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, ZedProFeatureFlag};
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, LanguageModelToolSchemaFormat,
18 ModelRequestLimitReachedError, RateLimiter, RequestUsage, ZED_CLOUD_PROVIDER_ID,
19};
20use language_model::{
21 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider, LlmApiToken,
22 MaxMonthlySpendReachedError, PaymentRequiredError, RefreshLlmTokenListener,
23};
24use proto::Plan;
25use release_channel::AppVersion;
26use schemars::JsonSchema;
27use serde::{Deserialize, Serialize, de::DeserializeOwned};
28use settings::{Settings, SettingsStore};
29use smol::Timer;
30use smol::io::{AsyncReadExt, BufReader};
31use std::pin::Pin;
32use std::str::FromStr as _;
33use std::{
34 sync::{Arc, LazyLock},
35 time::Duration,
36};
37use strum::IntoEnumIterator;
38use thiserror::Error;
39use ui::{TintColor, prelude::*};
40use zed_llm_client::{
41 CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, CURRENT_PLAN_HEADER_NAME, CompletionBody,
42 CompletionRequestStatus, CountTokensBody, CountTokensResponse, EXPIRED_LLM_TOKEN_HEADER_NAME,
43 MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME, MODEL_REQUESTS_RESOURCE_HEADER_VALUE,
44 SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
45 TOOL_USE_LIMIT_REACHED_HEADER_NAME, ZED_VERSION_HEADER_NAME,
46};
47
48use crate::AllLanguageModelSettings;
49use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
50use crate::provider::google::{GoogleEventMapper, into_google};
51use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
52
53pub const PROVIDER_NAME: &str = "Zed";
54
55const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
56 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
57
58fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
59 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
60 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
61 .map(|json| serde_json::from_str(json).unwrap())
62 .unwrap_or_default()
63 });
64 ADDITIONAL_MODELS.as_slice()
65}
66
67#[derive(Default, Clone, Debug, PartialEq)]
68pub struct ZedDotDevSettings {
69 pub available_models: Vec<AvailableModel>,
70}
71
72#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
73#[serde(rename_all = "lowercase")]
74pub enum AvailableProvider {
75 Anthropic,
76 OpenAi,
77 Google,
78}
79
80#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
81pub struct AvailableModel {
82 /// The provider of the language model.
83 pub provider: AvailableProvider,
84 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
85 pub name: String,
86 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
87 pub display_name: Option<String>,
88 /// The size of the context window, indicating the maximum number of tokens the model can process.
89 pub max_tokens: usize,
90 /// The maximum number of output tokens allowed by the model.
91 pub max_output_tokens: Option<u32>,
92 /// The maximum number of completion tokens allowed by the model (o1-* only)
93 pub max_completion_tokens: Option<u32>,
94 /// Override this model with a different Anthropic model for tool calls.
95 pub tool_override: Option<String>,
96 /// Indicates whether this custom model supports caching.
97 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
98 /// The default temperature to use for this model.
99 pub default_temperature: Option<f32>,
100 /// Any extra beta headers to provide when using the model.
101 #[serde(default)]
102 pub extra_beta_headers: Vec<String>,
103 /// The model's mode (e.g. thinking)
104 pub mode: Option<ModelMode>,
105}
106
107#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
108#[serde(tag = "type", rename_all = "lowercase")]
109pub enum ModelMode {
110 #[default]
111 Default,
112 Thinking {
113 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
114 budget_tokens: Option<u32>,
115 },
116}
117
118impl From<ModelMode> for AnthropicModelMode {
119 fn from(value: ModelMode) -> Self {
120 match value {
121 ModelMode::Default => AnthropicModelMode::Default,
122 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
123 }
124 }
125}
126
127pub struct CloudLanguageModelProvider {
128 client: Arc<Client>,
129 state: gpui::Entity<State>,
130 _maintain_client_status: Task<()>,
131}
132
133pub struct State {
134 client: Arc<Client>,
135 llm_api_token: LlmApiToken,
136 user_store: Entity<UserStore>,
137 status: client::Status,
138 accept_terms: Option<Task<Result<()>>>,
139 _settings_subscription: Subscription,
140 _llm_token_subscription: Subscription,
141}
142
143impl State {
144 fn new(
145 client: Arc<Client>,
146 user_store: Entity<UserStore>,
147 status: client::Status,
148 cx: &mut Context<Self>,
149 ) -> Self {
150 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
151
152 Self {
153 client: client.clone(),
154 llm_api_token: LlmApiToken::default(),
155 user_store,
156 status,
157 accept_terms: None,
158 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
159 cx.notify();
160 }),
161 _llm_token_subscription: cx.subscribe(
162 &refresh_llm_token_listener,
163 |this, _listener, _event, cx| {
164 let client = this.client.clone();
165 let llm_api_token = this.llm_api_token.clone();
166 cx.spawn(async move |_this, _cx| {
167 llm_api_token.refresh(&client).await?;
168 anyhow::Ok(())
169 })
170 .detach_and_log_err(cx);
171 },
172 ),
173 }
174 }
175
176 fn is_signed_out(&self) -> bool {
177 self.status.is_signed_out()
178 }
179
180 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
181 let client = self.client.clone();
182 cx.spawn(async move |this, cx| {
183 client.authenticate_and_connect(true, &cx).await?;
184 this.update(cx, |_, cx| cx.notify())
185 })
186 }
187
188 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
189 self.user_store
190 .read(cx)
191 .current_user_has_accepted_terms()
192 .unwrap_or(false)
193 }
194
195 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
196 let user_store = self.user_store.clone();
197 self.accept_terms = Some(cx.spawn(async move |this, cx| {
198 let _ = user_store
199 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
200 .await;
201 this.update(cx, |this, cx| {
202 this.accept_terms = None;
203 cx.notify()
204 })
205 }));
206 }
207}
208
209impl CloudLanguageModelProvider {
210 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
211 let mut status_rx = client.status();
212 let status = *status_rx.borrow();
213
214 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
215
216 let state_ref = state.downgrade();
217 let maintain_client_status = cx.spawn(async move |cx| {
218 while let Some(status) = status_rx.next().await {
219 if let Some(this) = state_ref.upgrade() {
220 _ = this.update(cx, |this, cx| {
221 if this.status != status {
222 this.status = status;
223 cx.notify();
224 }
225 });
226 } else {
227 break;
228 }
229 }
230 });
231
232 Self {
233 client,
234 state: state.clone(),
235 _maintain_client_status: maintain_client_status,
236 }
237 }
238
239 fn create_language_model(
240 &self,
241 model: CloudModel,
242 llm_api_token: LlmApiToken,
243 ) -> Arc<dyn LanguageModel> {
244 Arc::new(CloudLanguageModel {
245 id: LanguageModelId::from(model.id().to_string()),
246 model,
247 llm_api_token: llm_api_token.clone(),
248 client: self.client.clone(),
249 request_limiter: RateLimiter::new(4),
250 })
251 }
252}
253
254impl LanguageModelProviderState for CloudLanguageModelProvider {
255 type ObservableEntity = State;
256
257 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
258 Some(self.state.clone())
259 }
260}
261
262impl LanguageModelProvider for CloudLanguageModelProvider {
263 fn id(&self) -> LanguageModelProviderId {
264 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
265 }
266
267 fn name(&self) -> LanguageModelProviderName {
268 LanguageModelProviderName(PROVIDER_NAME.into())
269 }
270
271 fn icon(&self) -> IconName {
272 IconName::AiZed
273 }
274
275 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
276 let llm_api_token = self.state.read(cx).llm_api_token.clone();
277 let model = CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet);
278 Some(self.create_language_model(model, llm_api_token))
279 }
280
281 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
282 let llm_api_token = self.state.read(cx).llm_api_token.clone();
283 let model = CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet);
284 Some(self.create_language_model(model, llm_api_token))
285 }
286
287 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
288 let llm_api_token = self.state.read(cx).llm_api_token.clone();
289 [
290 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
291 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
292 ]
293 .into_iter()
294 .map(|model| self.create_language_model(model, llm_api_token.clone()))
295 .collect()
296 }
297
298 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
299 let mut models = BTreeMap::default();
300
301 if cx.is_staff() {
302 for model in anthropic::Model::iter() {
303 if !matches!(model, anthropic::Model::Custom { .. }) {
304 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
305 }
306 }
307 for model in open_ai::Model::iter() {
308 if !matches!(model, open_ai::Model::Custom { .. }) {
309 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
310 }
311 }
312 for model in google_ai::Model::iter() {
313 if !matches!(model, google_ai::Model::Custom { .. }) {
314 models.insert(model.id().to_string(), CloudModel::Google(model));
315 }
316 }
317 } else {
318 models.insert(
319 anthropic::Model::Claude3_5Sonnet.id().to_string(),
320 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
321 );
322 models.insert(
323 anthropic::Model::Claude3_7Sonnet.id().to_string(),
324 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
325 );
326 models.insert(
327 anthropic::Model::Claude3_7SonnetThinking.id().to_string(),
328 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
329 );
330 }
331
332 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBetaFeatureFlag>() {
333 zed_cloud_provider_additional_models()
334 } else {
335 &[]
336 };
337
338 // Override with available models from settings
339 for model in AllLanguageModelSettings::get_global(cx)
340 .zed_dot_dev
341 .available_models
342 .iter()
343 .chain(llm_closed_beta_models)
344 .cloned()
345 {
346 let model = match model.provider {
347 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
348 name: model.name.clone(),
349 display_name: model.display_name.clone(),
350 max_tokens: model.max_tokens,
351 tool_override: model.tool_override.clone(),
352 cache_configuration: model.cache_configuration.as_ref().map(|config| {
353 anthropic::AnthropicModelCacheConfiguration {
354 max_cache_anchors: config.max_cache_anchors,
355 should_speculate: config.should_speculate,
356 min_total_token: config.min_total_token,
357 }
358 }),
359 default_temperature: model.default_temperature,
360 max_output_tokens: model.max_output_tokens,
361 extra_beta_headers: model.extra_beta_headers.clone(),
362 mode: model.mode.unwrap_or_default().into(),
363 }),
364 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
365 name: model.name.clone(),
366 display_name: model.display_name.clone(),
367 max_tokens: model.max_tokens,
368 max_output_tokens: model.max_output_tokens,
369 max_completion_tokens: model.max_completion_tokens,
370 }),
371 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
372 name: model.name.clone(),
373 display_name: model.display_name.clone(),
374 max_tokens: model.max_tokens,
375 }),
376 };
377 models.insert(model.id().to_string(), model.clone());
378 }
379
380 let llm_api_token = self.state.read(cx).llm_api_token.clone();
381 models
382 .into_values()
383 .map(|model| self.create_language_model(model, llm_api_token.clone()))
384 .collect()
385 }
386
387 fn is_authenticated(&self, cx: &App) -> bool {
388 !self.state.read(cx).is_signed_out()
389 }
390
391 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
392 Task::ready(Ok(()))
393 }
394
395 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
396 cx.new(|_| ConfigurationView {
397 state: self.state.clone(),
398 })
399 .into()
400 }
401
402 fn must_accept_terms(&self, cx: &App) -> bool {
403 !self.state.read(cx).has_accepted_terms_of_service(cx)
404 }
405
406 fn render_accept_terms(
407 &self,
408 view: LanguageModelProviderTosView,
409 cx: &mut App,
410 ) -> Option<AnyElement> {
411 render_accept_terms(self.state.clone(), view, cx)
412 }
413
414 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
415 Task::ready(Ok(()))
416 }
417}
418
419fn render_accept_terms(
420 state: Entity<State>,
421 view_kind: LanguageModelProviderTosView,
422 cx: &mut App,
423) -> Option<AnyElement> {
424 if state.read(cx).has_accepted_terms_of_service(cx) {
425 return None;
426 }
427
428 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
429
430 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
431 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadtEmptyState);
432
433 let terms_button = Button::new("terms_of_service", "Terms of Service")
434 .style(ButtonStyle::Subtle)
435 .icon(IconName::ArrowUpRight)
436 .icon_color(Color::Muted)
437 .icon_size(IconSize::XSmall)
438 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
439 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
440
441 let button_container = h_flex().child(
442 Button::new("accept_terms", "I accept the Terms of Service")
443 .when(!thread_empty_state, |this| {
444 this.full_width()
445 .style(ButtonStyle::Tinted(TintColor::Accent))
446 .icon(IconName::Check)
447 .icon_position(IconPosition::Start)
448 .icon_size(IconSize::Small)
449 })
450 .when(thread_empty_state, |this| {
451 this.style(ButtonStyle::Tinted(TintColor::Warning))
452 .label_size(LabelSize::Small)
453 })
454 .disabled(accept_terms_disabled)
455 .on_click({
456 let state = state.downgrade();
457 move |_, _window, cx| {
458 state
459 .update(cx, |state, cx| state.accept_terms_of_service(cx))
460 .ok();
461 }
462 }),
463 );
464
465 let form = if thread_empty_state {
466 h_flex()
467 .w_full()
468 .flex_wrap()
469 .justify_between()
470 .child(
471 h_flex()
472 .child(
473 Label::new("To start using Zed AI, please read and accept the")
474 .size(LabelSize::Small),
475 )
476 .child(terms_button),
477 )
478 .child(button_container)
479 } else {
480 v_flex()
481 .w_full()
482 .gap_2()
483 .child(
484 h_flex()
485 .flex_wrap()
486 .when(thread_fresh_start, |this| this.justify_center())
487 .child(Label::new(
488 "To start using Zed AI, please read and accept the",
489 ))
490 .child(terms_button),
491 )
492 .child({
493 match view_kind {
494 LanguageModelProviderTosView::PromptEditorPopup => {
495 button_container.w_full().justify_end()
496 }
497 LanguageModelProviderTosView::Configuration => {
498 button_container.w_full().justify_start()
499 }
500 LanguageModelProviderTosView::ThreadFreshStart => {
501 button_container.w_full().justify_center()
502 }
503 LanguageModelProviderTosView::ThreadtEmptyState => div().w_0(),
504 }
505 })
506 };
507
508 Some(form.into_any())
509}
510
511pub struct CloudLanguageModel {
512 id: LanguageModelId,
513 model: CloudModel,
514 llm_api_token: LlmApiToken,
515 client: Arc<Client>,
516 request_limiter: RateLimiter,
517}
518
519struct PerformLlmCompletionResponse {
520 response: Response<AsyncBody>,
521 usage: Option<RequestUsage>,
522 tool_use_limit_reached: bool,
523 includes_status_messages: bool,
524}
525
526impl CloudLanguageModel {
527 const MAX_RETRIES: usize = 3;
528
529 async fn perform_llm_completion(
530 client: Arc<Client>,
531 llm_api_token: LlmApiToken,
532 app_version: Option<SemanticVersion>,
533 body: CompletionBody,
534 ) -> Result<PerformLlmCompletionResponse> {
535 let http_client = &client.http_client();
536
537 let mut token = llm_api_token.acquire(&client).await?;
538 let mut retries_remaining = Self::MAX_RETRIES;
539 let mut retry_delay = Duration::from_secs(1);
540
541 loop {
542 let request_builder = http_client::Request::builder().method(Method::POST);
543 let request_builder = if let Ok(completions_url) = std::env::var("ZED_COMPLETIONS_URL")
544 {
545 request_builder.uri(completions_url)
546 } else {
547 request_builder.uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
548 };
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::Free => 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 telemetry_id(&self) -> String {
690 format!("zed.dev/{}", self.model.id())
691 }
692
693 fn availability(&self) -> LanguageModelAvailability {
694 self.model.availability()
695 }
696
697 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
698 self.model.tool_input_format()
699 }
700
701 fn max_token_count(&self) -> usize {
702 self.model.max_token_count()
703 }
704
705 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
706 match &self.model {
707 CloudModel::Anthropic(model) => {
708 model
709 .cache_configuration()
710 .map(|cache| LanguageModelCacheConfiguration {
711 max_cache_anchors: cache.max_cache_anchors,
712 should_speculate: cache.should_speculate,
713 min_total_token: cache.min_total_token,
714 })
715 }
716 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
717 }
718 }
719
720 fn count_tokens(
721 &self,
722 request: LanguageModelRequest,
723 cx: &App,
724 ) -> BoxFuture<'static, Result<usize>> {
725 match self.model.clone() {
726 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
727 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
728 CloudModel::Google(model) => {
729 let client = self.client.clone();
730 let llm_api_token = self.llm_api_token.clone();
731 let model_id = model.id().to_string();
732 let generate_content_request = into_google(request, model_id.clone());
733 async move {
734 let http_client = &client.http_client();
735 let token = llm_api_token.acquire(&client).await?;
736
737 let request_builder = http_client::Request::builder().method(Method::POST);
738 let request_builder =
739 if let Ok(completions_url) = std::env::var("ZED_COUNT_TOKENS_URL") {
740 request_builder.uri(completions_url)
741 } else {
742 request_builder.uri(
743 http_client
744 .build_zed_llm_url("/count_tokens", &[])?
745 .as_ref(),
746 )
747 };
748 let request_body = CountTokensBody {
749 provider: zed_llm_client::LanguageModelProvider::Google,
750 model: model_id,
751 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
752 generate_content_request,
753 })?,
754 };
755 let request = request_builder
756 .header("Content-Type", "application/json")
757 .header("Authorization", format!("Bearer {token}"))
758 .body(serde_json::to_string(&request_body)?.into())?;
759 let mut response = http_client.send(request).await?;
760 let status = response.status();
761 let mut response_body = String::new();
762 response
763 .body_mut()
764 .read_to_string(&mut response_body)
765 .await?;
766
767 if status.is_success() {
768 let response_body: CountTokensResponse =
769 serde_json::from_str(&response_body)?;
770
771 Ok(response_body.tokens)
772 } else {
773 Err(anyhow!(ApiError {
774 status,
775 body: response_body
776 }))
777 }
778 }
779 .boxed()
780 }
781 }
782 }
783
784 fn stream_completion(
785 &self,
786 request: LanguageModelRequest,
787 cx: &AsyncApp,
788 ) -> BoxFuture<
789 'static,
790 Result<
791 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
792 >,
793 > {
794 let thread_id = request.thread_id.clone();
795 let prompt_id = request.prompt_id.clone();
796 let mode = request.mode;
797 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
798 match &self.model {
799 CloudModel::Anthropic(model) => {
800 let request = into_anthropic(
801 request,
802 model.request_id().into(),
803 model.default_temperature(),
804 model.max_output_tokens(),
805 model.mode(),
806 );
807 let client = self.client.clone();
808 let llm_api_token = self.llm_api_token.clone();
809 let future = self.request_limiter.stream(async move {
810 let PerformLlmCompletionResponse {
811 response,
812 usage,
813 includes_status_messages,
814 tool_use_limit_reached,
815 } = Self::perform_llm_completion(
816 client.clone(),
817 llm_api_token,
818 app_version,
819 CompletionBody {
820 thread_id,
821 prompt_id,
822 mode,
823 provider: zed_llm_client::LanguageModelProvider::Anthropic,
824 model: request.model.clone(),
825 provider_request: serde_json::to_value(&request)?,
826 },
827 )
828 .await
829 .map_err(|err| match err.downcast::<ApiError>() {
830 Ok(api_err) => {
831 if api_err.status == StatusCode::BAD_REQUEST {
832 if let Some(tokens) = parse_prompt_too_long(&api_err.body) {
833 return anyhow!(
834 LanguageModelKnownError::ContextWindowLimitExceeded {
835 tokens
836 }
837 );
838 }
839 }
840 anyhow!(api_err)
841 }
842 Err(err) => anyhow!(err),
843 })?;
844
845 let mut mapper = AnthropicEventMapper::new();
846 Ok(map_cloud_completion_events(
847 Box::pin(
848 response_lines(response, includes_status_messages)
849 .chain(usage_updated_event(usage))
850 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
851 ),
852 move |event| mapper.map_event(event),
853 ))
854 });
855 async move { Ok(future.await?.boxed()) }.boxed()
856 }
857 CloudModel::OpenAi(model) => {
858 let client = self.client.clone();
859 let request = into_open_ai(request, model, model.max_output_tokens());
860 let llm_api_token = self.llm_api_token.clone();
861 let future = self.request_limiter.stream(async move {
862 let PerformLlmCompletionResponse {
863 response,
864 usage,
865 includes_status_messages,
866 tool_use_limit_reached,
867 } = Self::perform_llm_completion(
868 client.clone(),
869 llm_api_token,
870 app_version,
871 CompletionBody {
872 thread_id,
873 prompt_id,
874 mode,
875 provider: zed_llm_client::LanguageModelProvider::OpenAi,
876 model: request.model.clone(),
877 provider_request: serde_json::to_value(&request)?,
878 },
879 )
880 .await?;
881
882 let mut mapper = OpenAiEventMapper::new();
883 Ok(map_cloud_completion_events(
884 Box::pin(
885 response_lines(response, includes_status_messages)
886 .chain(usage_updated_event(usage))
887 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
888 ),
889 move |event| mapper.map_event(event),
890 ))
891 });
892 async move { Ok(future.await?.boxed()) }.boxed()
893 }
894 CloudModel::Google(model) => {
895 let client = self.client.clone();
896 let request = into_google(request, model.id().into());
897 let llm_api_token = self.llm_api_token.clone();
898 let future = self.request_limiter.stream(async move {
899 let PerformLlmCompletionResponse {
900 response,
901 usage,
902 includes_status_messages,
903 tool_use_limit_reached,
904 } = Self::perform_llm_completion(
905 client.clone(),
906 llm_api_token,
907 app_version,
908 CompletionBody {
909 thread_id,
910 prompt_id,
911 mode,
912 provider: zed_llm_client::LanguageModelProvider::Google,
913 model: request.model.model_id.clone(),
914 provider_request: serde_json::to_value(&request)?,
915 },
916 )
917 .await?;
918
919 let mut mapper = GoogleEventMapper::new();
920 Ok(map_cloud_completion_events(
921 Box::pin(
922 response_lines(response, includes_status_messages)
923 .chain(usage_updated_event(usage))
924 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
925 ),
926 move |event| mapper.map_event(event),
927 ))
928 });
929 async move { Ok(future.await?.boxed()) }.boxed()
930 }
931 }
932 }
933}
934
935#[derive(Serialize, Deserialize)]
936#[serde(rename_all = "snake_case")]
937pub enum CloudCompletionEvent<T> {
938 Status(CompletionRequestStatus),
939 Event(T),
940}
941
942fn map_cloud_completion_events<T, F>(
943 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
944 mut map_callback: F,
945) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
946where
947 T: DeserializeOwned + 'static,
948 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
949 + Send
950 + 'static,
951{
952 stream
953 .flat_map(move |event| {
954 futures::stream::iter(match event {
955 Err(error) => {
956 vec![Err(LanguageModelCompletionError::Other(error))]
957 }
958 Ok(CloudCompletionEvent::Status(event)) => {
959 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
960 }
961 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
962 })
963 })
964 .boxed()
965}
966
967fn usage_updated_event<T>(
968 usage: Option<RequestUsage>,
969) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
970 futures::stream::iter(usage.map(|usage| {
971 Ok(CloudCompletionEvent::Status(
972 CompletionRequestStatus::UsageUpdated {
973 amount: usage.amount as usize,
974 limit: usage.limit,
975 },
976 ))
977 }))
978}
979
980fn tool_use_limit_reached_event<T>(
981 tool_use_limit_reached: bool,
982) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
983 futures::stream::iter(tool_use_limit_reached.then(|| {
984 Ok(CloudCompletionEvent::Status(
985 CompletionRequestStatus::ToolUseLimitReached,
986 ))
987 }))
988}
989
990fn response_lines<T: DeserializeOwned>(
991 response: Response<AsyncBody>,
992 includes_status_messages: bool,
993) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
994 futures::stream::try_unfold(
995 (String::new(), BufReader::new(response.into_body())),
996 move |(mut line, mut body)| async move {
997 match body.read_line(&mut line).await {
998 Ok(0) => Ok(None),
999 Ok(_) => {
1000 let event = if includes_status_messages {
1001 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
1002 } else {
1003 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1004 };
1005
1006 line.clear();
1007 Ok(Some((event, (line, body))))
1008 }
1009 Err(e) => Err(e.into()),
1010 }
1011 },
1012 )
1013}
1014
1015struct ConfigurationView {
1016 state: gpui::Entity<State>,
1017}
1018
1019impl ConfigurationView {
1020 fn authenticate(&mut self, cx: &mut Context<Self>) {
1021 self.state.update(cx, |state, cx| {
1022 state.authenticate(cx).detach_and_log_err(cx);
1023 });
1024 cx.notify();
1025 }
1026}
1027
1028impl Render for ConfigurationView {
1029 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1030 const ZED_AI_URL: &str = "https://zed.dev/ai";
1031
1032 let is_connected = !self.state.read(cx).is_signed_out();
1033 let plan = self.state.read(cx).user_store.read(cx).current_plan();
1034 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
1035
1036 let is_pro = plan == Some(proto::Plan::ZedPro);
1037 let subscription_text = Label::new(if is_pro {
1038 "You have full access to Zed's hosted LLMs, which include models from Anthropic, OpenAI, and Google. They come with faster speeds and higher limits through Zed Pro."
1039 } else {
1040 "You have basic access to models from Anthropic through the Zed AI Free plan."
1041 });
1042 let manage_subscription_button = if is_pro {
1043 Some(
1044 h_flex().child(
1045 Button::new("manage_settings", "Manage Subscription")
1046 .style(ButtonStyle::Tinted(TintColor::Accent))
1047 .on_click(
1048 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1049 ),
1050 ),
1051 )
1052 } else if cx.has_flag::<ZedProFeatureFlag>() {
1053 Some(
1054 h_flex()
1055 .gap_2()
1056 .child(
1057 Button::new("learn_more", "Learn more")
1058 .style(ButtonStyle::Subtle)
1059 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
1060 )
1061 .child(
1062 Button::new("upgrade", "Upgrade")
1063 .style(ButtonStyle::Subtle)
1064 .color(Color::Accent)
1065 .on_click(
1066 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1067 ),
1068 ),
1069 )
1070 } else {
1071 None
1072 };
1073
1074 if is_connected {
1075 v_flex()
1076 .gap_3()
1077 .w_full()
1078 .children(render_accept_terms(
1079 self.state.clone(),
1080 LanguageModelProviderTosView::Configuration,
1081 cx,
1082 ))
1083 .when(has_accepted_terms, |this| {
1084 this.child(subscription_text)
1085 .children(manage_subscription_button)
1086 })
1087 } else {
1088 v_flex()
1089 .gap_2()
1090 .child(Label::new("Use Zed AI to access hosted language models."))
1091 .child(
1092 Button::new("sign_in", "Sign In")
1093 .icon_color(Color::Muted)
1094 .icon(IconName::Github)
1095 .icon_position(IconPosition::Start)
1096 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
1097 )
1098 }
1099 }
1100}