiktoken_rs::num_tokens_from_messages(model, &messages).map(|tokens| tokens as u64)
})
.boxed()
}
fn stream_completion(
&self,
request: LanguageModelRequest,
cx: &AsyncApp,
) -> BoxFuture<
'static,
Result<
futures::stream::BoxStream<
'static,
Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
>,
LanguageModelCompletionError,
>,
> {
let request = into_open_ai(
request,
&self.model.name,
self.model.capabilities.parallel_tool_calls,
self.model.capabilities.prompt_cache_key,
self.max_output_tokens(),
None,
);
let completions = self.stream_completion(request, cx);
async move {
let mapper = OpenAiEventMapper::new();
Ok(mapper.map_stream(completions.await?).boxed())
}
.boxed()
}
}
struct ConfigurationView {
api_key_editor: Entity<SingleLineInput>,
state: gpui::Entity<State>,
load_credentials_task: Option<Task<()>>,
}
impl ConfigurationView {
fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
let api_key_editor = cx.new(|cx| {
SingleLineInput::new(
window,
cx,
"000000000000000000000000000000000000000000000000000",
)
});
cx.observe(&state, |_, _, cx| {
cx.notify();
})
.detach();
let load_credentials_task = Some(cx.spawn_in(window, {
let state = state.clone();
async move |this, cx| {
if let Some(task) = state
.update(cx, |state, cx| state.authenticate(cx))
.log_err()
{
// We don't log an error, because "not signed in" is also an error.
let _ = task.await;
}
this.update(cx, |this, cx| {
this.load_credentials_task = None;
cx.notify();
})
.log_err();
}
}));
Self {
api_key_editor,
state,
load_credentials_task,
}
}
fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
let api_key = self
.api_key_editor
.read(cx)
.editor()
.read(cx)
.text(cx)
.trim()
.to_string();
// Don't proceed if no API key is provided and we're not authenticated
if api_key.is_empty() && !self.state.read(cx).is_authenticated() {
return;
}
let state = self.state.clone();
cx.spawn_in(window, async move |_, cx| {
state
.update(cx, |state, cx| state.set_api_key(api_key, cx))?
.await
})
.detach_and_log_err(cx);
cx.notify();
}
fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
self.api_key_editor.update(cx, |input, cx| {
input.editor.update(cx, |editor, cx| {
editor.set_text("", window, cx);
});
});
let state = self.state.clone();
cx.spawn_in(window, async move |_, cx| {
state.update(cx, |state, cx| state.reset_api_key(cx))?.await
})
.detach_and_log_err(cx);
cx.notify();
}
fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
!self.state.read(cx).is_authenticated()
}
}
impl Render for ConfigurationView {
fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
let env_var_set = self.state.read(cx).api_key_from_env;
let env_var_name = self.state.read(cx).env_var_name.clone();
let api_key_section = if self.should_render_editor(cx) {
v_flex()
.on_action(cx.listener(Self::save_api_key))
.child(Label::new("To use Zed's agent with an OpenAI-compatible provider, you need to add an API key."))
.child(
div()
.pt(DynamicSpacing::Base04.rems(cx))
.child(self.api_key_editor.clone())
)
.child(
Label::new(
format!("You can also assign the {env_var_name} environment variable and restart Zed."),
)
.size(LabelSize::Small).color(Color::Muted),
)
.into_any()
} else {
h_flex()
.mt_1()
.p_1()
.justify_between()
.rounded_md()
.border_1()
.border_color(cx.theme().colors().border)
.bg(cx.theme().colors().background)
.child(
h_flex()
.gap_1()
.child(Icon::new(IconName::Check).color(Color::Success))
.child(Label::new(if env_var_set {
format!("API key set in {env_var_name} environment variable.")
} else {
"API key configured.".to_string()
})),
)
.child(
Button::new("reset-api-key", "Reset API Key")
.label_size(LabelSize::Small)
.icon(IconName::Undo)
.icon_size(IconSize::Small)
.icon_position(IconPosition::Start)
.layer(ElevationIndex::ModalSurface)
.when(env_var_set, |this| {
this.tooltip(Tooltip::text(format!("To reset your API key, unset the {env_var_name} environment variable.")))
})
.on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx))),
)
.into_any()
};
if self.load_credentials_task.is_some() {
div().child(Label::new("Loading credentials…")).into_any()
} else {
v_flex().size_full().child(api_key_section).into_any()
}
}
}