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 |this, cx| {
184 client.authenticate_and_connect(true, &cx).await?;
185 this.update(cx, |_, cx| cx.notify())
186 })
187 }
188
189 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
190 self.user_store
191 .read(cx)
192 .current_user_has_accepted_terms()
193 .unwrap_or(false)
194 }
195
196 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
197 let user_store = self.user_store.clone();
198 self.accept_terms = Some(cx.spawn(async move |this, cx| {
199 let _ = user_store
200 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
201 .await;
202 this.update(cx, |this, cx| {
203 this.accept_terms = None;
204 cx.notify()
205 })
206 }));
207 }
208}
209
210impl CloudLanguageModelProvider {
211 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
212 let mut status_rx = client.status();
213 let status = *status_rx.borrow();
214
215 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
216
217 let state_ref = state.downgrade();
218 let maintain_client_status = cx.spawn(async move |cx| {
219 while let Some(status) = status_rx.next().await {
220 if let Some(this) = state_ref.upgrade() {
221 _ = this.update(cx, |this, cx| {
222 if this.status != status {
223 this.status = status;
224 cx.notify();
225 }
226 });
227 } else {
228 break;
229 }
230 }
231 });
232
233 Self {
234 client,
235 state: state.clone(),
236 _maintain_client_status: maintain_client_status,
237 }
238 }
239
240 fn create_language_model(
241 &self,
242 model: CloudModel,
243 llm_api_token: LlmApiToken,
244 ) -> Arc<dyn LanguageModel> {
245 Arc::new(CloudLanguageModel {
246 id: LanguageModelId::from(model.id().to_string()),
247 model,
248 llm_api_token: llm_api_token.clone(),
249 client: self.client.clone(),
250 request_limiter: RateLimiter::new(4),
251 })
252 }
253}
254
255impl LanguageModelProviderState for CloudLanguageModelProvider {
256 type ObservableEntity = State;
257
258 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
259 Some(self.state.clone())
260 }
261}
262
263impl LanguageModelProvider for CloudLanguageModelProvider {
264 fn id(&self) -> LanguageModelProviderId {
265 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
266 }
267
268 fn name(&self) -> LanguageModelProviderName {
269 LanguageModelProviderName(PROVIDER_NAME.into())
270 }
271
272 fn icon(&self) -> IconName {
273 IconName::AiZed
274 }
275
276 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
277 let llm_api_token = self.state.read(cx).llm_api_token.clone();
278 let model = CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet);
279 Some(self.create_language_model(model, llm_api_token))
280 }
281
282 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
283 let llm_api_token = self.state.read(cx).llm_api_token.clone();
284 let model = CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet);
285 Some(self.create_language_model(model, llm_api_token))
286 }
287
288 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
289 let llm_api_token = self.state.read(cx).llm_api_token.clone();
290 [
291 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
292 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
293 ]
294 .into_iter()
295 .map(|model| self.create_language_model(model, llm_api_token.clone()))
296 .collect()
297 }
298
299 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
300 let mut models = BTreeMap::default();
301
302 if cx.is_staff() {
303 for model in anthropic::Model::iter() {
304 if !matches!(model, anthropic::Model::Custom { .. }) {
305 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
306 }
307 }
308 for model in open_ai::Model::iter() {
309 if !matches!(model, open_ai::Model::Custom { .. }) {
310 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
311 }
312 }
313 for model in google_ai::Model::iter() {
314 if !matches!(model, google_ai::Model::Custom { .. }) {
315 models.insert(model.id().to_string(), CloudModel::Google(model));
316 }
317 }
318 } else {
319 models.insert(
320 anthropic::Model::Claude3_5Sonnet.id().to_string(),
321 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
322 );
323 models.insert(
324 anthropic::Model::Claude3_7Sonnet.id().to_string(),
325 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
326 );
327 models.insert(
328 anthropic::Model::Claude3_7SonnetThinking.id().to_string(),
329 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
330 );
331 }
332
333 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBetaFeatureFlag>() {
334 zed_cloud_provider_additional_models()
335 } else {
336 &[]
337 };
338
339 // Override with available models from settings
340 for model in AllLanguageModelSettings::get_global(cx)
341 .zed_dot_dev
342 .available_models
343 .iter()
344 .chain(llm_closed_beta_models)
345 .cloned()
346 {
347 let model = match model.provider {
348 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
349 name: model.name.clone(),
350 display_name: model.display_name.clone(),
351 max_tokens: model.max_tokens,
352 tool_override: model.tool_override.clone(),
353 cache_configuration: model.cache_configuration.as_ref().map(|config| {
354 anthropic::AnthropicModelCacheConfiguration {
355 max_cache_anchors: config.max_cache_anchors,
356 should_speculate: config.should_speculate,
357 min_total_token: config.min_total_token,
358 }
359 }),
360 default_temperature: model.default_temperature,
361 max_output_tokens: model.max_output_tokens,
362 extra_beta_headers: model.extra_beta_headers.clone(),
363 mode: model.mode.unwrap_or_default().into(),
364 }),
365 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
366 name: model.name.clone(),
367 display_name: model.display_name.clone(),
368 max_tokens: model.max_tokens,
369 max_output_tokens: model.max_output_tokens,
370 max_completion_tokens: model.max_completion_tokens,
371 }),
372 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
373 name: model.name.clone(),
374 display_name: model.display_name.clone(),
375 max_tokens: model.max_tokens,
376 }),
377 };
378 models.insert(model.id().to_string(), model.clone());
379 }
380
381 let llm_api_token = self.state.read(cx).llm_api_token.clone();
382 models
383 .into_values()
384 .map(|model| self.create_language_model(model, llm_api_token.clone()))
385 .collect()
386 }
387
388 fn is_authenticated(&self, cx: &App) -> bool {
389 !self.state.read(cx).is_signed_out()
390 }
391
392 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
393 Task::ready(Ok(()))
394 }
395
396 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
397 cx.new(|_| ConfigurationView {
398 state: self.state.clone(),
399 })
400 .into()
401 }
402
403 fn must_accept_terms(&self, cx: &App) -> bool {
404 !self.state.read(cx).has_accepted_terms_of_service(cx)
405 }
406
407 fn render_accept_terms(
408 &self,
409 view: LanguageModelProviderTosView,
410 cx: &mut App,
411 ) -> Option<AnyElement> {
412 render_accept_terms(self.state.clone(), view, cx)
413 }
414
415 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
416 Task::ready(Ok(()))
417 }
418}
419
420fn render_accept_terms(
421 state: Entity<State>,
422 view_kind: LanguageModelProviderTosView,
423 cx: &mut App,
424) -> Option<AnyElement> {
425 if state.read(cx).has_accepted_terms_of_service(cx) {
426 return None;
427 }
428
429 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
430
431 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
432 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadtEmptyState);
433
434 let terms_button = Button::new("terms_of_service", "Terms of Service")
435 .style(ButtonStyle::Subtle)
436 .icon(IconName::ArrowUpRight)
437 .icon_color(Color::Muted)
438 .icon_size(IconSize::XSmall)
439 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
440 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
441
442 let button_container = h_flex().child(
443 Button::new("accept_terms", "I accept the Terms of Service")
444 .when(!thread_empty_state, |this| {
445 this.full_width()
446 .style(ButtonStyle::Tinted(TintColor::Accent))
447 .icon(IconName::Check)
448 .icon_position(IconPosition::Start)
449 .icon_size(IconSize::Small)
450 })
451 .when(thread_empty_state, |this| {
452 this.style(ButtonStyle::Tinted(TintColor::Warning))
453 .label_size(LabelSize::Small)
454 })
455 .disabled(accept_terms_disabled)
456 .on_click({
457 let state = state.downgrade();
458 move |_, _window, cx| {
459 state
460 .update(cx, |state, cx| state.accept_terms_of_service(cx))
461 .ok();
462 }
463 }),
464 );
465
466 let form = if thread_empty_state {
467 h_flex()
468 .w_full()
469 .flex_wrap()
470 .justify_between()
471 .child(
472 h_flex()
473 .child(
474 Label::new("To start using Zed AI, please read and accept the")
475 .size(LabelSize::Small),
476 )
477 .child(terms_button),
478 )
479 .child(button_container)
480 } else {
481 v_flex()
482 .w_full()
483 .gap_2()
484 .child(
485 h_flex()
486 .flex_wrap()
487 .when(thread_fresh_start, |this| this.justify_center())
488 .child(Label::new(
489 "To start using Zed AI, please read and accept the",
490 ))
491 .child(terms_button),
492 )
493 .child({
494 match view_kind {
495 LanguageModelProviderTosView::PromptEditorPopup => {
496 button_container.w_full().justify_end()
497 }
498 LanguageModelProviderTosView::Configuration => {
499 button_container.w_full().justify_start()
500 }
501 LanguageModelProviderTosView::ThreadFreshStart => {
502 button_container.w_full().justify_center()
503 }
504 LanguageModelProviderTosView::ThreadtEmptyState => div().w_0(),
505 }
506 })
507 };
508
509 Some(form.into_any())
510}
511
512pub struct CloudLanguageModel {
513 id: LanguageModelId,
514 model: CloudModel,
515 llm_api_token: LlmApiToken,
516 client: Arc<Client>,
517 request_limiter: RateLimiter,
518}
519
520struct PerformLlmCompletionResponse {
521 response: Response<AsyncBody>,
522 usage: Option<RequestUsage>,
523 tool_use_limit_reached: bool,
524 includes_status_messages: bool,
525}
526
527impl CloudLanguageModel {
528 const MAX_RETRIES: usize = 3;
529
530 async fn perform_llm_completion(
531 client: Arc<Client>,
532 llm_api_token: LlmApiToken,
533 app_version: Option<SemanticVersion>,
534 body: CompletionBody,
535 ) -> Result<PerformLlmCompletionResponse> {
536 let http_client = &client.http_client();
537
538 let mut token = llm_api_token.acquire(&client).await?;
539 let mut retries_remaining = Self::MAX_RETRIES;
540 let mut retry_delay = Duration::from_secs(1);
541
542 loop {
543 let request_builder = http_client::Request::builder().method(Method::POST);
544 let request_builder = if let Ok(completions_url) = std::env::var("ZED_COMPLETIONS_URL")
545 {
546 request_builder.uri(completions_url)
547 } else {
548 request_builder.uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
549 };
550 let request_builder = if let Some(app_version) = app_version {
551 request_builder.header(ZED_VERSION_HEADER_NAME, app_version.to_string())
552 } else {
553 request_builder
554 };
555
556 let request = request_builder
557 .header("Content-Type", "application/json")
558 .header("Authorization", format!("Bearer {token}"))
559 .header(CLIENT_SUPPORTS_STATUS_MESSAGES_HEADER_NAME, "true")
560 .body(serde_json::to_string(&body)?.into())?;
561 let mut response = http_client.send(request).await?;
562 let status = response.status();
563 if status.is_success() {
564 let includes_status_messages = response
565 .headers()
566 .get(SERVER_SUPPORTS_STATUS_MESSAGES_HEADER_NAME)
567 .is_some();
568
569 let tool_use_limit_reached = response
570 .headers()
571 .get(TOOL_USE_LIMIT_REACHED_HEADER_NAME)
572 .is_some();
573
574 let usage = if includes_status_messages {
575 None
576 } else {
577 RequestUsage::from_headers(response.headers()).ok()
578 };
579
580 return Ok(PerformLlmCompletionResponse {
581 response,
582 usage,
583 includes_status_messages,
584 tool_use_limit_reached,
585 });
586 } else if response
587 .headers()
588 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
589 .is_some()
590 {
591 retries_remaining -= 1;
592 token = llm_api_token.refresh(&client).await?;
593 } else if status == StatusCode::FORBIDDEN
594 && response
595 .headers()
596 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
597 .is_some()
598 {
599 return Err(anyhow!(MaxMonthlySpendReachedError));
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::Free => 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 return Err(anyhow!("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 return Err(anyhow!(
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 availability(&self) -> LanguageModelAvailability {
711 self.model.availability()
712 }
713
714 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
715 self.model.tool_input_format()
716 }
717
718 fn max_token_count(&self) -> usize {
719 self.model.max_token_count()
720 }
721
722 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
723 match &self.model {
724 CloudModel::Anthropic(model) => {
725 model
726 .cache_configuration()
727 .map(|cache| LanguageModelCacheConfiguration {
728 max_cache_anchors: cache.max_cache_anchors,
729 should_speculate: cache.should_speculate,
730 min_total_token: cache.min_total_token,
731 })
732 }
733 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
734 }
735 }
736
737 fn count_tokens(
738 &self,
739 request: LanguageModelRequest,
740 cx: &App,
741 ) -> BoxFuture<'static, Result<usize>> {
742 match self.model.clone() {
743 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
744 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
745 CloudModel::Google(model) => {
746 let client = self.client.clone();
747 let llm_api_token = self.llm_api_token.clone();
748 let model_id = model.id().to_string();
749 let generate_content_request = into_google(request, model_id.clone());
750 async move {
751 let http_client = &client.http_client();
752 let token = llm_api_token.acquire(&client).await?;
753
754 let request_builder = http_client::Request::builder().method(Method::POST);
755 let request_builder =
756 if let Ok(completions_url) = std::env::var("ZED_COUNT_TOKENS_URL") {
757 request_builder.uri(completions_url)
758 } else {
759 request_builder.uri(
760 http_client
761 .build_zed_llm_url("/count_tokens", &[])?
762 .as_ref(),
763 )
764 };
765 let request_body = CountTokensBody {
766 provider: zed_llm_client::LanguageModelProvider::Google,
767 model: model_id,
768 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
769 generate_content_request,
770 })?,
771 };
772 let request = request_builder
773 .header("Content-Type", "application/json")
774 .header("Authorization", format!("Bearer {token}"))
775 .body(serde_json::to_string(&request_body)?.into())?;
776 let mut response = http_client.send(request).await?;
777 let status = response.status();
778 let mut response_body = String::new();
779 response
780 .body_mut()
781 .read_to_string(&mut response_body)
782 .await?;
783
784 if status.is_success() {
785 let response_body: CountTokensResponse =
786 serde_json::from_str(&response_body)?;
787
788 Ok(response_body.tokens)
789 } else {
790 Err(anyhow!(ApiError {
791 status,
792 body: response_body
793 }))
794 }
795 }
796 .boxed()
797 }
798 }
799 }
800
801 fn stream_completion(
802 &self,
803 request: LanguageModelRequest,
804 cx: &AsyncApp,
805 ) -> BoxFuture<
806 'static,
807 Result<
808 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
809 >,
810 > {
811 let thread_id = request.thread_id.clone();
812 let prompt_id = request.prompt_id.clone();
813 let mode = request.mode;
814 let app_version = cx.update(|cx| AppVersion::global(cx)).ok();
815 match &self.model {
816 CloudModel::Anthropic(model) => {
817 let request = into_anthropic(
818 request,
819 model.request_id().into(),
820 model.default_temperature(),
821 model.max_output_tokens(),
822 model.mode(),
823 );
824 let client = self.client.clone();
825 let llm_api_token = self.llm_api_token.clone();
826 let future = self.request_limiter.stream(async move {
827 let PerformLlmCompletionResponse {
828 response,
829 usage,
830 includes_status_messages,
831 tool_use_limit_reached,
832 } = Self::perform_llm_completion(
833 client.clone(),
834 llm_api_token,
835 app_version,
836 CompletionBody {
837 thread_id,
838 prompt_id,
839 mode,
840 provider: zed_llm_client::LanguageModelProvider::Anthropic,
841 model: request.model.clone(),
842 provider_request: serde_json::to_value(&request)?,
843 },
844 )
845 .await
846 .map_err(|err| match err.downcast::<ApiError>() {
847 Ok(api_err) => {
848 if api_err.status == StatusCode::BAD_REQUEST {
849 if let Some(tokens) = parse_prompt_too_long(&api_err.body) {
850 return anyhow!(
851 LanguageModelKnownError::ContextWindowLimitExceeded {
852 tokens
853 }
854 );
855 }
856 }
857 anyhow!(api_err)
858 }
859 Err(err) => anyhow!(err),
860 })?;
861
862 let mut mapper = AnthropicEventMapper::new();
863 Ok(map_cloud_completion_events(
864 Box::pin(
865 response_lines(response, includes_status_messages)
866 .chain(usage_updated_event(usage))
867 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
868 ),
869 move |event| mapper.map_event(event),
870 ))
871 });
872 async move { Ok(future.await?.boxed()) }.boxed()
873 }
874 CloudModel::OpenAi(model) => {
875 let client = self.client.clone();
876 let request = into_open_ai(request, model, model.max_output_tokens());
877 let llm_api_token = self.llm_api_token.clone();
878 let future = self.request_limiter.stream(async move {
879 let PerformLlmCompletionResponse {
880 response,
881 usage,
882 includes_status_messages,
883 tool_use_limit_reached,
884 } = Self::perform_llm_completion(
885 client.clone(),
886 llm_api_token,
887 app_version,
888 CompletionBody {
889 thread_id,
890 prompt_id,
891 mode,
892 provider: zed_llm_client::LanguageModelProvider::OpenAi,
893 model: request.model.clone(),
894 provider_request: serde_json::to_value(&request)?,
895 },
896 )
897 .await?;
898
899 let mut mapper = OpenAiEventMapper::new();
900 Ok(map_cloud_completion_events(
901 Box::pin(
902 response_lines(response, includes_status_messages)
903 .chain(usage_updated_event(usage))
904 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
905 ),
906 move |event| mapper.map_event(event),
907 ))
908 });
909 async move { Ok(future.await?.boxed()) }.boxed()
910 }
911 CloudModel::Google(model) => {
912 let client = self.client.clone();
913 let request = into_google(request, model.id().into());
914 let llm_api_token = self.llm_api_token.clone();
915 let future = self.request_limiter.stream(async move {
916 let PerformLlmCompletionResponse {
917 response,
918 usage,
919 includes_status_messages,
920 tool_use_limit_reached,
921 } = Self::perform_llm_completion(
922 client.clone(),
923 llm_api_token,
924 app_version,
925 CompletionBody {
926 thread_id,
927 prompt_id,
928 mode,
929 provider: zed_llm_client::LanguageModelProvider::Google,
930 model: request.model.model_id.clone(),
931 provider_request: serde_json::to_value(&request)?,
932 },
933 )
934 .await?;
935
936 let mut mapper = GoogleEventMapper::new();
937 Ok(map_cloud_completion_events(
938 Box::pin(
939 response_lines(response, includes_status_messages)
940 .chain(usage_updated_event(usage))
941 .chain(tool_use_limit_reached_event(tool_use_limit_reached)),
942 ),
943 move |event| mapper.map_event(event),
944 ))
945 });
946 async move { Ok(future.await?.boxed()) }.boxed()
947 }
948 }
949 }
950}
951
952#[derive(Serialize, Deserialize)]
953#[serde(rename_all = "snake_case")]
954pub enum CloudCompletionEvent<T> {
955 Status(CompletionRequestStatus),
956 Event(T),
957}
958
959fn map_cloud_completion_events<T, F>(
960 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
961 mut map_callback: F,
962) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
963where
964 T: DeserializeOwned + 'static,
965 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
966 + Send
967 + 'static,
968{
969 stream
970 .flat_map(move |event| {
971 futures::stream::iter(match event {
972 Err(error) => {
973 vec![Err(LanguageModelCompletionError::Other(error))]
974 }
975 Ok(CloudCompletionEvent::Status(event)) => {
976 vec![Ok(LanguageModelCompletionEvent::StatusUpdate(event))]
977 }
978 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
979 })
980 })
981 .boxed()
982}
983
984fn usage_updated_event<T>(
985 usage: Option<RequestUsage>,
986) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
987 futures::stream::iter(usage.map(|usage| {
988 Ok(CloudCompletionEvent::Status(
989 CompletionRequestStatus::UsageUpdated {
990 amount: usage.amount as usize,
991 limit: usage.limit,
992 },
993 ))
994 }))
995}
996
997fn tool_use_limit_reached_event<T>(
998 tool_use_limit_reached: bool,
999) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1000 futures::stream::iter(tool_use_limit_reached.then(|| {
1001 Ok(CloudCompletionEvent::Status(
1002 CompletionRequestStatus::ToolUseLimitReached,
1003 ))
1004 }))
1005}
1006
1007fn response_lines<T: DeserializeOwned>(
1008 response: Response<AsyncBody>,
1009 includes_status_messages: bool,
1010) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
1011 futures::stream::try_unfold(
1012 (String::new(), BufReader::new(response.into_body())),
1013 move |(mut line, mut body)| async move {
1014 match body.read_line(&mut line).await {
1015 Ok(0) => Ok(None),
1016 Ok(_) => {
1017 let event = if includes_status_messages {
1018 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
1019 } else {
1020 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
1021 };
1022
1023 line.clear();
1024 Ok(Some((event, (line, body))))
1025 }
1026 Err(e) => Err(e.into()),
1027 }
1028 },
1029 )
1030}
1031
1032struct ConfigurationView {
1033 state: gpui::Entity<State>,
1034}
1035
1036impl ConfigurationView {
1037 fn authenticate(&mut self, cx: &mut Context<Self>) {
1038 self.state.update(cx, |state, cx| {
1039 state.authenticate(cx).detach_and_log_err(cx);
1040 });
1041 cx.notify();
1042 }
1043}
1044
1045impl Render for ConfigurationView {
1046 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1047 const ZED_PRICING_URL: &str = "https://zed.dev/pricing";
1048
1049 let is_connected = !self.state.read(cx).is_signed_out();
1050 let user_store = self.state.read(cx).user_store.read(cx);
1051 let plan = user_store.current_plan();
1052 let subscription_period = user_store.subscription_period();
1053 let eligible_for_trial = user_store.trial_started_at().is_none();
1054 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
1055
1056 let is_pro = plan == Some(proto::Plan::ZedPro);
1057 let subscription_text = match (plan, subscription_period) {
1058 (Some(proto::Plan::ZedPro), Some(_)) => {
1059 "You have access to Zed's hosted LLMs through your Zed Pro subscription."
1060 }
1061 (Some(proto::Plan::ZedProTrial), Some(_)) => {
1062 "You have access to Zed's hosted LLMs through your Zed Pro trial."
1063 }
1064 (Some(proto::Plan::Free), Some(_)) => {
1065 "You have basic access to Zed's hosted LLMs through your Zed Free subscription."
1066 }
1067 _ => {
1068 if eligible_for_trial {
1069 "Subscribe for access to Zed's hosted LLMs. Start with a 14 day free trial."
1070 } else {
1071 "Subscribe for access to Zed's hosted LLMs."
1072 }
1073 }
1074 };
1075 let manage_subscription_buttons = if is_pro {
1076 h_flex().child(
1077 Button::new("manage_settings", "Manage Subscription")
1078 .style(ButtonStyle::Tinted(TintColor::Accent))
1079 .on_click(cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx)))),
1080 )
1081 } else {
1082 h_flex()
1083 .gap_2()
1084 .child(
1085 Button::new("learn_more", "Learn more")
1086 .style(ButtonStyle::Subtle)
1087 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_PRICING_URL))),
1088 )
1089 .child(
1090 Button::new("upgrade", "Upgrade")
1091 .style(ButtonStyle::Subtle)
1092 .color(Color::Accent)
1093 .on_click(
1094 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1095 ),
1096 )
1097 };
1098
1099 if is_connected {
1100 v_flex()
1101 .gap_3()
1102 .w_full()
1103 .children(render_accept_terms(
1104 self.state.clone(),
1105 LanguageModelProviderTosView::Configuration,
1106 cx,
1107 ))
1108 .when(has_accepted_terms, |this| {
1109 this.child(subscription_text)
1110 .child(manage_subscription_buttons)
1111 })
1112 } else {
1113 v_flex()
1114 .gap_2()
1115 .child(Label::new("Use Zed AI to access hosted language models."))
1116 .child(
1117 Button::new("sign_in", "Sign In")
1118 .icon_color(Color::Muted)
1119 .icon(IconName::Github)
1120 .icon_position(IconPosition::Start)
1121 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
1122 )
1123 }
1124 }
1125}