1use anyhow::Result;
2use convert_case::{Case, Casing};
3use futures::{FutureExt, StreamExt, future::BoxFuture};
4use gpui::{AnyView, App, AsyncApp, Context, Entity, SharedString, Task, Window};
5use http_client::HttpClient;
6use language_model::{
7 ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel, LanguageModelCompletionError,
8 LanguageModelCompletionEvent, LanguageModelId, LanguageModelName, LanguageModelProvider,
9 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
10 LanguageModelRequest, LanguageModelToolChoice, LanguageModelToolSchemaFormat, RateLimiter,
11};
12use menu;
13use open_ai::{
14 ResponseStreamEvent,
15 responses::{Request as ResponseRequest, StreamEvent as ResponsesStreamEvent, stream_response},
16 stream_completion,
17};
18use settings::{Settings, SettingsStore};
19use std::sync::Arc;
20use ui::{ElevationIndex, Tooltip, prelude::*};
21use ui_input::InputField;
22use util::ResultExt;
23
24use crate::provider::open_ai::{
25 OpenAiEventMapper, OpenAiResponseEventMapper, into_open_ai, into_open_ai_response,
26};
27pub use settings::OpenAiCompatibleAvailableModel as AvailableModel;
28pub use settings::OpenAiCompatibleModelCapabilities as ModelCapabilities;
29
30#[derive(Default, Clone, Debug, PartialEq)]
31pub struct OpenAiCompatibleSettings {
32 pub api_url: String,
33 pub available_models: Vec<AvailableModel>,
34}
35
36pub struct OpenAiCompatibleLanguageModelProvider {
37 id: LanguageModelProviderId,
38 name: LanguageModelProviderName,
39 http_client: Arc<dyn HttpClient>,
40 state: Entity<State>,
41}
42
43pub struct State {
44 id: Arc<str>,
45 api_key_state: ApiKeyState,
46 settings: OpenAiCompatibleSettings,
47}
48
49impl State {
50 fn is_authenticated(&self) -> bool {
51 self.api_key_state.has_key()
52 }
53
54 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
55 let api_url = SharedString::new(self.settings.api_url.as_str());
56 self.api_key_state
57 .store(api_url, api_key, |this| &mut this.api_key_state, cx)
58 }
59
60 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
61 let api_url = SharedString::new(self.settings.api_url.clone());
62 self.api_key_state
63 .load_if_needed(api_url, |this| &mut this.api_key_state, cx)
64 }
65}
66
67impl OpenAiCompatibleLanguageModelProvider {
68 pub fn new(id: Arc<str>, http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
69 fn resolve_settings<'a>(id: &'a str, cx: &'a App) -> Option<&'a OpenAiCompatibleSettings> {
70 crate::AllLanguageModelSettings::get_global(cx)
71 .openai_compatible
72 .get(id)
73 }
74
75 let api_key_env_var_name = format!("{}_API_KEY", id).to_case(Case::UpperSnake).into();
76 let state = cx.new(|cx| {
77 cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
78 let Some(settings) = resolve_settings(&this.id, cx).cloned() else {
79 return;
80 };
81 if &this.settings != &settings {
82 let api_url = SharedString::new(settings.api_url.as_str());
83 this.api_key_state.handle_url_change(
84 api_url,
85 |this| &mut this.api_key_state,
86 cx,
87 );
88 this.settings = settings;
89 cx.notify();
90 }
91 })
92 .detach();
93 let settings = resolve_settings(&id, cx).cloned().unwrap_or_default();
94 State {
95 id: id.clone(),
96 api_key_state: ApiKeyState::new(
97 SharedString::new(settings.api_url.as_str()),
98 EnvVar::new(api_key_env_var_name),
99 ),
100 settings,
101 }
102 });
103
104 Self {
105 id: id.clone().into(),
106 name: id.into(),
107 http_client,
108 state,
109 }
110 }
111
112 fn create_language_model(&self, model: AvailableModel) -> Arc<dyn LanguageModel> {
113 Arc::new(OpenAiCompatibleLanguageModel {
114 id: LanguageModelId::from(model.name.clone()),
115 provider_id: self.id.clone(),
116 provider_name: self.name.clone(),
117 model,
118 state: self.state.clone(),
119 http_client: self.http_client.clone(),
120 request_limiter: RateLimiter::new(4),
121 })
122 }
123}
124
125impl LanguageModelProviderState for OpenAiCompatibleLanguageModelProvider {
126 type ObservableEntity = State;
127
128 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
129 Some(self.state.clone())
130 }
131}
132
133impl LanguageModelProvider for OpenAiCompatibleLanguageModelProvider {
134 fn id(&self) -> LanguageModelProviderId {
135 self.id.clone()
136 }
137
138 fn name(&self) -> LanguageModelProviderName {
139 self.name.clone()
140 }
141
142 fn icon(&self) -> IconOrSvg {
143 IconOrSvg::Icon(IconName::AiOpenAiCompat)
144 }
145
146 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
147 self.state
148 .read(cx)
149 .settings
150 .available_models
151 .first()
152 .map(|model| self.create_language_model(model.clone()))
153 }
154
155 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
156 None
157 }
158
159 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
160 self.state
161 .read(cx)
162 .settings
163 .available_models
164 .iter()
165 .map(|model| self.create_language_model(model.clone()))
166 .collect()
167 }
168
169 fn is_authenticated(&self, cx: &App) -> bool {
170 self.state.read(cx).is_authenticated()
171 }
172
173 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
174 self.state.update(cx, |state, cx| state.authenticate(cx))
175 }
176
177 fn configuration_view(
178 &self,
179 _target_agent: language_model::ConfigurationViewTargetAgent,
180 window: &mut Window,
181 cx: &mut App,
182 ) -> AnyView {
183 cx.new(|cx| ConfigurationView::new(self.state.clone(), window, cx))
184 .into()
185 }
186
187 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
188 self.state
189 .update(cx, |state, cx| state.set_api_key(None, cx))
190 }
191}
192
193pub struct OpenAiCompatibleLanguageModel {
194 id: LanguageModelId,
195 provider_id: LanguageModelProviderId,
196 provider_name: LanguageModelProviderName,
197 model: AvailableModel,
198 state: Entity<State>,
199 http_client: Arc<dyn HttpClient>,
200 request_limiter: RateLimiter,
201}
202
203impl OpenAiCompatibleLanguageModel {
204 fn stream_completion(
205 &self,
206 request: open_ai::Request,
207 bypass_rate_limit: bool,
208 cx: &AsyncApp,
209 ) -> BoxFuture<
210 'static,
211 Result<
212 futures::stream::BoxStream<'static, Result<ResponseStreamEvent>>,
213 LanguageModelCompletionError,
214 >,
215 > {
216 let http_client = self.http_client.clone();
217
218 let (api_key, api_url) = self.state.read_with(cx, |state, _cx| {
219 let api_url = &state.settings.api_url;
220 (
221 state.api_key_state.key(api_url),
222 state.settings.api_url.clone(),
223 )
224 });
225
226 let provider = self.provider_name.clone();
227 let future = self.request_limiter.stream_with_bypass(
228 async move {
229 let Some(api_key) = api_key else {
230 return Err(LanguageModelCompletionError::NoApiKey { provider });
231 };
232 let request = stream_completion(
233 http_client.as_ref(),
234 provider.0.as_str(),
235 &api_url,
236 &api_key,
237 request,
238 );
239 let response = request.await?;
240 Ok(response)
241 },
242 bypass_rate_limit,
243 );
244
245 async move { Ok(future.await?.boxed()) }.boxed()
246 }
247
248 fn stream_response(
249 &self,
250 request: ResponseRequest,
251 bypass_rate_limit: bool,
252 cx: &AsyncApp,
253 ) -> BoxFuture<'static, Result<futures::stream::BoxStream<'static, Result<ResponsesStreamEvent>>>>
254 {
255 let http_client = self.http_client.clone();
256
257 let (api_key, api_url) = self.state.read_with(cx, |state, _cx| {
258 let api_url = &state.settings.api_url;
259 (
260 state.api_key_state.key(api_url),
261 state.settings.api_url.clone(),
262 )
263 });
264
265 let provider = self.provider_name.clone();
266 let future = self.request_limiter.stream_with_bypass(
267 async move {
268 let Some(api_key) = api_key else {
269 return Err(LanguageModelCompletionError::NoApiKey { provider });
270 };
271 let request = stream_response(
272 http_client.as_ref(),
273 provider.0.as_str(),
274 &api_url,
275 &api_key,
276 request,
277 );
278 let response = request.await?;
279 Ok(response)
280 },
281 bypass_rate_limit,
282 );
283
284 async move { Ok(future.await?.boxed()) }.boxed()
285 }
286}
287
288impl LanguageModel for OpenAiCompatibleLanguageModel {
289 fn id(&self) -> LanguageModelId {
290 self.id.clone()
291 }
292
293 fn name(&self) -> LanguageModelName {
294 LanguageModelName::from(
295 self.model
296 .display_name
297 .clone()
298 .unwrap_or_else(|| self.model.name.clone()),
299 )
300 }
301
302 fn provider_id(&self) -> LanguageModelProviderId {
303 self.provider_id.clone()
304 }
305
306 fn provider_name(&self) -> LanguageModelProviderName {
307 self.provider_name.clone()
308 }
309
310 fn supports_tools(&self) -> bool {
311 self.model.capabilities.tools
312 }
313
314 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
315 LanguageModelToolSchemaFormat::JsonSchemaSubset
316 }
317
318 fn supports_images(&self) -> bool {
319 self.model.capabilities.images
320 }
321
322 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
323 match choice {
324 LanguageModelToolChoice::Auto => self.model.capabilities.tools,
325 LanguageModelToolChoice::Any => self.model.capabilities.tools,
326 LanguageModelToolChoice::None => true,
327 }
328 }
329
330 fn supports_split_token_display(&self) -> bool {
331 true
332 }
333
334 fn telemetry_id(&self) -> String {
335 format!("openai/{}", self.model.name)
336 }
337
338 fn max_token_count(&self) -> u64 {
339 self.model.max_tokens
340 }
341
342 fn max_output_tokens(&self) -> Option<u64> {
343 self.model.max_output_tokens
344 }
345
346 fn count_tokens(
347 &self,
348 request: LanguageModelRequest,
349 cx: &App,
350 ) -> BoxFuture<'static, Result<u64>> {
351 let max_token_count = self.max_token_count();
352 cx.background_spawn(async move {
353 let messages = super::open_ai::collect_tiktoken_messages(request);
354 let model = if max_token_count >= 100_000 {
355 // If the max tokens is 100k or more, it is likely the o200k_base tokenizer from gpt4o
356 "gpt-4o"
357 } else {
358 // Otherwise fallback to gpt-4, since only cl100k_base and o200k_base are
359 // supported with this tiktoken method
360 "gpt-4"
361 };
362 tiktoken_rs::num_tokens_from_messages(model, &messages).map(|tokens| tokens as u64)
363 })
364 .boxed()
365 }
366
367 fn stream_completion(
368 &self,
369 request: LanguageModelRequest,
370 cx: &AsyncApp,
371 ) -> BoxFuture<
372 'static,
373 Result<
374 futures::stream::BoxStream<
375 'static,
376 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
377 >,
378 LanguageModelCompletionError,
379 >,
380 > {
381 let bypass_rate_limit = request.bypass_rate_limit;
382 if self.model.capabilities.chat_completions {
383 let request = into_open_ai(
384 request,
385 &self.model.name,
386 self.model.capabilities.parallel_tool_calls,
387 self.model.capabilities.prompt_cache_key,
388 self.max_output_tokens(),
389 None,
390 );
391 let completions = self.stream_completion(request, bypass_rate_limit, cx);
392 async move {
393 let mapper = OpenAiEventMapper::new();
394 Ok(mapper.map_stream(completions.await?).boxed())
395 }
396 .boxed()
397 } else {
398 let request = into_open_ai_response(
399 request,
400 &self.model.name,
401 self.model.capabilities.parallel_tool_calls,
402 self.model.capabilities.prompt_cache_key,
403 self.max_output_tokens(),
404 None,
405 );
406 let completions = self.stream_response(request, bypass_rate_limit, cx);
407 async move {
408 let mapper = OpenAiResponseEventMapper::new();
409 Ok(mapper.map_stream(completions.await?).boxed())
410 }
411 .boxed()
412 }
413 }
414}
415
416struct ConfigurationView {
417 api_key_editor: Entity<InputField>,
418 state: Entity<State>,
419 load_credentials_task: Option<Task<()>>,
420}
421
422impl ConfigurationView {
423 fn new(state: Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
424 let api_key_editor = cx.new(|cx| {
425 InputField::new(
426 window,
427 cx,
428 "000000000000000000000000000000000000000000000000000",
429 )
430 });
431
432 cx.observe(&state, |_, _, cx| {
433 cx.notify();
434 })
435 .detach();
436
437 let load_credentials_task = Some(cx.spawn_in(window, {
438 let state = state.clone();
439 async move |this, cx| {
440 if let Some(task) = Some(state.update(cx, |state, cx| state.authenticate(cx))) {
441 // We don't log an error, because "not signed in" is also an error.
442 let _ = task.await;
443 }
444 this.update(cx, |this, cx| {
445 this.load_credentials_task = None;
446 cx.notify();
447 })
448 .log_err();
449 }
450 }));
451
452 Self {
453 api_key_editor,
454 state,
455 load_credentials_task,
456 }
457 }
458
459 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
460 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
461 if api_key.is_empty() {
462 return;
463 }
464
465 // url changes can cause the editor to be displayed again
466 self.api_key_editor
467 .update(cx, |input, cx| input.set_text("", window, cx));
468
469 let state = self.state.clone();
470 cx.spawn_in(window, async move |_, cx| {
471 state
472 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
473 .await
474 })
475 .detach_and_log_err(cx);
476 }
477
478 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
479 self.api_key_editor
480 .update(cx, |input, cx| input.set_text("", window, cx));
481
482 let state = self.state.clone();
483 cx.spawn_in(window, async move |_, cx| {
484 state
485 .update(cx, |state, cx| state.set_api_key(None, cx))
486 .await
487 })
488 .detach_and_log_err(cx);
489 }
490
491 fn should_render_editor(&self, cx: &Context<Self>) -> bool {
492 !self.state.read(cx).is_authenticated()
493 }
494}
495
496impl Render for ConfigurationView {
497 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
498 let state = self.state.read(cx);
499 let env_var_set = state.api_key_state.is_from_env_var();
500 let env_var_name = state.api_key_state.env_var_name();
501
502 let api_key_section = if self.should_render_editor(cx) {
503 v_flex()
504 .on_action(cx.listener(Self::save_api_key))
505 .child(Label::new("To use Zed's agent with an OpenAI-compatible provider, you need to add an API key."))
506 .child(
507 div()
508 .pt(DynamicSpacing::Base04.rems(cx))
509 .child(self.api_key_editor.clone())
510 )
511 .child(
512 Label::new(
513 format!("You can also set the {env_var_name} environment variable and restart Zed."),
514 )
515 .size(LabelSize::Small).color(Color::Muted),
516 )
517 .into_any()
518 } else {
519 h_flex()
520 .mt_1()
521 .p_1()
522 .justify_between()
523 .rounded_md()
524 .border_1()
525 .border_color(cx.theme().colors().border)
526 .bg(cx.theme().colors().background)
527 .child(
528 h_flex()
529 .flex_1()
530 .min_w_0()
531 .gap_1()
532 .child(Icon::new(IconName::Check).color(Color::Success))
533 .child(
534 div()
535 .w_full()
536 .overflow_x_hidden()
537 .text_ellipsis()
538 .child(Label::new(
539 if env_var_set {
540 format!("API key set in {env_var_name} environment variable")
541 } else {
542 format!("API key configured for {}", &state.settings.api_url)
543 }
544 ))
545 ),
546 )
547 .child(
548 h_flex()
549 .flex_shrink_0()
550 .child(
551 Button::new("reset-api-key", "Reset API Key")
552 .label_size(LabelSize::Small)
553 .icon(IconName::Undo)
554 .icon_size(IconSize::Small)
555 .icon_position(IconPosition::Start)
556 .layer(ElevationIndex::ModalSurface)
557 .when(env_var_set, |this| {
558 this.tooltip(Tooltip::text(format!("To reset your API key, unset the {env_var_name} environment variable.")))
559 })
560 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx))),
561 ),
562 )
563 .into_any()
564 };
565
566 if self.load_credentials_task.is_some() {
567 div().child(Label::new("Loading credentials…")).into_any()
568 } else {
569 v_flex().size_full().child(api_key_section).into_any()
570 }
571 }
572}