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