1use anthropic::AnthropicModelMode;
2use anyhow::{Context as _, Result, anyhow};
3use chrono::{DateTime, Utc};
4use client::{Client, ModelRequestUsage, UserStore, zed_urls};
5use feature_flags::{FeatureFlagAppExt as _, ZedCloudFeatureFlag};
6use futures::{
7 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
8};
9use google_ai::GoogleModelMode;
10use gpui::{
11 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
12};
13use http_client::http::{HeaderMap, HeaderValue};
14use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
15use language_model::{
16 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
17 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
18 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
19 LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest,
20 LanguageModelToolChoice, LanguageModelToolSchemaFormat, LlmApiToken,
21 ModelRequestLimitReachedError, PaymentRequiredError, RateLimiter, RefreshLlmTokenListener,
22};
23use proto::Plan;
24use release_channel::AppVersion;
25use schemars::JsonSchema;
26use serde::{Deserialize, Serialize, de::DeserializeOwned};
27use settings::SettingsStore;
28use smol::io::{AsyncReadExt, BufReader};
29use std::pin::Pin;
30use std::str::FromStr as _;
31use std::sync::Arc;
32use std::time::Duration;
33use thiserror::Error;
34use ui::{TintColor, prelude::*};
35use util::{ResultExt as _, maybe};
36use zed_llm_client::{
37 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
38 CompletionRequestStatus, CountTokensBody, CountTokensResponse, EXPIRED_LLM_TOKEN_HEADER_NAME,
39 ListModelsResponse, MODEL_REQUESTS_RESOURCE_HEADER_VALUE,
40 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
41 TOOL_USE_LIMIT_REACHED_HEADER_NAME, ZED_VERSION_HEADER_NAME,
42};
43
44use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
45use crate::provider::google::{GoogleEventMapper, into_google};
46use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
47
48const PROVIDER_ID: LanguageModelProviderId = language_model::ZED_CLOUD_PROVIDER_ID;
49const PROVIDER_NAME: LanguageModelProviderName = language_model::ZED_CLOUD_PROVIDER_NAME;
50
51#[derive(Default, Clone, Debug, PartialEq)]
52pub struct ZedDotDevSettings {
53 pub available_models: Vec<AvailableModel>,
54}
55
56#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
57#[serde(rename_all = "lowercase")]
58pub enum AvailableProvider {
59 Anthropic,
60 OpenAi,
61 Google,
62}
63
64#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
65pub struct AvailableModel {
66 /// The provider of the language model.
67 pub provider: AvailableProvider,
68 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
69 pub name: String,
70 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
71 pub display_name: Option<String>,
72 /// The size of the context window, indicating the maximum number of tokens the model can process.
73 pub max_tokens: usize,
74 /// The maximum number of output tokens allowed by the model.
75 pub max_output_tokens: Option<u64>,
76 /// The maximum number of completion tokens allowed by the model (o1-* only)
77 pub max_completion_tokens: Option<u64>,
78 /// Override this model with a different Anthropic model for tool calls.
79 pub tool_override: Option<String>,
80 /// Indicates whether this custom model supports caching.
81 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
82 /// The default temperature to use for this model.
83 pub default_temperature: Option<f32>,
84 /// Any extra beta headers to provide when using the model.
85 #[serde(default)]
86 pub extra_beta_headers: Vec<String>,
87 /// The model's mode (e.g. thinking)
88 pub mode: Option<ModelMode>,
89}
90
91#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
92#[serde(tag = "type", rename_all = "lowercase")]
93pub enum ModelMode {
94 #[default]
95 Default,
96 Thinking {
97 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
98 budget_tokens: Option<u32>,
99 },
100}
101
102impl From<ModelMode> for AnthropicModelMode {
103 fn from(value: ModelMode) -> Self {
104 match value {
105 ModelMode::Default => AnthropicModelMode::Default,
106 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
107 }
108 }
109}
110
111pub struct CloudLanguageModelProvider {
112 client: Arc<Client>,
113 state: gpui::Entity<State>,
114 _maintain_client_status: Task<()>,
115}
116
117pub struct State {
118 client: Arc<Client>,
119 llm_api_token: LlmApiToken,
120 user_store: Entity<UserStore>,
121 status: client::Status,
122 accept_terms_of_service_task: Option<Task<Result<()>>>,
123 models: Vec<Arc<zed_llm_client::LanguageModel>>,
124 default_model: Option<Arc<zed_llm_client::LanguageModel>>,
125 default_fast_model: Option<Arc<zed_llm_client::LanguageModel>>,
126 recommended_models: Vec<Arc<zed_llm_client::LanguageModel>>,
127 _fetch_models_task: Task<()>,
128 _settings_subscription: Subscription,
129 _llm_token_subscription: Subscription,
130}
131
132impl State {
133 fn new(
134 client: Arc<Client>,
135 user_store: Entity<UserStore>,
136 status: client::Status,
137 cx: &mut Context<Self>,
138 ) -> Self {
139 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
140 let use_cloud = cx.has_flag::<ZedCloudFeatureFlag>();
141
142 Self {
143 client: client.clone(),
144 llm_api_token: LlmApiToken::default(),
145 user_store,
146 status,
147 accept_terms_of_service_task: None,
148 models: Vec::new(),
149 default_model: None,
150 default_fast_model: None,
151 recommended_models: Vec::new(),
152 _fetch_models_task: cx.spawn(async move |this, cx| {
153 maybe!(async move {
154 let (client, llm_api_token) = this
155 .read_with(cx, |this, _cx| (client.clone(), this.llm_api_token.clone()))?;
156
157 loop {
158 let status = this.read_with(cx, |this, _cx| this.status)?;
159 if matches!(status, client::Status::Connected { .. }) {
160 break;
161 }
162
163 cx.background_executor()
164 .timer(Duration::from_millis(100))
165 .await;
166 }
167
168 let response = Self::fetch_models(client, llm_api_token, use_cloud).await?;
169 this.update(cx, |this, cx| {
170 this.update_models(response, cx);
171 })
172 })
173 .await
174 .context("failed to fetch Zed models")
175 .log_err();
176 }),
177 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
178 cx.notify();
179 }),
180 _llm_token_subscription: cx.subscribe(
181 &refresh_llm_token_listener,
182 move |this, _listener, _event, cx| {
183 let client = this.client.clone();
184 let llm_api_token = this.llm_api_token.clone();
185 cx.spawn(async move |this, cx| {
186 llm_api_token.refresh(&client).await?;
187 let response = Self::fetch_models(client, llm_api_token, use_cloud).await?;
188 this.update(cx, |this, cx| {
189 this.update_models(response, cx);
190 })
191 })
192 .detach_and_log_err(cx);
193 },
194 ),
195 }
196 }
197
198 fn is_signed_out(&self) -> bool {
199 self.status.is_signed_out()
200 }
201
202 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
203 let client = self.client.clone();
204 cx.spawn(async move |state, cx| {
205 client
206 .authenticate_and_connect(true, &cx)
207 .await
208 .into_response()?;
209 state.update(cx, |_, cx| cx.notify())
210 })
211 }
212
213 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
214 self.user_store
215 .read(cx)
216 .current_user_has_accepted_terms()
217 .unwrap_or(false)
218 }
219
220 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
221 let user_store = self.user_store.clone();
222 self.accept_terms_of_service_task = Some(cx.spawn(async move |this, cx| {
223 let _ = user_store
224 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
225 .await;
226 this.update(cx, |this, cx| {
227 this.accept_terms_of_service_task = None;
228 cx.notify()
229 })
230 }));
231 }
232
233 fn update_models(&mut self, response: ListModelsResponse, cx: &mut Context<Self>) {
234 let mut models = Vec::new();
235
236 for model in response.models {
237 models.push(Arc::new(model.clone()));
238
239 // Right now we represent thinking variants of models as separate models on the client,
240 // so we need to insert variants for any model that supports thinking.
241 if model.supports_thinking {
242 models.push(Arc::new(zed_llm_client::LanguageModel {
243 id: zed_llm_client::LanguageModelId(format!("{}-thinking", model.id).into()),
244 display_name: format!("{} Thinking", model.display_name),
245 ..model
246 }));
247 }
248 }
249
250 self.default_model = models
251 .iter()
252 .find(|model| model.id == response.default_model)
253 .cloned();
254 self.default_fast_model = models
255 .iter()
256 .find(|model| model.id == response.default_fast_model)
257 .cloned();
258 self.recommended_models = response
259 .recommended_models
260 .iter()
261 .filter_map(|id| models.iter().find(|model| &model.id == id))
262 .cloned()
263 .collect();
264 self.models = models;
265 cx.notify();
266 }
267
268 async fn fetch_models(
269 client: Arc<Client>,
270 llm_api_token: LlmApiToken,
271 use_cloud: bool,
272 ) -> Result<ListModelsResponse> {
273 let http_client = &client.http_client();
274 let token = llm_api_token.acquire(&client).await?;
275
276 let request = http_client::Request::builder()
277 .method(Method::GET)
278 .uri(
279 http_client
280 .build_zed_llm_url("/models", &[], use_cloud)?
281 .as_ref(),
282 )
283 .header("Authorization", format!("Bearer {token}"))
284 .body(AsyncBody::empty())?;
285 let mut response = http_client
286 .send(request)
287 .await
288 .context("failed to send list models request")?;
289
290 if response.status().is_success() {
291 let mut body = String::new();
292 response.body_mut().read_to_string(&mut body).await?;
293 return Ok(serde_json::from_str(&body)?);
294 } else {
295 let mut body = String::new();
296 response.body_mut().read_to_string(&mut body).await?;
297 anyhow::bail!(
298 "error listing models.\nStatus: {:?}\nBody: {body}",
299 response.status(),
300 );
301 }
302 }
303}
304
305impl CloudLanguageModelProvider {
306 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
307 let mut status_rx = client.status();
308 let status = *status_rx.borrow();
309
310 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
311
312 let state_ref = state.downgrade();
313 let maintain_client_status = cx.spawn(async move |cx| {
314 while let Some(status) = status_rx.next().await {
315 if let Some(this) = state_ref.upgrade() {
316 _ = this.update(cx, |this, cx| {
317 if this.status != status {
318 this.status = status;
319 cx.notify();
320 }
321 });
322 } else {
323 break;
324 }
325 }
326 });
327
328 Self {
329 client,
330 state: state.clone(),
331 _maintain_client_status: maintain_client_status,
332 }
333 }
334
335 fn create_language_model(
336 &self,
337 model: Arc<zed_llm_client::LanguageModel>,
338 llm_api_token: LlmApiToken,
339 ) -> Arc<dyn LanguageModel> {
340 Arc::new(CloudLanguageModel {
341 id: LanguageModelId(SharedString::from(model.id.0.clone())),
342 model,
343 llm_api_token: llm_api_token.clone(),
344 client: self.client.clone(),
345 request_limiter: RateLimiter::new(4),
346 })
347 }
348}
349
350impl LanguageModelProviderState for CloudLanguageModelProvider {
351 type ObservableEntity = State;
352
353 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
354 Some(self.state.clone())
355 }
356}
357
358impl LanguageModelProvider for CloudLanguageModelProvider {
359 fn id(&self) -> LanguageModelProviderId {
360 PROVIDER_ID
361 }
362
363 fn name(&self) -> LanguageModelProviderName {
364 PROVIDER_NAME
365 }
366
367 fn icon(&self) -> IconName {
368 IconName::AiZed
369 }
370
371 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
372 let default_model = self.state.read(cx).default_model.clone()?;
373 let llm_api_token = self.state.read(cx).llm_api_token.clone();
374 Some(self.create_language_model(default_model, llm_api_token))
375 }
376
377 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
378 let default_fast_model = self.state.read(cx).default_fast_model.clone()?;
379 let llm_api_token = self.state.read(cx).llm_api_token.clone();
380 Some(self.create_language_model(default_fast_model, llm_api_token))
381 }
382
383 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
384 let llm_api_token = self.state.read(cx).llm_api_token.clone();
385 self.state
386 .read(cx)
387 .recommended_models
388 .iter()
389 .cloned()
390 .map(|model| self.create_language_model(model, llm_api_token.clone()))
391 .collect()
392 }
393
394 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
395 let llm_api_token = self.state.read(cx).llm_api_token.clone();
396 self.state
397 .read(cx)
398 .models
399 .iter()
400 .cloned()
401 .map(|model| self.create_language_model(model, llm_api_token.clone()))
402 .collect()
403 }
404
405 fn is_authenticated(&self, cx: &App) -> bool {
406 let state = self.state.read(cx);
407 !state.is_signed_out() && state.has_accepted_terms_of_service(cx)
408 }
409
410 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
411 Task::ready(Ok(()))
412 }
413
414 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
415 cx.new(|_| ConfigurationView::new(self.state.clone()))
416 .into()
417 }
418
419 fn must_accept_terms(&self, cx: &App) -> bool {
420 !self.state.read(cx).has_accepted_terms_of_service(cx)
421 }
422
423 fn render_accept_terms(
424 &self,
425 view: LanguageModelProviderTosView,
426 cx: &mut App,
427 ) -> Option<AnyElement> {
428 let state = self.state.read(cx);
429 if state.has_accepted_terms_of_service(cx) {
430 return None;
431 }
432 Some(
433 render_accept_terms(view, state.accept_terms_of_service_task.is_some(), {
434 let state = self.state.clone();
435 move |_window, cx| {
436 state.update(cx, |state, cx| state.accept_terms_of_service(cx));
437 }
438 })
439 .into_any_element(),
440 )
441 }
442
443 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
444 Task::ready(Ok(()))
445 }
446}
447
448fn render_accept_terms(
449 view_kind: LanguageModelProviderTosView,
450 accept_terms_of_service_in_progress: bool,
451 accept_terms_callback: impl Fn(&mut Window, &mut App) + 'static,
452) -> impl IntoElement {
453 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
454 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadEmptyState);
455
456 let terms_button = Button::new("terms_of_service", "Terms of Service")
457 .style(ButtonStyle::Subtle)
458 .icon(IconName::ArrowUpRight)
459 .icon_color(Color::Muted)
460 .icon_size(IconSize::XSmall)
461 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
462 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
463
464 let button_container = h_flex().child(
465 Button::new("accept_terms", "I accept the Terms of Service")
466 .when(!thread_empty_state, |this| {
467 this.full_width()
468 .style(ButtonStyle::Tinted(TintColor::Accent))
469 .icon(IconName::Check)
470 .icon_position(IconPosition::Start)
471 .icon_size(IconSize::Small)
472 })
473 .when(thread_empty_state, |this| {
474 this.style(ButtonStyle::Tinted(TintColor::Warning))
475 .label_size(LabelSize::Small)
476 })
477 .disabled(accept_terms_of_service_in_progress)
478 .on_click(move |_, window, cx| (accept_terms_callback)(window, cx)),
479 );
480
481 if thread_empty_state {
482 h_flex()
483 .w_full()
484 .flex_wrap()
485 .justify_between()
486 .child(
487 h_flex()
488 .child(
489 Label::new("To start using Zed AI, please read and accept the")
490 .size(LabelSize::Small),
491 )
492 .child(terms_button),
493 )
494 .child(button_container)
495 } else {
496 v_flex()
497 .w_full()
498 .gap_2()
499 .child(
500 h_flex()
501 .flex_wrap()
502 .when(thread_fresh_start, |this| this.justify_center())
503 .child(Label::new(
504 "To start using Zed AI, please read and accept the",
505 ))
506 .child(terms_button),
507 )
508 .child({
509 match view_kind {
510 LanguageModelProviderTosView::PromptEditorPopup => {
511 button_container.w_full().justify_end()
512 }
513 LanguageModelProviderTosView::Configuration => {
514 button_container.w_full().justify_start()
515 }
516 LanguageModelProviderTosView::ThreadFreshStart => {
517 button_container.w_full().justify_center()
518 }
519 LanguageModelProviderTosView::ThreadEmptyState => div().w_0(),
520 }
521 })
522 }
523}
524
525pub struct CloudLanguageModel {
526 id: LanguageModelId,
527 model: Arc<zed_llm_client::LanguageModel>,
528 llm_api_token: LlmApiToken,
529 client: Arc<Client>,
530 request_limiter: RateLimiter,
531}
532
533struct PerformLlmCompletionResponse {
534 response: Response<AsyncBody>,
535 usage: Option<ModelRequestUsage>,
536 tool_use_limit_reached: bool,
537 includes_status_messages: bool,
538}
539
540impl CloudLanguageModel {
541 async fn perform_llm_completion(
542 client: Arc<Client>,
543 llm_api_token: LlmApiToken,
544 app_version: Option<SemanticVersion>,
545 body: CompletionBody,
546 use_cloud: bool,
547 ) -> Result<PerformLlmCompletionResponse> {
548 let http_client = &client.http_client();
549
550 let mut token = llm_api_token.acquire(&client).await?;
551 let mut refreshed_token = false;
552
553 loop {
554 let request_builder = http_client::Request::builder().method(Method::POST).uri(
555 http_client
556 .build_zed_llm_url("/completions", &[], use_cloud)?
557 .as_ref(),
558 );
559 let request_builder = if let Some(app_version) = app_version {
560 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
561 } else {
562 request_builder
563 };
564
565 let request = request_builder
566 .header("Content-Type", "application/json")
567 .header("Authorization", format!("Bearer {token}"))
568 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
569 .body(serde_json::to_string(&body)?.into())?;
570 let mut response = http_client.send(request).await?;
571 let status = response.status();
572 if status.is_success() {
573 let includes_status_messages = response
574 .headers()
575 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
576 .is_some();
577
578 let tool_use_limit_reached = response
579 .headers()
580 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
581 .is_some();
582
583 let usage = if includes_status_messages {
584 None
585 } else {
586 ModelRequestUsage::from_headers(response.headers()).ok()
587 };
588
589 return Ok(PerformLlmCompletionResponse {
590 response,
591 usage,
592 includes_status_messages,
593 tool_use_limit_reached,
594 });
595 }
596
597 if !refreshed_token
598 && response
599 .headers()
600 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
601 .is_some()
602 {
603 token = llm_api_token.refresh(&client).await?;
604 refreshed_token = true;
605 continue;
606 }
607
608 if status == StatusCode::FORBIDDEN
609 && response
610 .headers()
611 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
612 .is_some()
613 {
614 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
615 .headers()
616 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
617 .and_then(|resource| resource.to_str().ok())
618 {
619 if let Some(plan) = response
620 .headers()
621 .get(CURRENT_PLAN_HEADER_NAME)
622 .and_then(|plan| plan.to_str().ok())
623 .and_then(|plan| zed_llm_client::Plan::from_str(plan).ok())
624 {
625 let plan = match plan {
626 zed_llm_client::Plan::ZedFree => Plan::Free,
627 zed_llm_client::Plan::ZedPro => Plan::ZedPro,
628 zed_llm_client::Plan::ZedProTrial => Plan::ZedProTrial,
629 };
630 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
631 }
632 }
633 } else if status == StatusCode::PAYMENT_REQUIRED {
634 return Err(anyhow!(PaymentRequiredError));
635 }
636
637 let mut body = String::new();
638 let headers = response.headers().clone();
639 response.body_mut().read_to_string(&mut body).await?;
640 return Err(anyhow!(ApiError {
641 status,
642 body,
643 headers
644 }));
645 }
646 }
647}
648
649#[derive(Debug, Error)]
650#[error("cloud language model request failed with status {status}: {body}")]
651struct ApiError {
652 status: StatusCode,
653 body: String,
654 headers: HeaderMap<HeaderValue>,
655}
656
657impl From<ApiError> for LanguageModelCompletionError {
658 fn from(error: ApiError) -> Self {
659 let retry_after = None;
660 LanguageModelCompletionError::from_http_status(
661 PROVIDER_NAME,
662 error.status,
663 error.body,
664 retry_after,
665 )
666 }
667}
668
669impl LanguageModel for CloudLanguageModel {
670 fn id(&self) -> LanguageModelId {
671 self.id.clone()
672 }
673
674 fn name(&self) -> LanguageModelName {
675 LanguageModelName::from(self.model.display_name.clone())
676 }
677
678 fn provider_id(&self) -> LanguageModelProviderId {
679 PROVIDER_ID
680 }
681
682 fn provider_name(&self) -> LanguageModelProviderName {
683 PROVIDER_NAME
684 }
685
686 fn upstream_provider_id(&self) -> LanguageModelProviderId {
687 use zed_llm_client::LanguageModelProvider::*;
688 match self.model.provider {
689 Anthropic => language_model::ANTHROPIC_PROVIDER_ID,
690 OpenAi => language_model::OPEN_AI_PROVIDER_ID,
691 Google => language_model::GOOGLE_PROVIDER_ID,
692 }
693 }
694
695 fn upstream_provider_name(&self) -> LanguageModelProviderName {
696 use zed_llm_client::LanguageModelProvider::*;
697 match self.model.provider {
698 Anthropic => language_model::ANTHROPIC_PROVIDER_NAME,
699 OpenAi => language_model::OPEN_AI_PROVIDER_NAME,
700 Google => language_model::GOOGLE_PROVIDER_NAME,
701 }
702 }
703
704 fn supports_tools(&self) -> bool {
705 self.model.supports_tools
706 }
707
708 fn supports_images(&self) -> bool {
709 self.model.supports_images
710 }
711
712 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
713 match choice {
714 LanguageModelToolChoice::Auto
715 | LanguageModelToolChoice::Any
716 | LanguageModelToolChoice::None => true,
717 }
718 }
719
720 fn supports_burn_mode(&self) -> bool {
721 self.model.supports_max_mode
722 }
723
724 fn telemetry_id(&self) -> String {
725 format!("zed.dev/{}", self.model.id)
726 }
727
728 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
729 match self.model.provider {
730 zed_llm_client::LanguageModelProvider::Anthropic
731 | zed_llm_client::LanguageModelProvider::OpenAi => {
732 LanguageModelToolSchemaFormat::JsonSchema
733 }
734 zed_llm_client::LanguageModelProvider::Google => {
735 LanguageModelToolSchemaFormat::JsonSchemaSubset
736 }
737 }
738 }
739
740 fn max_token_count(&self) -> u64 {
741 self.model.max_token_count as u64
742 }
743
744 fn max_token_count_in_burn_mode(&self) -> Option<u64> {
745 self.model
746 .max_token_count_in_max_mode
747 .filter(|_| self.model.supports_max_mode)
748 .map(|max_token_count| max_token_count as u64)
749 }
750
751 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
752 match &self.model.provider {
753 zed_llm_client::LanguageModelProvider::Anthropic => {
754 Some(LanguageModelCacheConfiguration {
755 min_total_token: 2_048,
756 should_speculate: true,
757 max_cache_anchors: 4,
758 })
759 }
760 zed_llm_client::LanguageModelProvider::OpenAi
761 | zed_llm_client::LanguageModelProvider::Google => None,
762 }
763 }
764
765 fn count_tokens(
766 &self,
767 request: LanguageModelRequest,
768 cx: &App,
769 ) -> BoxFuture<'static, Result<u64>> {
770 match self.model.provider {
771 zed_llm_client::LanguageModelProvider::Anthropic => count_anthropic_tokens(request, cx),
772 zed_llm_client::LanguageModelProvider::OpenAi => {
773 let model = match open_ai::Model::from_id(&self.model.id.0) {
774 Ok(model) => model,
775 Err(err) => return async move { Err(anyhow!(err)) }.boxed(),
776 };
777 count_open_ai_tokens(request, model, cx)
778 }
779 zed_llm_client::LanguageModelProvider::Google => {
780 let client = self.client.clone();
781 let llm_api_token = self.llm_api_token.clone();
782 let model_id = self.model.id.to_string();
783 let generate_content_request =
784 into_google(request, model_id.clone(), GoogleModelMode::Default);
785 let use_cloud = cx.has_flag::<ZedCloudFeatureFlag>();
786 async move {
787 let http_client = &client.http_client();
788 let token = llm_api_token.acquire(&client).await?;
789
790 let request_body = CountTokensBody {
791 provider: zed_llm_client::LanguageModelProvider::Google,
792 model: model_id,
793 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
794 generate_content_request,
795 })?,
796 };
797 let request = http_client::Request::builder()
798 .method(Method::POST)
799 .uri(
800 http_client
801 .build_zed_llm_url("/count_tokens", &[], use_cloud)?
802 .as_ref(),
803 )
804 .header("Content-Type", "application/json")
805 .header("Authorization", format!("Bearer {token}"))
806 .body(serde_json::to_string(&request_body)?.into())?;
807 let mut response = http_client.send(request).await?;
808 let status = response.status();
809 let headers = response.headers().clone();
810 let mut response_body = String::new();
811 response
812 .body_mut()
813 .read_to_string(&mut response_body)
814 .await?;
815
816 if status.is_success() {
817 let response_body: CountTokensResponse =
818 serde_json::from_str(&response_body)?;
819
820 Ok(response_body.tokens as u64)
821 } else {
822 Err(anyhow!(ApiError {
823 status,
824 body: response_body,
825 headers
826 }))
827 }
828 }
829 .boxed()
830 }
831 }
832 }
833
834 fn stream_completion(
835 &self,
836 request: LanguageModelRequest,
837 cx: &AsyncApp,
838 ) -> BoxFuture<
839 'static,
840 Result<
841 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
842 LanguageModelCompletionError,
843 >,
844 > {
845 let thread_id = request.thread_id.clone();
846 let prompt_id = request.prompt_id.clone();
847 let intent = request.intent;
848 let mode = request.mode;
849 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
850 let use_cloud = cx
851 .update(|cx| cx.has_flag::<ZedCloudFeatureFlag>())
852 .unwrap_or(false);
853 let thinking_allowed = request.thinking_allowed;
854 match self.model.provider {
855 zed_llm_client::LanguageModelProvider::Anthropic => {
856 let request = into_anthropic(
857 request,
858 self.model.id.to_string(),
859 1.0,
860 self.model.max_output_tokens as u64,
861 if thinking_allowed && self.model.id.0.ends_with("-thinking") {
862 AnthropicModelMode::Thinking {
863 budget_tokens: Some(4_096),
864 }
865 } else {
866 AnthropicModelMode::Default
867 },
868 );
869 let client = self.client.clone();
870 let llm_api_token = self.llm_api_token.clone();
871 let future = self.request_limiter.stream(async move {
872 let PerformLlmCompletionResponse {
873 response,
874 usage,
875 includes_status_messages,
876 tool_use_limit_reached,
877 } = Self::perform_llm_completion(
878 client.clone(),
879 llm_api_token,
880 app_version,
881 CompletionBody {
882 thread_id,
883 prompt_id,
884 intent,
885 mode,
886 provider: zed_llm_client::LanguageModelProvider::Anthropic,
887 model: request.model.clone(),
888 provider_request: serde_json::to_value(&request)
889 .map_err(|e| anyhow!(e))?,
890 },
891 use_cloud,
892 )
893 .await
894 .map_err(|err| match err.downcast::<ApiError>() {
895 Ok(api_err) => anyhow!(LanguageModelCompletionError::from(api_err)),
896 Err(err) => anyhow!(err),
897 })?;
898
899 let mut mapper = AnthropicEventMapper::new();
900 Ok(map_cloud_completion_events(
901 Box::pin(
902 response_lines(response, includes_status_messages)
903 .chain(usage_updated_event(usage))
904 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
905 ),
906 move |event| mapper.map_event(event),
907 ))
908 });
909 async move { Ok(future.await?.boxed()) }.boxed()
910 }
911 zed_llm_client::LanguageModelProvider::OpenAi => {
912 let client = self.client.clone();
913 let model = match open_ai::Model::from_id(&self.model.id.0) {
914 Ok(model) => model,
915 Err(err) => return async move { Err(anyhow!(err).into()) }.boxed(),
916 };
917 let request = into_open_ai(
918 request,
919 model.id(),
920 model.supports_parallel_tool_calls(),
921 None,
922 );
923 let llm_api_token = self.llm_api_token.clone();
924 let future = self.request_limiter.stream(async move {
925 let PerformLlmCompletionResponse {
926 response,
927 usage,
928 includes_status_messages,
929 tool_use_limit_reached,
930 } = Self::perform_llm_completion(
931 client.clone(),
932 llm_api_token,
933 app_version,
934 CompletionBody {
935 thread_id,
936 prompt_id,
937 intent,
938 mode,
939 provider: zed_llm_client::LanguageModelProvider::OpenAi,
940 model: request.model.clone(),
941 provider_request: serde_json::to_value(&request)
942 .map_err(|e| anyhow!(e))?,
943 },
944 use_cloud,
945 )
946 .await?;
947
948 let mut mapper = OpenAiEventMapper::new();
949 Ok(map_cloud_completion_events(
950 Box::pin(
951 response_lines(response, includes_status_messages)
952 .chain(usage_updated_event(usage))
953 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
954 ),
955 move |event| mapper.map_event(event),
956 ))
957 });
958 async move { Ok(future.await?.boxed()) }.boxed()
959 }
960 zed_llm_client::LanguageModelProvider::Google => {
961 let client = self.client.clone();
962 let request =
963 into_google(request, self.model.id.to_string(), GoogleModelMode::Default);
964 let llm_api_token = self.llm_api_token.clone();
965 let future = self.request_limiter.stream(async move {
966 let PerformLlmCompletionResponse {
967 response,
968 usage,
969 includes_status_messages,
970 tool_use_limit_reached,
971 } = Self::perform_llm_completion(
972 client.clone(),
973 llm_api_token,
974 app_version,
975 CompletionBody {
976 thread_id,
977 prompt_id,
978 intent,
979 mode,
980 provider: zed_llm_client::LanguageModelProvider::Google,
981 model: request.model.model_id.clone(),
982 provider_request: serde_json::to_value(&request)
983 .map_err(|e| anyhow!(e))?,
984 },
985 use_cloud,
986 )
987 .await?;
988
989 let mut mapper = GoogleEventMapper::new();
990 Ok(map_cloud_completion_events(
991 Box::pin(
992 response_lines(response, includes_status_messages)
993 .chain(usage_updated_event(usage))
994 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
995 ),
996 move |event| mapper.map_event(event),
997 ))
998 });
999 async move { Ok(future.await?.boxed()) }.boxed()
1000 }
1001 }
1002 }
1003}
1004
1005#[derive(Serialize, Deserialize)]
1006#[serde(rename_all = "snake_case")]
1007pub enum CloudCompletionEvent<T> {
1008 Status(CompletionRequestStatus),
1009 Event(T),
1010}
1011
1012fn map_cloud_completion_events<T, F>(
1013 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
1014 mut map_callback: F,
1015) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1016where
1017 T: DeserializeOwned + 'static,
1018 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
1019 + Send
1020 + 'static,
1021{
1022 stream
1023 .flat_map(move |event| {
1024 futures::stream::iter(match event {
1025 Err(error) => {
1026 vec![Err(LanguageModelCompletionError::from(error))]
1027 }
1028 Ok(CloudCompletionEvent::Status(event)) => {
1029 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
1030 }
1031 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
1032 })
1033 })
1034 .boxed()
1035}
1036
1037fn usage_updated_event<T>(
1038 usage: Option<ModelRequestUsage>,
1039) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1040 futures::stream::iter(usage.map(|usage| {
1041 Ok(CloudCompletionEvent::Status(
1042 CompletionRequestStatus::UsageUpdated {
1043 amount: usage.amount as usize,
1044 limit: usage.limit,
1045 },
1046 ))
1047 }))
1048}
1049
1050fn tool_use_limit_reached_event<T>(
1051 tool_use_limit_reached: bool,
1052) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1053 futures::stream::iter(tool_use_limit_reached.then(|| {
1054 Ok(CloudCompletionEvent::Status(
1055 CompletionRequestStatus::ToolUseLimitReached,
1056 ))
1057 }))
1058}
1059
1060fn response_lines<T: DeserializeOwned>(
1061 response: Response<AsyncBody>,
1062 includes_status_messages: bool,
1063) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1064 futures::stream::try_unfold(
1065 (String::new(), BufReader::new(response.into_body())),
1066 move |(mut line, mut body)| async move {
1067 match body.read_line(&mut line).await {
1068 Ok(0) => Ok(None),
1069 Ok(_) => {
1070 let event = if includes_status_messages {
1071 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
1072 } else {
1073 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1074 };
1075
1076 line.clear();
1077 Ok(Some((event, (line, body))))
1078 }
1079 Err(e) => Err(e.into()),
1080 }
1081 },
1082 )
1083}
1084
1085#[derive(IntoElement, RegisterComponent)]
1086struct ZedAiConfiguration {
1087 is_connected: bool,
1088 plan: Option<proto::Plan>,
1089 subscription_period: Option<(DateTime<Utc>, DateTime<Utc>)>,
1090 eligible_for_trial: bool,
1091 has_accepted_terms_of_service: bool,
1092 accept_terms_of_service_in_progress: bool,
1093 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1094 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1095}
1096
1097impl RenderOnce for ZedAiConfiguration {
1098 fn render(self, _window: &mut Window, _cx: &mut App) -> impl IntoElement {
1099 const ZED_PRICING_URL: &str = "https://zed.dev/pricing";
1100
1101 let is_pro = self.plan == Some(proto::Plan::ZedPro);
1102 let subscription_text = match (self.plan, self.subscription_period) {
1103 (Some(proto::Plan::ZedPro), Some(_)) => {
1104 "You have access to Zed's hosted LLMs through your Zed Pro subscription."
1105 }
1106 (Some(proto::Plan::ZedProTrial), Some(_)) => {
1107 "You have access to Zed's hosted LLMs through your Zed Pro trial."
1108 }
1109 (Some(proto::Plan::Free), Some(_)) => {
1110 "You have basic access to Zed's hosted LLMs through your Zed Free subscription."
1111 }
1112 _ => {
1113 if self.eligible_for_trial {
1114 "Subscribe for access to Zed's hosted LLMs. Start with a 14 day free trial."
1115 } else {
1116 "Subscribe for access to Zed's hosted LLMs."
1117 }
1118 }
1119 };
1120 let manage_subscription_buttons = if is_pro {
1121 h_flex().child(
1122 Button::new("manage_settings", "Manage Subscription")
1123 .style(ButtonStyle::Tinted(TintColor::Accent))
1124 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1125 )
1126 } else {
1127 h_flex()
1128 .gap_2()
1129 .child(
1130 Button::new("learn_more", "Learn more")
1131 .style(ButtonStyle::Subtle)
1132 .on_click(|_, _, cx| cx.open_url(ZED_PRICING_URL)),
1133 )
1134 .child(
1135 Button::new(
1136 "upgrade",
1137 if self.plan.is_none() && self.eligible_for_trial {
1138 "Start Trial"
1139 } else {
1140 "Upgrade"
1141 },
1142 )
1143 .style(ButtonStyle::Subtle)
1144 .color(Color::Accent)
1145 .on_click(|_, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1146 )
1147 };
1148
1149 if self.is_connected {
1150 v_flex()
1151 .gap_3()
1152 .w_full()
1153 .when(!self.has_accepted_terms_of_service, |this| {
1154 this.child(render_accept_terms(
1155 LanguageModelProviderTosView::Configuration,
1156 self.accept_terms_of_service_in_progress,
1157 {
1158 let callback = self.accept_terms_of_service_callback.clone();
1159 move |window, cx| (callback)(window, cx)
1160 },
1161 ))
1162 })
1163 .when(self.has_accepted_terms_of_service, |this| {
1164 this.child(subscription_text)
1165 .child(manage_subscription_buttons)
1166 })
1167 } else {
1168 v_flex()
1169 .gap_2()
1170 .child(Label::new("Use Zed AI to access hosted language models."))
1171 .child(
1172 Button::new("sign_in", "Sign In")
1173 .icon_color(Color::Muted)
1174 .icon(IconName::Github)
1175 .icon_position(IconPosition::Start)
1176 .on_click({
1177 let callback = self.sign_in_callback.clone();
1178 move |_, window, cx| (callback)(window, cx)
1179 }),
1180 )
1181 }
1182 }
1183}
1184
1185struct ConfigurationView {
1186 state: Entity<State>,
1187 accept_terms_of_service_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1188 sign_in_callback: Arc<dyn Fn(&mut Window, &mut App) + Send + Sync>,
1189}
1190
1191impl ConfigurationView {
1192 fn new(state: Entity<State>) -> Self {
1193 let accept_terms_of_service_callback = Arc::new({
1194 let state = state.clone();
1195 move |_window: &mut Window, cx: &mut App| {
1196 state.update(cx, |state, cx| {
1197 state.accept_terms_of_service(cx);
1198 });
1199 }
1200 });
1201
1202 let sign_in_callback = Arc::new({
1203 let state = state.clone();
1204 move |_window: &mut Window, cx: &mut App| {
1205 state.update(cx, |state, cx| {
1206 state.authenticate(cx).detach_and_log_err(cx);
1207 });
1208 }
1209 });
1210
1211 Self {
1212 state,
1213 accept_terms_of_service_callback,
1214 sign_in_callback,
1215 }
1216 }
1217}
1218
1219impl Render for ConfigurationView {
1220 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1221 let state = self.state.read(cx);
1222 let user_store = state.user_store.read(cx);
1223
1224 ZedAiConfiguration {
1225 is_connected: !state.is_signed_out(),
1226 plan: user_store.current_plan(),
1227 subscription_period: user_store.subscription_period(),
1228 eligible_for_trial: user_store.trial_started_at().is_none(),
1229 has_accepted_terms_of_service: state.has_accepted_terms_of_service(cx),
1230 accept_terms_of_service_in_progress: state.accept_terms_of_service_task.is_some(),
1231 accept_terms_of_service_callback: self.accept_terms_of_service_callback.clone(),
1232 sign_in_callback: self.sign_in_callback.clone(),
1233 }
1234 }
1235}
1236
1237impl Component for ZedAiConfiguration {
1238 fn scope() -> ComponentScope {
1239 ComponentScope::Agent
1240 }
1241
1242 fn preview(_window: &mut Window, _cx: &mut App) -> Option<AnyElement> {
1243 fn configuration(
1244 is_connected: bool,
1245 plan: Option<proto::Plan>,
1246 eligible_for_trial: bool,
1247 has_accepted_terms_of_service: bool,
1248 ) -> AnyElement {
1249 ZedAiConfiguration {
1250 is_connected,
1251 plan,
1252 subscription_period: plan
1253 .is_some()
1254 .then(|| (Utc::now(), Utc::now() + chrono::Duration::days(7))),
1255 eligible_for_trial,
1256 has_accepted_terms_of_service,
1257 accept_terms_of_service_in_progress: false,
1258 accept_terms_of_service_callback: Arc::new(|_, _| {}),
1259 sign_in_callback: Arc::new(|_, _| {}),
1260 }
1261 .into_any_element()
1262 }
1263
1264 Some(
1265 v_flex()
1266 .p_4()
1267 .gap_4()
1268 .children(vec![
1269 single_example("Not connected", configuration(false, None, false, true)),
1270 single_example(
1271 "Accept Terms of Service",
1272 configuration(true, None, true, false),
1273 ),
1274 single_example(
1275 "No Plan - Not eligible for trial",
1276 configuration(true, None, false, true),
1277 ),
1278 single_example(
1279 "No Plan - Eligible for trial",
1280 configuration(true, None, true, true),
1281 ),
1282 single_example(
1283 "Free Plan",
1284 configuration(true, Some(proto::Plan::Free), true, true),
1285 ),
1286 single_example(
1287 "Zed Pro Trial Plan",
1288 configuration(true, Some(proto::Plan::ZedProTrial), true, true),
1289 ),
1290 single_example(
1291 "Zed Pro Plan",
1292 configuration(true, Some(proto::Plan::ZedPro), true, true),
1293 ),
1294 ])
1295 .into_any_element(),
1296 )
1297 }
1298}