1use anthropic::{AnthropicModelMode, parse_prompt_too_long};
2use anyhow::{Result, anyhow};
3use client::{Client, UserStore, zed_urls};
4use collections::BTreeMap;
5use feature_flags::{FeatureFlagAppExt, LlmClosedBetaFeatureFlag};
6use futures::{
7 AsyncBufReadExt, FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream,
8};
9use gpui::{
10 AnyElement, AnyView, App, AsyncApp, Context, Entity, SemanticVersion, Subscription, Task,
11};
12use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
13use language_model::{
14 AuthenticateError, CloudModel, LanguageModel, LanguageModelCacheConfiguration,
15 LanguageModelCompletionError, LanguageModelId, LanguageModelKnownError, LanguageModelName,
16 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
17 LanguageModelProviderTosView, LanguageModelRequest, LanguageModelToolChoice,
18 LanguageModelToolSchemaFormat, ModelRequestLimitReachedError, RateLimiter, RequestUsage,
19 ZED_CLOUD_PROVIDER_ID,
20};
21use language_model::{
22 LanguageModelCompletionEvent, LanguageModelProvider, LlmApiToken, PaymentRequiredError,
23 RefreshLlmTokenListener,
24};
25use proto::Plan;
26use release_channel::AppVersion;
27use schemars::JsonSchema;
28use serde::{Deserialize, Serialize, de::DeserializeOwned};
29use settings::{Settings, SettingsStore};
30use smol::Timer;
31use smol::io::{AsyncReadExt, BufReader};
32use std::pin::Pin;
33use std::str::FromStr as _;
34use std::{
35 sync::{Arc, LazyLock},
36 time::Duration,
37};
38use strum::IntoEnumIterator;
39use thiserror::Error;
40use ui::{TintColor, prelude::*};
41use zed_llm_client::{
42 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
43 CompletionRequestStatus, CountTokensBody, CountTokensResponse, EXPIRED_LLM_TOKEN_HEADER_NAME,
44 MODEL_REQUESTS_RESOURCE_HEADER_VALUE, SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME,
45 SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME, TOOL_USE_LIMIT_REACHED_HEADER_NAME,
46 ZED_VERSION_HEADER_NAME,
47};
48
49use crate::AllLanguageModelSettings;
50use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
51use crate::provider::google::{GoogleEventMapper, into_google};
52use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
53
54pub const PROVIDER_NAME: &str = "Zed";
55
56const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
57 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
58
59fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
60 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
61 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
62 .map(|json| serde_json::from_str(json).unwrap())
63 .unwrap_or_default()
64 });
65 ADDITIONAL_MODELS.as_slice()
66}
67
68#[derive(Default, Clone, Debug, PartialEq)]
69pub struct ZedDotDevSettings {
70 pub available_models: Vec<AvailableModel>,
71}
72
73#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
74#[serde(rename_all = "lowercase")]
75pub enum AvailableProvider {
76 Anthropic,
77 OpenAi,
78 Google,
79}
80
81#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
82pub struct AvailableModel {
83 /// The provider of the language model.
84 pub provider: AvailableProvider,
85 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
86 pub name: String,
87 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
88 pub display_name: Option<String>,
89 /// The size of the context window, indicating the maximum number of tokens the model can process.
90 pub max_tokens: usize,
91 /// The maximum number of output tokens allowed by the model.
92 pub max_output_tokens: Option<u32>,
93 /// The maximum number of completion tokens allowed by the model (o1-* only)
94 pub max_completion_tokens: Option<u32>,
95 /// Override this model with a different Anthropic model for tool calls.
96 pub tool_override: Option<String>,
97 /// Indicates whether this custom model supports caching.
98 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
99 /// The default temperature to use for this model.
100 pub default_temperature: Option<f32>,
101 /// Any extra beta headers to provide when using the model.
102 #[serde(default)]
103 pub extra_beta_headers: Vec<String>,
104 /// The model's mode (e.g. thinking)
105 pub mode: Option<ModelMode>,
106}
107
108#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
109#[serde(tag = "type", rename_all = "lowercase")]
110pub enum ModelMode {
111 #[default]
112 Default,
113 Thinking {
114 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
115 budget_tokens: Option<u32>,
116 },
117}
118
119impl From<ModelMode> for AnthropicModelMode {
120 fn from(value: ModelMode) -> Self {
121 match value {
122 ModelMode::Default => AnthropicModelMode::Default,
123 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
124 }
125 }
126}
127
128pub struct CloudLanguageModelProvider {
129 client: Arc<Client>,
130 state: gpui::Entity<State>,
131 _maintain_client_status: Task<()>,
132}
133
134pub struct State {
135 client: Arc<Client>,
136 llm_api_token: LlmApiToken,
137 user_store: Entity<UserStore>,
138 status: client::Status,
139 accept_terms: Option<Task<Result<()>>>,
140 _settings_subscription: Subscription,
141 _llm_token_subscription: Subscription,
142}
143
144impl State {
145 fn new(
146 client: Arc<Client>,
147 user_store: Entity<UserStore>,
148 status: client::Status,
149 cx: &mut Context<Self>,
150 ) -> Self {
151 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
152
153 Self {
154 client: client.clone(),
155 llm_api_token: LlmApiToken::default(),
156 user_store,
157 status,
158 accept_terms: None,
159 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
160 cx.notify();
161 }),
162 _llm_token_subscription: cx.subscribe(
163 &refresh_llm_token_listener,
164 |this, _listener, _event, cx| {
165 let client = this.client.clone();
166 let llm_api_token = this.llm_api_token.clone();
167 cx.spawn(async move |_this, _cx| {
168 llm_api_token.refresh(&client).await?;
169 anyhow::Ok(())
170 })
171 .detach_and_log_err(cx);
172 },
173 ),
174 }
175 }
176
177 fn is_signed_out(&self) -> bool {
178 self.status.is_signed_out()
179 }
180
181 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
182 let client = self.client.clone();
183 cx.spawn(async move |state, cx| {
184 client
185 .authenticate_and_connect(true, &cx)
186 .await
187 .into_response()?;
188 state.update(cx, |_, cx| cx.notify())
189 })
190 }
191
192 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
193 self.user_store
194 .read(cx)
195 .current_user_has_accepted_terms()
196 .unwrap_or(false)
197 }
198
199 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
200 let user_store = self.user_store.clone();
201 self.accept_terms = Some(cx.spawn(async move |this, cx| {
202 let _ = user_store
203 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
204 .await;
205 this.update(cx, |this, cx| {
206 this.accept_terms = None;
207 cx.notify()
208 })
209 }));
210 }
211}
212
213impl CloudLanguageModelProvider {
214 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
215 let mut status_rx = client.status();
216 let status = *status_rx.borrow();
217
218 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
219
220 let state_ref = state.downgrade();
221 let maintain_client_status = cx.spawn(async move |cx| {
222 while let Some(status) = status_rx.next().await {
223 if let Some(this) = state_ref.upgrade() {
224 _ = this.update(cx, |this, cx| {
225 if this.status != status {
226 this.status = status;
227 cx.notify();
228 }
229 });
230 } else {
231 break;
232 }
233 }
234 });
235
236 Self {
237 client,
238 state: state.clone(),
239 _maintain_client_status: maintain_client_status,
240 }
241 }
242
243 fn create_language_model(
244 &self,
245 model: CloudModel,
246 llm_api_token: LlmApiToken,
247 ) -> Arc<dyn LanguageModel> {
248 Arc::new(CloudLanguageModel {
249 id: LanguageModelId::from(model.id().to_string()),
250 model,
251 llm_api_token: llm_api_token.clone(),
252 client: self.client.clone(),
253 request_limiter: RateLimiter::new(4),
254 })
255 }
256}
257
258impl LanguageModelProviderState for CloudLanguageModelProvider {
259 type ObservableEntity = State;
260
261 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
262 Some(self.state.clone())
263 }
264}
265
266impl LanguageModelProvider for CloudLanguageModelProvider {
267 fn id(&self) -> LanguageModelProviderId {
268 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
269 }
270
271 fn name(&self) -> LanguageModelProviderName {
272 LanguageModelProviderName(PROVIDER_NAME.into())
273 }
274
275 fn icon(&self) -> IconName {
276 IconName::AiZed
277 }
278
279 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
280 let llm_api_token = self.state.read(cx).llm_api_token.clone();
281 let model = CloudModel::Anthropic(anthropic::Model::ClaudeSonnet4);
282 Some(self.create_language_model(model, llm_api_token))
283 }
284
285 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
286 let llm_api_token = self.state.read(cx).llm_api_token.clone();
287 let model = CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet);
288 Some(self.create_language_model(model, llm_api_token))
289 }
290
291 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
292 let llm_api_token = self.state.read(cx).llm_api_token.clone();
293 [
294 CloudModel::Anthropic(anthropic::Model::ClaudeSonnet4),
295 CloudModel::Anthropic(anthropic::Model::ClaudeSonnet4Thinking),
296 ]
297 .into_iter()
298 .map(|model| self.create_language_model(model, llm_api_token.clone()))
299 .collect()
300 }
301
302 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
303 let mut models = BTreeMap::default();
304
305 if cx.is_staff() {
306 for model in anthropic::Model::iter() {
307 if !matches!(model, anthropic::Model::Custom { .. }) {
308 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
309 }
310 }
311 for model in open_ai::Model::iter() {
312 if !matches!(model, open_ai::Model::Custom { .. }) {
313 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
314 }
315 }
316 for model in google_ai::Model::iter() {
317 if !matches!(model, google_ai::Model::Custom { .. }) {
318 models.insert(model.id().to_string(), CloudModel::Google(model));
319 }
320 }
321 } else {
322 models.insert(
323 anthropic::Model::Claude3_5Sonnet.id().to_string(),
324 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
325 );
326 models.insert(
327 anthropic::Model::Claude3_7Sonnet.id().to_string(),
328 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
329 );
330 models.insert(
331 anthropic::Model::Claude3_7SonnetThinking.id().to_string(),
332 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
333 );
334 models.insert(
335 anthropic::Model::ClaudeSonnet4.id().to_string(),
336 CloudModel::Anthropic(anthropic::Model::ClaudeSonnet4),
337 );
338 models.insert(
339 anthropic::Model::ClaudeSonnet4Thinking.id().to_string(),
340 CloudModel::Anthropic(anthropic::Model::ClaudeSonnet4Thinking),
341 );
342 }
343
344 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBetaFeatureFlag>() {
345 zed_cloud_provider_additional_models()
346 } else {
347 &[]
348 };
349
350 // Override with available models from settings
351 for model in AllLanguageModelSettings::get_global(cx)
352 .zed_dot_dev
353 .available_models
354 .iter()
355 .chain(llm_closed_beta_models)
356 .cloned()
357 {
358 let model = match model.provider {
359 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
360 name: model.name.clone(),
361 display_name: model.display_name.clone(),
362 max_tokens: model.max_tokens,
363 tool_override: model.tool_override.clone(),
364 cache_configuration: model.cache_configuration.as_ref().map(|config| {
365 anthropic::AnthropicModelCacheConfiguration {
366 max_cache_anchors: config.max_cache_anchors,
367 should_speculate: config.should_speculate,
368 min_total_token: config.min_total_token,
369 }
370 }),
371 default_temperature: model.default_temperature,
372 max_output_tokens: model.max_output_tokens,
373 extra_beta_headers: model.extra_beta_headers.clone(),
374 mode: model.mode.unwrap_or_default().into(),
375 }),
376 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
377 name: model.name.clone(),
378 display_name: model.display_name.clone(),
379 max_tokens: model.max_tokens,
380 max_output_tokens: model.max_output_tokens,
381 max_completion_tokens: model.max_completion_tokens,
382 }),
383 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
384 name: model.name.clone(),
385 display_name: model.display_name.clone(),
386 max_tokens: model.max_tokens,
387 }),
388 };
389 models.insert(model.id().to_string(), model.clone());
390 }
391
392 let llm_api_token = self.state.read(cx).llm_api_token.clone();
393 models
394 .into_values()
395 .map(|model| self.create_language_model(model, llm_api_token.clone()))
396 .collect()
397 }
398
399 fn is_authenticated(&self, cx: &App) -> bool {
400 !self.state.read(cx).is_signed_out()
401 }
402
403 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
404 Task::ready(Ok(()))
405 }
406
407 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
408 cx.new(|_| ConfigurationView {
409 state: self.state.clone(),
410 })
411 .into()
412 }
413
414 fn must_accept_terms(&self, cx: &App) -> bool {
415 !self.state.read(cx).has_accepted_terms_of_service(cx)
416 }
417
418 fn render_accept_terms(
419 &self,
420 view: LanguageModelProviderTosView,
421 cx: &mut App,
422 ) -> Option<AnyElement> {
423 render_accept_terms(self.state.clone(), view, cx)
424 }
425
426 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
427 Task::ready(Ok(()))
428 }
429}
430
431fn render_accept_terms(
432 state: Entity<State>,
433 view_kind: LanguageModelProviderTosView,
434 cx: &mut App,
435) -> Option<AnyElement> {
436 if state.read(cx).has_accepted_terms_of_service(cx) {
437 return None;
438 }
439
440 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
441
442 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
443 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadtEmptyState);
444
445 let terms_button = Button::new("terms_of_service", "Terms of Service")
446 .style(ButtonStyle::Subtle)
447 .icon(IconName::ArrowUpRight)
448 .icon_color(Color::Muted)
449 .icon_size(IconSize::XSmall)
450 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
451 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
452
453 let button_container = h_flex().child(
454 Button::new("accept_terms", "I accept the Terms of Service")
455 .when(!thread_empty_state, |this| {
456 this.full_width()
457 .style(ButtonStyle::Tinted(TintColor::Accent))
458 .icon(IconName::Check)
459 .icon_position(IconPosition::Start)
460 .icon_size(IconSize::Small)
461 })
462 .when(thread_empty_state, |this| {
463 this.style(ButtonStyle::Tinted(TintColor::Warning))
464 .label_size(LabelSize::Small)
465 })
466 .disabled(accept_terms_disabled)
467 .on_click({
468 let state = state.downgrade();
469 move |_, _window, cx| {
470 state
471 .update(cx, |state, cx| state.accept_terms_of_service(cx))
472 .ok();
473 }
474 }),
475 );
476
477 let form = if thread_empty_state {
478 h_flex()
479 .w_full()
480 .flex_wrap()
481 .justify_between()
482 .child(
483 h_flex()
484 .child(
485 Label::new("To start using Zed AI, please read and accept the")
486 .size(LabelSize::Small),
487 )
488 .child(terms_button),
489 )
490 .child(button_container)
491 } else {
492 v_flex()
493 .w_full()
494 .gap_2()
495 .child(
496 h_flex()
497 .flex_wrap()
498 .when(thread_fresh_start, |this| this.justify_center())
499 .child(Label::new(
500 "To start using Zed AI, please read and accept the",
501 ))
502 .child(terms_button),
503 )
504 .child({
505 match view_kind {
506 LanguageModelProviderTosView::PromptEditorPopup => {
507 button_container.w_full().justify_end()
508 }
509 LanguageModelProviderTosView::Configuration => {
510 button_container.w_full().justify_start()
511 }
512 LanguageModelProviderTosView::ThreadFreshStart => {
513 button_container.w_full().justify_center()
514 }
515 LanguageModelProviderTosView::ThreadtEmptyState => div().w_0(),
516 }
517 })
518 };
519
520 Some(form.into_any())
521}
522
523pub struct CloudLanguageModel {
524 id: LanguageModelId,
525 model: CloudModel,
526 llm_api_token: LlmApiToken,
527 client: Arc<Client>,
528 request_limiter: RateLimiter,
529}
530
531struct PerformLlmCompletionResponse {
532 response: Response<AsyncBody>,
533 usage: Option<RequestUsage>,
534 tool_use_limit_reached: bool,
535 includes_status_messages: bool,
536}
537
538impl CloudLanguageModel {
539 const MAX_RETRIES: usize = 3;
540
541 async fn perform_llm_completion(
542 client: Arc<Client>,
543 llm_api_token: LlmApiToken,
544 app_version: Option<SemanticVersion>,
545 body: CompletionBody,
546 ) -> Result<PerformLlmCompletionResponse> {
547 let http_client = &client.http_client();
548
549 let mut token = llm_api_token.acquire(&client).await?;
550 let mut retries_remaining = Self::MAX_RETRIES;
551 let mut retry_delay = Duration::from_secs(1);
552
553 loop {
554 let request_builder = http_client::Request::builder()
555 .method(Method::POST)
556 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
557 let request_builder = if let Some(app_version) = app_version {
558 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
559 } else {
560 request_builder
561 };
562
563 let request = request_builder
564 .header("Content-Type", "application/json")
565 .header("Authorization", format!("Bearer {token}"))
566 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
567 .body(serde_json::to_string(&body)?.into())?;
568 let mut response = http_client.send(request).await?;
569 let status = response.status();
570 if status.is_success() {
571 let includes_status_messages = response
572 .headers()
573 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
574 .is_some();
575
576 let tool_use_limit_reached = response
577 .headers()
578 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
579 .is_some();
580
581 let usage = if includes_status_messages {
582 None
583 } else {
584 RequestUsage::from_headers(response.headers()).ok()
585 };
586
587 return Ok(PerformLlmCompletionResponse {
588 response,
589 usage,
590 includes_status_messages,
591 tool_use_limit_reached,
592 });
593 } else if response
594 .headers()
595 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
596 .is_some()
597 {
598 retries_remaining -= 1;
599 token = llm_api_token.refresh(&client).await?;
600 } else if status == StatusCode::FORBIDDEN
601 && response
602 .headers()
603 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
604 .is_some()
605 {
606 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
607 .headers()
608 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
609 .and_then(|resource| resource.to_str().ok())
610 {
611 if let Some(plan) = response
612 .headers()
613 .get(CURRENT_PLAN_HEADER_NAME)
614 .and_then(|plan| plan.to_str().ok())
615 .and_then(|plan| zed_llm_client::Plan::from_str(plan).ok())
616 {
617 let plan = match plan {
618 zed_llm_client::Plan::ZedFree => Plan::Free,
619 zed_llm_client::Plan::ZedPro => Plan::ZedPro,
620 zed_llm_client::Plan::ZedProTrial => Plan::ZedProTrial,
621 };
622 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
623 }
624 }
625
626 anyhow::bail!("Forbidden");
627 } else if status.as_u16() >= 500 && status.as_u16() < 600 {
628 // If we encounter an error in the 500 range, retry after a delay.
629 // We've seen at least these in the wild from API providers:
630 // * 500 Internal Server Error
631 // * 502 Bad Gateway
632 // * 529 Service Overloaded
633
634 if retries_remaining == 0 {
635 let mut body = String::new();
636 response.body_mut().read_to_string(&mut body).await?;
637 anyhow::bail!(
638 "cloud language model completion failed after {} retries with status {status}: {body}",
639 Self::MAX_RETRIES
640 );
641 }
642
643 Timer::after(retry_delay).await;
644
645 retries_remaining -= 1;
646 retry_delay *= 2; // If it fails again, wait longer.
647 } else if status == StatusCode::PAYMENT_REQUIRED {
648 return Err(anyhow!(PaymentRequiredError));
649 } else {
650 let mut body = String::new();
651 response.body_mut().read_to_string(&mut body).await?;
652 return Err(anyhow!(ApiError { status, body }));
653 }
654 }
655 }
656}
657
658#[derive(Debug, Error)]
659#[error("cloud language model request failed with status {status}: {body}")]
660struct ApiError {
661 status: StatusCode,
662 body: String,
663}
664
665impl LanguageModel for CloudLanguageModel {
666 fn id(&self) -> LanguageModelId {
667 self.id.clone()
668 }
669
670 fn name(&self) -> LanguageModelName {
671 LanguageModelName::from(self.model.display_name().to_string())
672 }
673
674 fn provider_id(&self) -> LanguageModelProviderId {
675 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
676 }
677
678 fn provider_name(&self) -> LanguageModelProviderName {
679 LanguageModelProviderName(PROVIDER_NAME.into())
680 }
681
682 fn supports_tools(&self) -> bool {
683 match self.model {
684 CloudModel::Anthropic(_) => true,
685 CloudModel::Google(_) => true,
686 CloudModel::OpenAi(_) => true,
687 }
688 }
689
690 fn supports_images(&self) -> bool {
691 match self.model {
692 CloudModel::Anthropic(_) => true,
693 CloudModel::Google(_) => true,
694 CloudModel::OpenAi(_) => false,
695 }
696 }
697
698 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
699 match choice {
700 LanguageModelToolChoice::Auto
701 | LanguageModelToolChoice::Any
702 | LanguageModelToolChoice::None => true,
703 }
704 }
705
706 fn telemetry_id(&self) -> String {
707 format!("zed.dev/{}", self.model.id())
708 }
709
710 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
711 self.model.tool_input_format()
712 }
713
714 fn max_token_count(&self) -> usize {
715 self.model.max_token_count()
716 }
717
718 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
719 match &self.model {
720 CloudModel::Anthropic(model) => {
721 model
722 .cache_configuration()
723 .map(|cache| LanguageModelCacheConfiguration {
724 max_cache_anchors: cache.max_cache_anchors,
725 should_speculate: cache.should_speculate,
726 min_total_token: cache.min_total_token,
727 })
728 }
729 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
730 }
731 }
732
733 fn count_tokens(
734 &self,
735 request: LanguageModelRequest,
736 cx: &App,
737 ) -> BoxFuture<'static, Result<usize>> {
738 match self.model.clone() {
739 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
740 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
741 CloudModel::Google(model) => {
742 let client = self.client.clone();
743 let llm_api_token = self.llm_api_token.clone();
744 let model_id = model.id().to_string();
745 let generate_content_request = into_google(request, model_id.clone());
746 async move {
747 let http_client = &client.http_client();
748 let token = llm_api_token.acquire(&client).await?;
749
750 let request_body = CountTokensBody {
751 provider: zed_llm_client::LanguageModelProvider::Google,
752 model: model_id,
753 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
754 generate_content_request,
755 })?,
756 };
757 let request = http_client::Request::builder()
758 .method(Method::POST)
759 .uri(
760 http_client
761 .build_zed_llm_url("/count_tokens", &[])?
762 .as_ref(),
763 )
764 .header("Content-Type", "application/json")
765 .header("Authorization", format!("Bearer {token}"))
766 .body(serde_json::to_string(&request_body)?.into())?;
767 let mut response = http_client.send(request).await?;
768 let status = response.status();
769 let mut response_body = String::new();
770 response
771 .body_mut()
772 .read_to_string(&mut response_body)
773 .await?;
774
775 if status.is_success() {
776 let response_body: CountTokensResponse =
777 serde_json::from_str(&response_body)?;
778
779 Ok(response_body.tokens)
780 } else {
781 Err(anyhow!(ApiError {
782 status,
783 body: response_body
784 }))
785 }
786 }
787 .boxed()
788 }
789 }
790 }
791
792 fn stream_completion(
793 &self,
794 request: LanguageModelRequest,
795 cx: &AsyncApp,
796 ) -> BoxFuture<
797 'static,
798 Result<
799 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
800 >,
801 > {
802 let thread_id = request.thread_id.clone();
803 let prompt_id = request.prompt_id.clone();
804 let intent = request.intent;
805 let mode = request.mode;
806 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
807 match &self.model {
808 CloudModel::Anthropic(model) => {
809 let request = into_anthropic(
810 request,
811 model.request_id().into(),
812 model.default_temperature(),
813 model.max_output_tokens(),
814 model.mode(),
815 );
816 let client = self.client.clone();
817 let llm_api_token = self.llm_api_token.clone();
818 let future = self.request_limiter.stream(async move {
819 let PerformLlmCompletionResponse {
820 response,
821 usage,
822 includes_status_messages,
823 tool_use_limit_reached,
824 } = Self::perform_llm_completion(
825 client.clone(),
826 llm_api_token,
827 app_version,
828 CompletionBody {
829 thread_id,
830 prompt_id,
831 intent,
832 mode,
833 provider: zed_llm_client::LanguageModelProvider::Anthropic,
834 model: request.model.clone(),
835 provider_request: serde_json::to_value(&request)?,
836 },
837 )
838 .await
839 .map_err(|err| match err.downcast::<ApiError>() {
840 Ok(api_err) => {
841 if api_err.status == StatusCode::BAD_REQUEST {
842 if let Some(tokens) = parse_prompt_too_long(&api_err.body) {
843 return anyhow!(
844 LanguageModelKnownError::ContextWindowLimitExceeded {
845 tokens
846 }
847 );
848 }
849 }
850 anyhow!(api_err)
851 }
852 Err(err) => anyhow!(err),
853 })?;
854
855 let mut mapper = AnthropicEventMapper::new();
856 Ok(map_cloud_completion_events(
857 Box::pin(
858 response_lines(response, includes_status_messages)
859 .chain(usage_updated_event(usage))
860 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
861 ),
862 move |event| mapper.map_event(event),
863 ))
864 });
865 async move { Ok(future.await?.boxed()) }.boxed()
866 }
867 CloudModel::OpenAi(model) => {
868 let client = self.client.clone();
869 let request = into_open_ai(request, model, model.max_output_tokens());
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::OpenAi,
887 model: request.model.clone(),
888 provider_request: serde_json::to_value(&request)?,
889 },
890 )
891 .await?;
892
893 let mut mapper = OpenAiEventMapper::new();
894 Ok(map_cloud_completion_events(
895 Box::pin(
896 response_lines(response, includes_status_messages)
897 .chain(usage_updated_event(usage))
898 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
899 ),
900 move |event| mapper.map_event(event),
901 ))
902 });
903 async move { Ok(future.await?.boxed()) }.boxed()
904 }
905 CloudModel::Google(model) => {
906 let client = self.client.clone();
907 let request = into_google(request, model.id().into());
908 let llm_api_token = self.llm_api_token.clone();
909 let future = self.request_limiter.stream(async move {
910 let PerformLlmCompletionResponse {
911 response,
912 usage,
913 includes_status_messages,
914 tool_use_limit_reached,
915 } = Self::perform_llm_completion(
916 client.clone(),
917 llm_api_token,
918 app_version,
919 CompletionBody {
920 thread_id,
921 prompt_id,
922 intent,
923 mode,
924 provider: zed_llm_client::LanguageModelProvider::Google,
925 model: request.model.model_id.clone(),
926 provider_request: serde_json::to_value(&request)?,
927 },
928 )
929 .await?;
930
931 let mut mapper = GoogleEventMapper::new();
932 Ok(map_cloud_completion_events(
933 Box::pin(
934 response_lines(response, includes_status_messages)
935 .chain(usage_updated_event(usage))
936 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
937 ),
938 move |event| mapper.map_event(event),
939 ))
940 });
941 async move { Ok(future.await?.boxed()) }.boxed()
942 }
943 }
944 }
945}
946
947#[derive(Serialize, Deserialize)]
948#[serde(rename_all = "snake_case")]
949pub enum CloudCompletionEvent<T> {
950 Status(CompletionRequestStatus),
951 Event(T),
952}
953
954fn map_cloud_completion_events<T, F>(
955 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
956 mut map_callback: F,
957) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
958where
959 T: DeserializeOwned + 'static,
960 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
961 + Send
962 + 'static,
963{
964 stream
965 .flat_map(move |event| {
966 futures::stream::iter(match event {
967 Err(error) => {
968 vec![Err(LanguageModelCompletionError::Other(error))]
969 }
970 Ok(CloudCompletionEvent::Status(event)) => {
971 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
972 }
973 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
974 })
975 })
976 .boxed()
977}
978
979fn usage_updated_event<T>(
980 usage: Option<RequestUsage>,
981) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
982 futures::stream::iter(usage.map(|usage| {
983 Ok(CloudCompletionEvent::Status(
984 CompletionRequestStatus::UsageUpdated {
985 amount: usage.amount as usize,
986 limit: usage.limit,
987 },
988 ))
989 }))
990}
991
992fn tool_use_limit_reached_event<T>(
993 tool_use_limit_reached: bool,
994) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
995 futures::stream::iter(tool_use_limit_reached.then(|| {
996 Ok(CloudCompletionEvent::Status(
997 CompletionRequestStatus::ToolUseLimitReached,
998 ))
999 }))
1000}
1001
1002fn response_lines<T: DeserializeOwned>(
1003 response: Response<AsyncBody>,
1004 includes_status_messages: bool,
1005) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1006 futures::stream::try_unfold(
1007 (String::new(), BufReader::new(response.into_body())),
1008 move |(mut line, mut body)| async move {
1009 match body.read_line(&mut line).await {
1010 Ok(0) => Ok(None),
1011 Ok(_) => {
1012 let event = if includes_status_messages {
1013 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
1014 } else {
1015 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1016 };
1017
1018 line.clear();
1019 Ok(Some((event, (line, body))))
1020 }
1021 Err(e) => Err(e.into()),
1022 }
1023 },
1024 )
1025}
1026
1027struct ConfigurationView {
1028 state: gpui::Entity<State>,
1029}
1030
1031impl ConfigurationView {
1032 fn authenticate(&mut self, cx: &mut Context<Self>) {
1033 self.state.update(cx, |state, cx| {
1034 state.authenticate(cx).detach_and_log_err(cx);
1035 });
1036 cx.notify();
1037 }
1038}
1039
1040impl Render for ConfigurationView {
1041 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1042 const ZED_PRICING_URL: &str = "https://zed.dev/pricing";
1043
1044 let is_connected = !self.state.read(cx).is_signed_out();
1045 let user_store = self.state.read(cx).user_store.read(cx);
1046 let plan = user_store.current_plan();
1047 let subscription_period = user_store.subscription_period();
1048 let eligible_for_trial = user_store.trial_started_at().is_none();
1049 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
1050
1051 let is_pro = plan == Some(proto::Plan::ZedPro);
1052 let subscription_text = match (plan, subscription_period) {
1053 (Some(proto::Plan::ZedPro), Some(_)) => {
1054 "You have access to Zed's hosted LLMs through your Zed Pro subscription."
1055 }
1056 (Some(proto::Plan::ZedProTrial), Some(_)) => {
1057 "You have access to Zed's hosted LLMs through your Zed Pro trial."
1058 }
1059 (Some(proto::Plan::Free), Some(_)) => {
1060 "You have basic access to Zed's hosted LLMs through your Zed Free subscription."
1061 }
1062 _ => {
1063 if eligible_for_trial {
1064 "Subscribe for access to Zed's hosted LLMs. Start with a 14 day free trial."
1065 } else {
1066 "Subscribe for access to Zed's hosted LLMs."
1067 }
1068 }
1069 };
1070 let manage_subscription_buttons = if is_pro {
1071 h_flex().child(
1072 Button::new("manage_settings", "Manage Subscription")
1073 .style(ButtonStyle::Tinted(TintColor::Accent))
1074 .on_click(cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx)))),
1075 )
1076 } else {
1077 h_flex()
1078 .gap_2()
1079 .child(
1080 Button::new("learn_more", "Learn more")
1081 .style(ButtonStyle::Subtle)
1082 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_PRICING_URL))),
1083 )
1084 .child(
1085 Button::new("upgrade", "Upgrade")
1086 .style(ButtonStyle::Subtle)
1087 .color(Color::Accent)
1088 .on_click(
1089 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1090 ),
1091 )
1092 };
1093
1094 if is_connected {
1095 v_flex()
1096 .gap_3()
1097 .w_full()
1098 .children(render_accept_terms(
1099 self.state.clone(),
1100 LanguageModelProviderTosView::Configuration,
1101 cx,
1102 ))
1103 .when(has_accepted_terms, |this| {
1104 this.child(subscription_text)
1105 .child(manage_subscription_buttons)
1106 })
1107 } else {
1108 v_flex()
1109 .gap_2()
1110 .child(Label::new("Use Zed AI to access hosted language models."))
1111 .child(
1112 Button::new("sign_in", "Sign In")
1113 .icon_color(Color::Muted)
1114 .icon(IconName::Github)
1115 .icon_position(IconPosition::Start)
1116 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
1117 )
1118 }
1119 }
1120}