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 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider, LlmApiToken,
23 MaxMonthlySpendReachedError, PaymentRequiredError, 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 MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME, MODEL_REQUESTS_RESOURCE_HEADER_VALUE,
45 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
46 TOOL_USE_LIMIT_REACHED_HEADER_NAME, 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::Claude3_7Sonnet);
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::Claude3_7Sonnet),
295 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
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 }
335
336 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBetaFeatureFlag>() {
337 zed_cloud_provider_additional_models()
338 } else {
339 &[]
340 };
341
342 // Override with available models from settings
343 for model in AllLanguageModelSettings::get_global(cx)
344 .zed_dot_dev
345 .available_models
346 .iter()
347 .chain(llm_closed_beta_models)
348 .cloned()
349 {
350 let model = match model.provider {
351 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
352 name: model.name.clone(),
353 display_name: model.display_name.clone(),
354 max_tokens: model.max_tokens,
355 tool_override: model.tool_override.clone(),
356 cache_configuration: model.cache_configuration.as_ref().map(|config| {
357 anthropic::AnthropicModelCacheConfiguration {
358 max_cache_anchors: config.max_cache_anchors,
359 should_speculate: config.should_speculate,
360 min_total_token: config.min_total_token,
361 }
362 }),
363 default_temperature: model.default_temperature,
364 max_output_tokens: model.max_output_tokens,
365 extra_beta_headers: model.extra_beta_headers.clone(),
366 mode: model.mode.unwrap_or_default().into(),
367 }),
368 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
369 name: model.name.clone(),
370 display_name: model.display_name.clone(),
371 max_tokens: model.max_tokens,
372 max_output_tokens: model.max_output_tokens,
373 max_completion_tokens: model.max_completion_tokens,
374 }),
375 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
376 name: model.name.clone(),
377 display_name: model.display_name.clone(),
378 max_tokens: model.max_tokens,
379 }),
380 };
381 models.insert(model.id().to_string(), model.clone());
382 }
383
384 let llm_api_token = self.state.read(cx).llm_api_token.clone();
385 models
386 .into_values()
387 .map(|model| self.create_language_model(model, llm_api_token.clone()))
388 .collect()
389 }
390
391 fn is_authenticated(&self, cx: &App) -> bool {
392 !self.state.read(cx).is_signed_out()
393 }
394
395 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
396 Task::ready(Ok(()))
397 }
398
399 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
400 cx.new(|_| ConfigurationView {
401 state: self.state.clone(),
402 })
403 .into()
404 }
405
406 fn must_accept_terms(&self, cx: &App) -> bool {
407 !self.state.read(cx).has_accepted_terms_of_service(cx)
408 }
409
410 fn render_accept_terms(
411 &self,
412 view: LanguageModelProviderTosView,
413 cx: &mut App,
414 ) -> Option<AnyElement> {
415 render_accept_terms(self.state.clone(), view, cx)
416 }
417
418 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
419 Task::ready(Ok(()))
420 }
421}
422
423fn render_accept_terms(
424 state: Entity<State>,
425 view_kind: LanguageModelProviderTosView,
426 cx: &mut App,
427) -> Option<AnyElement> {
428 if state.read(cx).has_accepted_terms_of_service(cx) {
429 return None;
430 }
431
432 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
433
434 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
435 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadtEmptyState);
436
437 let terms_button = Button::new("terms_of_service", "Terms of Service")
438 .style(ButtonStyle::Subtle)
439 .icon(IconName::ArrowUpRight)
440 .icon_color(Color::Muted)
441 .icon_size(IconSize::XSmall)
442 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
443 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
444
445 let button_container = h_flex().child(
446 Button::new("accept_terms", "I accept the Terms of Service")
447 .when(!thread_empty_state, |this| {
448 this.full_width()
449 .style(ButtonStyle::Tinted(TintColor::Accent))
450 .icon(IconName::Check)
451 .icon_position(IconPosition::Start)
452 .icon_size(IconSize::Small)
453 })
454 .when(thread_empty_state, |this| {
455 this.style(ButtonStyle::Tinted(TintColor::Warning))
456 .label_size(LabelSize::Small)
457 })
458 .disabled(accept_terms_disabled)
459 .on_click({
460 let state = state.downgrade();
461 move |_, _window, cx| {
462 state
463 .update(cx, |state, cx| state.accept_terms_of_service(cx))
464 .ok();
465 }
466 }),
467 );
468
469 let form = if thread_empty_state {
470 h_flex()
471 .w_full()
472 .flex_wrap()
473 .justify_between()
474 .child(
475 h_flex()
476 .child(
477 Label::new("To start using Zed AI, please read and accept the")
478 .size(LabelSize::Small),
479 )
480 .child(terms_button),
481 )
482 .child(button_container)
483 } else {
484 v_flex()
485 .w_full()
486 .gap_2()
487 .child(
488 h_flex()
489 .flex_wrap()
490 .when(thread_fresh_start, |this| this.justify_center())
491 .child(Label::new(
492 "To start using Zed AI, please read and accept the",
493 ))
494 .child(terms_button),
495 )
496 .child({
497 match view_kind {
498 LanguageModelProviderTosView::PromptEditorPopup => {
499 button_container.w_full().justify_end()
500 }
501 LanguageModelProviderTosView::Configuration => {
502 button_container.w_full().justify_start()
503 }
504 LanguageModelProviderTosView::ThreadFreshStart => {
505 button_container.w_full().justify_center()
506 }
507 LanguageModelProviderTosView::ThreadtEmptyState => div().w_0(),
508 }
509 })
510 };
511
512 Some(form.into_any())
513}
514
515pub struct CloudLanguageModel {
516 id: LanguageModelId,
517 model: CloudModel,
518 llm_api_token: LlmApiToken,
519 client: Arc<Client>,
520 request_limiter: RateLimiter,
521}
522
523struct PerformLlmCompletionResponse {
524 response: Response<AsyncBody>,
525 usage: Option<RequestUsage>,
526 tool_use_limit_reached: bool,
527 includes_status_messages: bool,
528}
529
530impl CloudLanguageModel {
531 const MAX_RETRIES: usize = 3;
532
533 async fn perform_llm_completion(
534 client: Arc<Client>,
535 llm_api_token: LlmApiToken,
536 app_version: Option<SemanticVersion>,
537 body: CompletionBody,
538 ) -> Result<PerformLlmCompletionResponse> {
539 let http_client = &client.http_client();
540
541 let mut token = llm_api_token.acquire(&client).await?;
542 let mut retries_remaining = Self::MAX_RETRIES;
543 let mut retry_delay = Duration::from_secs(1);
544
545 loop {
546 let request_builder = http_client::Request::builder()
547 .method(Method::POST)
548 .uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref());
549 let request_builder = if let Some(app_version) = app_version {
550 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
551 } else {
552 request_builder
553 };
554
555 let request = request_builder
556 .header("Content-Type", "application/json")
557 .header("Authorization", format!("Bearer {token}"))
558 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
559 .body(serde_json::to_string(&body)?.into())?;
560 let mut response = http_client.send(request).await?;
561 let status = response.status();
562 if status.is_success() {
563 let includes_status_messages = response
564 .headers()
565 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
566 .is_some();
567
568 let tool_use_limit_reached = response
569 .headers()
570 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
571 .is_some();
572
573 let usage = if includes_status_messages {
574 None
575 } else {
576 RequestUsage::from_headers(response.headers()).ok()
577 };
578
579 return Ok(PerformLlmCompletionResponse {
580 response,
581 usage,
582 includes_status_messages,
583 tool_use_limit_reached,
584 });
585 } else if response
586 .headers()
587 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
588 .is_some()
589 {
590 retries_remaining -= 1;
591 token = llm_api_token.refresh(&client).await?;
592 } else if status == StatusCode::FORBIDDEN
593 && response
594 .headers()
595 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
596 .is_some()
597 {
598 return Err(anyhow!(MaxMonthlySpendReachedError));
599 } else if status == StatusCode::FORBIDDEN
600 && response
601 .headers()
602 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
603 .is_some()
604 {
605 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
606 .headers()
607 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
608 .and_then(|resource| resource.to_str().ok())
609 {
610 if let Some(plan) = response
611 .headers()
612 .get(CURRENT_PLAN_HEADER_NAME)
613 .and_then(|plan| plan.to_str().ok())
614 .and_then(|plan| zed_llm_client::Plan::from_str(plan).ok())
615 {
616 let plan = match plan {
617 zed_llm_client::Plan::ZedFree => Plan::Free,
618 zed_llm_client::Plan::ZedPro => Plan::ZedPro,
619 zed_llm_client::Plan::ZedProTrial => Plan::ZedProTrial,
620 };
621 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
622 }
623 }
624
625 return Err(anyhow!("Forbidden"));
626 } else if status.as_u16() >= 500 && status.as_u16() < 600 {
627 // If we encounter an error in the 500 range, retry after a delay.
628 // We've seen at least these in the wild from API providers:
629 // * 500 Internal Server Error
630 // * 502 Bad Gateway
631 // * 529 Service Overloaded
632
633 if retries_remaining == 0 {
634 let mut body = String::new();
635 response.body_mut().read_to_string(&mut body).await?;
636 return Err(anyhow!(
637 "cloud language model completion failed after {} retries with status {status}: {body}",
638 Self::MAX_RETRIES
639 ));
640 }
641
642 Timer::after(retry_delay).await;
643
644 retries_remaining -= 1;
645 retry_delay *= 2; // If it fails again, wait longer.
646 } else if status == StatusCode::PAYMENT_REQUIRED {
647 return Err(anyhow!(PaymentRequiredError));
648 } else {
649 let mut body = String::new();
650 response.body_mut().read_to_string(&mut body).await?;
651 return Err(anyhow!(ApiError { status, body }));
652 }
653 }
654 }
655}
656
657#[derive(Debug, Error)]
658#[error("cloud language model request failed with status {status}: {body}")]
659struct ApiError {
660 status: StatusCode,
661 body: String,
662}
663
664impl LanguageModel for CloudLanguageModel {
665 fn id(&self) -> LanguageModelId {
666 self.id.clone()
667 }
668
669 fn name(&self) -> LanguageModelName {
670 LanguageModelName::from(self.model.display_name().to_string())
671 }
672
673 fn provider_id(&self) -> LanguageModelProviderId {
674 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
675 }
676
677 fn provider_name(&self) -> LanguageModelProviderName {
678 LanguageModelProviderName(PROVIDER_NAME.into())
679 }
680
681 fn supports_tools(&self) -> bool {
682 match self.model {
683 CloudModel::Anthropic(_) => true,
684 CloudModel::Google(_) => true,
685 CloudModel::OpenAi(_) => true,
686 }
687 }
688
689 fn supports_images(&self) -> bool {
690 match self.model {
691 CloudModel::Anthropic(_) => true,
692 CloudModel::Google(_) => true,
693 CloudModel::OpenAi(_) => false,
694 }
695 }
696
697 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
698 match choice {
699 LanguageModelToolChoice::Auto
700 | LanguageModelToolChoice::Any
701 | LanguageModelToolChoice::None => true,
702 }
703 }
704
705 fn telemetry_id(&self) -> String {
706 format!("zed.dev/{}", self.model.id())
707 }
708
709 fn availability(&self) -> LanguageModelAvailability {
710 self.model.availability()
711 }
712
713 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
714 self.model.tool_input_format()
715 }
716
717 fn max_token_count(&self) -> usize {
718 self.model.max_token_count()
719 }
720
721 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
722 match &self.model {
723 CloudModel::Anthropic(model) => {
724 model
725 .cache_configuration()
726 .map(|cache| LanguageModelCacheConfiguration {
727 max_cache_anchors: cache.max_cache_anchors,
728 should_speculate: cache.should_speculate,
729 min_total_token: cache.min_total_token,
730 })
731 }
732 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
733 }
734 }
735
736 fn count_tokens(
737 &self,
738 request: LanguageModelRequest,
739 cx: &App,
740 ) -> BoxFuture<'static, Result<usize>> {
741 match self.model.clone() {
742 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
743 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
744 CloudModel::Google(model) => {
745 let client = self.client.clone();
746 let llm_api_token = self.llm_api_token.clone();
747 let model_id = model.id().to_string();
748 let generate_content_request = into_google(request, model_id.clone());
749 async move {
750 let http_client = &client.http_client();
751 let token = llm_api_token.acquire(&client).await?;
752
753 let request_body = CountTokensBody {
754 provider: zed_llm_client::LanguageModelProvider::Google,
755 model: model_id,
756 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
757 generate_content_request,
758 })?,
759 };
760 let request = http_client::Request::builder()
761 .method(Method::POST)
762 .uri(
763 http_client
764 .build_zed_llm_url("/count_tokens", &[])?
765 .as_ref(),
766 )
767 .header("Content-Type", "application/json")
768 .header("Authorization", format!("Bearer {token}"))
769 .body(serde_json::to_string(&request_body)?.into())?;
770 let mut response = http_client.send(request).await?;
771 let status = response.status();
772 let mut response_body = String::new();
773 response
774 .body_mut()
775 .read_to_string(&mut response_body)
776 .await?;
777
778 if status.is_success() {
779 let response_body: CountTokensResponse =
780 serde_json::from_str(&response_body)?;
781
782 Ok(response_body.tokens)
783 } else {
784 Err(anyhow!(ApiError {
785 status,
786 body: response_body
787 }))
788 }
789 }
790 .boxed()
791 }
792 }
793 }
794
795 fn stream_completion(
796 &self,
797 request: LanguageModelRequest,
798 cx: &AsyncApp,
799 ) -> BoxFuture<
800 'static,
801 Result<
802 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
803 >,
804 > {
805 let thread_id = request.thread_id.clone();
806 let prompt_id = request.prompt_id.clone();
807 let mode = request.mode;
808 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
809 match &self.model {
810 CloudModel::Anthropic(model) => {
811 let request = into_anthropic(
812 request,
813 model.request_id().into(),
814 model.default_temperature(),
815 model.max_output_tokens(),
816 model.mode(),
817 );
818 let client = self.client.clone();
819 let llm_api_token = self.llm_api_token.clone();
820 let future = self.request_limiter.stream(async move {
821 let PerformLlmCompletionResponse {
822 response,
823 usage,
824 includes_status_messages,
825 tool_use_limit_reached,
826 } = Self::perform_llm_completion(
827 client.clone(),
828 llm_api_token,
829 app_version,
830 CompletionBody {
831 thread_id,
832 prompt_id,
833 mode,
834 provider: zed_llm_client::LanguageModelProvider::Anthropic,
835 model: request.model.clone(),
836 provider_request: serde_json::to_value(&request)?,
837 },
838 )
839 .await
840 .map_err(|err| match err.downcast::<ApiError>() {
841 Ok(api_err) => {
842 if api_err.status == StatusCode::BAD_REQUEST {
843 if let Some(tokens) = parse_prompt_too_long(&api_err.body) {
844 return anyhow!(
845 LanguageModelKnownError::ContextWindowLimitExceeded {
846 tokens
847 }
848 );
849 }
850 }
851 anyhow!(api_err)
852 }
853 Err(err) => anyhow!(err),
854 })?;
855
856 let mut mapper = AnthropicEventMapper::new();
857 Ok(map_cloud_completion_events(
858 Box::pin(
859 response_lines(response, includes_status_messages)
860 .chain(usage_updated_event(usage))
861 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
862 ),
863 move |event| mapper.map_event(event),
864 ))
865 });
866 async move { Ok(future.await?.boxed()) }.boxed()
867 }
868 CloudModel::OpenAi(model) => {
869 let client = self.client.clone();
870 let request = into_open_ai(request, model, model.max_output_tokens());
871 let llm_api_token = self.llm_api_token.clone();
872 let future = self.request_limiter.stream(async move {
873 let PerformLlmCompletionResponse {
874 response,
875 usage,
876 includes_status_messages,
877 tool_use_limit_reached,
878 } = Self::perform_llm_completion(
879 client.clone(),
880 llm_api_token,
881 app_version,
882 CompletionBody {
883 thread_id,
884 prompt_id,
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 mode,
923 provider: zed_llm_client::LanguageModelProvider::Google,
924 model: request.model.model_id.clone(),
925 provider_request: serde_json::to_value(&request)?,
926 },
927 )
928 .await?;
929
930 let mut mapper = GoogleEventMapper::new();
931 Ok(map_cloud_completion_events(
932 Box::pin(
933 response_lines(response, includes_status_messages)
934 .chain(usage_updated_event(usage))
935 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
936 ),
937 move |event| mapper.map_event(event),
938 ))
939 });
940 async move { Ok(future.await?.boxed()) }.boxed()
941 }
942 }
943 }
944}
945
946#[derive(Serialize, Deserialize)]
947#[serde(rename_all = "snake_case")]
948pub enum CloudCompletionEvent<T> {
949 Status(CompletionRequestStatus),
950 Event(T),
951}
952
953fn map_cloud_completion_events<T, F>(
954 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
955 mut map_callback: F,
956) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
957where
958 T: DeserializeOwned + 'static,
959 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
960 + Send
961 + 'static,
962{
963 stream
964 .flat_map(move |event| {
965 futures::stream::iter(match event {
966 Err(error) => {
967 vec![Err(LanguageModelCompletionError::Other(error))]
968 }
969 Ok(CloudCompletionEvent::Status(event)) => {
970 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
971 }
972 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
973 })
974 })
975 .boxed()
976}
977
978fn usage_updated_event<T>(
979 usage: Option<RequestUsage>,
980) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
981 futures::stream::iter(usage.map(|usage| {
982 Ok(CloudCompletionEvent::Status(
983 CompletionRequestStatus::UsageUpdated {
984 amount: usage.amount as usize,
985 limit: usage.limit,
986 },
987 ))
988 }))
989}
990
991fn tool_use_limit_reached_event<T>(
992 tool_use_limit_reached: bool,
993) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
994 futures::stream::iter(tool_use_limit_reached.then(|| {
995 Ok(CloudCompletionEvent::Status(
996 CompletionRequestStatus::ToolUseLimitReached,
997 ))
998 }))
999}
1000
1001fn response_lines<T: DeserializeOwned>(
1002 response: Response<AsyncBody>,
1003 includes_status_messages: bool,
1004) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1005 futures::stream::try_unfold(
1006 (String::new(), BufReader::new(response.into_body())),
1007 move |(mut line, mut body)| async move {
1008 match body.read_line(&mut line).await {
1009 Ok(0) => Ok(None),
1010 Ok(_) => {
1011 let event = if includes_status_messages {
1012 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
1013 } else {
1014 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1015 };
1016
1017 line.clear();
1018 Ok(Some((event, (line, body))))
1019 }
1020 Err(e) => Err(e.into()),
1021 }
1022 },
1023 )
1024}
1025
1026struct ConfigurationView {
1027 state: gpui::Entity<State>,
1028}
1029
1030impl ConfigurationView {
1031 fn authenticate(&mut self, cx: &mut Context<Self>) {
1032 self.state.update(cx, |state, cx| {
1033 state.authenticate(cx).detach_and_log_err(cx);
1034 });
1035 cx.notify();
1036 }
1037}
1038
1039impl Render for ConfigurationView {
1040 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1041 const ZED_PRICING_URL: &str = "https://zed.dev/pricing";
1042
1043 let is_connected = !self.state.read(cx).is_signed_out();
1044 let user_store = self.state.read(cx).user_store.read(cx);
1045 let plan = user_store.current_plan();
1046 let subscription_period = user_store.subscription_period();
1047 let eligible_for_trial = user_store.trial_started_at().is_none();
1048 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
1049
1050 let is_pro = plan == Some(proto::Plan::ZedPro);
1051 let subscription_text = match (plan, subscription_period) {
1052 (Some(proto::Plan::ZedPro), Some(_)) => {
1053 "You have access to Zed's hosted LLMs through your Zed Pro subscription."
1054 }
1055 (Some(proto::Plan::ZedProTrial), Some(_)) => {
1056 "You have access to Zed's hosted LLMs through your Zed Pro trial."
1057 }
1058 (Some(proto::Plan::Free), Some(_)) => {
1059 "You have basic access to Zed's hosted LLMs through your Zed Free subscription."
1060 }
1061 _ => {
1062 if eligible_for_trial {
1063 "Subscribe for access to Zed's hosted LLMs. Start with a 14 day free trial."
1064 } else {
1065 "Subscribe for access to Zed's hosted LLMs."
1066 }
1067 }
1068 };
1069 let manage_subscription_buttons = if is_pro {
1070 h_flex().child(
1071 Button::new("manage_settings", "Manage Subscription")
1072 .style(ButtonStyle::Tinted(TintColor::Accent))
1073 .on_click(cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx)))),
1074 )
1075 } else {
1076 h_flex()
1077 .gap_2()
1078 .child(
1079 Button::new("learn_more", "Learn more")
1080 .style(ButtonStyle::Subtle)
1081 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_PRICING_URL))),
1082 )
1083 .child(
1084 Button::new("upgrade", "Upgrade")
1085 .style(ButtonStyle::Subtle)
1086 .color(Color::Accent)
1087 .on_click(
1088 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1089 ),
1090 )
1091 };
1092
1093 if is_connected {
1094 v_flex()
1095 .gap_3()
1096 .w_full()
1097 .children(render_accept_terms(
1098 self.state.clone(),
1099 LanguageModelProviderTosView::Configuration,
1100 cx,
1101 ))
1102 .when(has_accepted_terms, |this| {
1103 this.child(subscription_text)
1104 .child(manage_subscription_buttons)
1105 })
1106 } else {
1107 v_flex()
1108 .gap_2()
1109 .child(Label::new("Use Zed AI to access hosted language models."))
1110 .child(
1111 Button::new("sign_in", "Sign In")
1112 .icon_color(Color::Muted)
1113 .icon(IconName::Github)
1114 .icon_position(IconPosition::Start)
1115 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
1116 )
1117 }
1118 }
1119}