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::{AnyElement, AnyView, App, AsyncApp, Context, Entity, Subscription, Task};
10use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
11use language_model::{
12 AuthenticateError, CloudModel, LanguageModel, LanguageModelCacheConfiguration,
13 LanguageModelCompletionError, LanguageModelId, LanguageModelKnownError, LanguageModelName,
14 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
15 LanguageModelProviderTosView, LanguageModelRequest, LanguageModelToolSchemaFormat,
16 ModelRequestLimitReachedError, QueueState, RateLimiter, RequestUsage, ZED_CLOUD_PROVIDER_ID,
17};
18use language_model::{
19 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider, LlmApiToken,
20 MaxMonthlySpendReachedError, PaymentRequiredError, RefreshLlmTokenListener,
21};
22use proto::Plan;
23use schemars::JsonSchema;
24use serde::{Deserialize, Serialize, de::DeserializeOwned};
25use settings::{Settings, SettingsStore};
26use smol::Timer;
27use smol::io::{AsyncReadExt, BufReader};
28use std::pin::Pin;
29use std::str::FromStr as _;
30use std::{
31 sync::{Arc, LazyLock},
32 time::Duration,
33};
34use strum::IntoEnumIterator;
35use thiserror::Error;
36use ui::{TintColor, prelude::*};
37use zed_llm_client::{
38 CURRENT_PLAN_HEADER_NAME, CompletionBody, CountTokensBody, CountTokensResponse,
39 EXPIRED_LLM_TOKEN_HEADER_NAME, MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
40 MODEL_REQUESTS_RESOURCE_HEADER_VALUE, SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME,
41};
42
43use crate::AllLanguageModelSettings;
44use crate::provider::anthropic::{AnthropicEventMapper, count_anthropic_tokens, into_anthropic};
45use crate::provider::google::{GoogleEventMapper, into_google};
46use crate::provider::open_ai::{OpenAiEventMapper, count_open_ai_tokens, into_open_ai};
47
48pub const PROVIDER_NAME: &str = "Zed";
49
50const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
51 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
52
53fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
54 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
55 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
56 .map(|json| serde_json::from_str(json).unwrap())
57 .unwrap_or_default()
58 });
59 ADDITIONAL_MODELS.as_slice()
60}
61
62#[derive(Default, Clone, Debug, PartialEq)]
63pub struct ZedDotDevSettings {
64 pub available_models: Vec<AvailableModel>,
65}
66
67#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
68#[serde(rename_all = "lowercase")]
69pub enum AvailableProvider {
70 Anthropic,
71 OpenAi,
72 Google,
73}
74
75#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
76pub struct AvailableModel {
77 /// The provider of the language model.
78 pub provider: AvailableProvider,
79 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
80 pub name: String,
81 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
82 pub display_name: Option<String>,
83 /// The size of the context window, indicating the maximum number of tokens the model can process.
84 pub max_tokens: usize,
85 /// The maximum number of output tokens allowed by the model.
86 pub max_output_tokens: Option<u32>,
87 /// The maximum number of completion tokens allowed by the model (o1-* only)
88 pub max_completion_tokens: Option<u32>,
89 /// Override this model with a different Anthropic model for tool calls.
90 pub tool_override: Option<String>,
91 /// Indicates whether this custom model supports caching.
92 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
93 /// The default temperature to use for this model.
94 pub default_temperature: Option<f32>,
95 /// Any extra beta headers to provide when using the model.
96 #[serde(default)]
97 pub extra_beta_headers: Vec<String>,
98 /// The model's mode (e.g. thinking)
99 pub mode: Option<ModelMode>,
100}
101
102#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
103#[serde(tag = "type", rename_all = "lowercase")]
104pub enum ModelMode {
105 #[default]
106 Default,
107 Thinking {
108 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
109 budget_tokens: Option<u32>,
110 },
111}
112
113impl From<ModelMode> for AnthropicModelMode {
114 fn from(value: ModelMode) -> Self {
115 match value {
116 ModelMode::Default => AnthropicModelMode::Default,
117 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
118 }
119 }
120}
121
122pub struct CloudLanguageModelProvider {
123 client: Arc<Client>,
124 state: gpui::Entity<State>,
125 _maintain_client_status: Task<()>,
126}
127
128pub struct State {
129 client: Arc<Client>,
130 llm_api_token: LlmApiToken,
131 user_store: Entity<UserStore>,
132 status: client::Status,
133 accept_terms: Option<Task<Result<()>>>,
134 _settings_subscription: Subscription,
135 _llm_token_subscription: Subscription,
136}
137
138impl State {
139 fn new(
140 client: Arc<Client>,
141 user_store: Entity<UserStore>,
142 status: client::Status,
143 cx: &mut Context<Self>,
144 ) -> Self {
145 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
146
147 Self {
148 client: client.clone(),
149 llm_api_token: LlmApiToken::default(),
150 user_store,
151 status,
152 accept_terms: None,
153 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
154 cx.notify();
155 }),
156 _llm_token_subscription: cx.subscribe(
157 &refresh_llm_token_listener,
158 |this, _listener, _event, cx| {
159 let client = this.client.clone();
160 let llm_api_token = this.llm_api_token.clone();
161 cx.spawn(async move |_this, _cx| {
162 llm_api_token.refresh(&client).await?;
163 anyhow::Ok(())
164 })
165 .detach_and_log_err(cx);
166 },
167 ),
168 }
169 }
170
171 fn is_signed_out(&self) -> bool {
172 self.status.is_signed_out()
173 }
174
175 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
176 let client = self.client.clone();
177 cx.spawn(async move |this, cx| {
178 client.authenticate_and_connect(true, &cx).await?;
179 this.update(cx, |_, cx| cx.notify())
180 })
181 }
182
183 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
184 self.user_store
185 .read(cx)
186 .current_user_has_accepted_terms()
187 .unwrap_or(false)
188 }
189
190 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
191 let user_store = self.user_store.clone();
192 self.accept_terms = Some(cx.spawn(async move |this, cx| {
193 let _ = user_store
194 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
195 .await;
196 this.update(cx, |this, cx| {
197 this.accept_terms = None;
198 cx.notify()
199 })
200 }));
201 }
202}
203
204impl CloudLanguageModelProvider {
205 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
206 let mut status_rx = client.status();
207 let status = *status_rx.borrow();
208
209 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
210
211 let state_ref = state.downgrade();
212 let maintain_client_status = cx.spawn(async move |cx| {
213 while let Some(status) = status_rx.next().await {
214 if let Some(this) = state_ref.upgrade() {
215 _ = this.update(cx, |this, cx| {
216 if this.status != status {
217 this.status = status;
218 cx.notify();
219 }
220 });
221 } else {
222 break;
223 }
224 }
225 });
226
227 Self {
228 client,
229 state: state.clone(),
230 _maintain_client_status: maintain_client_status,
231 }
232 }
233
234 fn create_language_model(
235 &self,
236 model: CloudModel,
237 llm_api_token: LlmApiToken,
238 ) -> Arc<dyn LanguageModel> {
239 Arc::new(CloudLanguageModel {
240 id: LanguageModelId::from(model.id().to_string()),
241 model,
242 llm_api_token: llm_api_token.clone(),
243 client: self.client.clone(),
244 request_limiter: RateLimiter::new(4),
245 })
246 }
247}
248
249impl LanguageModelProviderState for CloudLanguageModelProvider {
250 type ObservableEntity = State;
251
252 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
253 Some(self.state.clone())
254 }
255}
256
257impl LanguageModelProvider for CloudLanguageModelProvider {
258 fn id(&self) -> LanguageModelProviderId {
259 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
260 }
261
262 fn name(&self) -> LanguageModelProviderName {
263 LanguageModelProviderName(PROVIDER_NAME.into())
264 }
265
266 fn icon(&self) -> IconName {
267 IconName::AiZed
268 }
269
270 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
271 let llm_api_token = self.state.read(cx).llm_api_token.clone();
272 let model = CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet);
273 Some(self.create_language_model(model, llm_api_token))
274 }
275
276 fn default_fast_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_5Sonnet);
279 Some(self.create_language_model(model, llm_api_token))
280 }
281
282 fn recommended_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
283 let llm_api_token = self.state.read(cx).llm_api_token.clone();
284 [
285 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
286 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
287 ]
288 .into_iter()
289 .map(|model| self.create_language_model(model, llm_api_token.clone()))
290 .collect()
291 }
292
293 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
294 let mut models = BTreeMap::default();
295
296 if cx.is_staff() {
297 for model in anthropic::Model::iter() {
298 if !matches!(model, anthropic::Model::Custom { .. }) {
299 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
300 }
301 }
302 for model in open_ai::Model::iter() {
303 if !matches!(model, open_ai::Model::Custom { .. }) {
304 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
305 }
306 }
307 for model in google_ai::Model::iter() {
308 if !matches!(model, google_ai::Model::Custom { .. }) {
309 models.insert(model.id().to_string(), CloudModel::Google(model));
310 }
311 }
312 } else {
313 models.insert(
314 anthropic::Model::Claude3_5Sonnet.id().to_string(),
315 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
316 );
317 models.insert(
318 anthropic::Model::Claude3_7Sonnet.id().to_string(),
319 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
320 );
321 models.insert(
322 anthropic::Model::Claude3_7SonnetThinking.id().to_string(),
323 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
324 );
325 }
326
327 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBetaFeatureFlag>() {
328 zed_cloud_provider_additional_models()
329 } else {
330 &[]
331 };
332
333 // Override with available models from settings
334 for model in AllLanguageModelSettings::get_global(cx)
335 .zed_dot_dev
336 .available_models
337 .iter()
338 .chain(llm_closed_beta_models)
339 .cloned()
340 {
341 let model = match model.provider {
342 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
343 name: model.name.clone(),
344 display_name: model.display_name.clone(),
345 max_tokens: model.max_tokens,
346 tool_override: model.tool_override.clone(),
347 cache_configuration: model.cache_configuration.as_ref().map(|config| {
348 anthropic::AnthropicModelCacheConfiguration {
349 max_cache_anchors: config.max_cache_anchors,
350 should_speculate: config.should_speculate,
351 min_total_token: config.min_total_token,
352 }
353 }),
354 default_temperature: model.default_temperature,
355 max_output_tokens: model.max_output_tokens,
356 extra_beta_headers: model.extra_beta_headers.clone(),
357 mode: model.mode.unwrap_or_default().into(),
358 }),
359 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
360 name: model.name.clone(),
361 display_name: model.display_name.clone(),
362 max_tokens: model.max_tokens,
363 max_output_tokens: model.max_output_tokens,
364 max_completion_tokens: model.max_completion_tokens,
365 }),
366 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
367 name: model.name.clone(),
368 display_name: model.display_name.clone(),
369 max_tokens: model.max_tokens,
370 }),
371 };
372 models.insert(model.id().to_string(), model.clone());
373 }
374
375 let llm_api_token = self.state.read(cx).llm_api_token.clone();
376 models
377 .into_values()
378 .map(|model| self.create_language_model(model, llm_api_token.clone()))
379 .collect()
380 }
381
382 fn is_authenticated(&self, cx: &App) -> bool {
383 !self.state.read(cx).is_signed_out()
384 }
385
386 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
387 Task::ready(Ok(()))
388 }
389
390 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
391 cx.new(|_| ConfigurationView {
392 state: self.state.clone(),
393 })
394 .into()
395 }
396
397 fn must_accept_terms(&self, cx: &App) -> bool {
398 !self.state.read(cx).has_accepted_terms_of_service(cx)
399 }
400
401 fn render_accept_terms(
402 &self,
403 view: LanguageModelProviderTosView,
404 cx: &mut App,
405 ) -> Option<AnyElement> {
406 render_accept_terms(self.state.clone(), view, cx)
407 }
408
409 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
410 Task::ready(Ok(()))
411 }
412}
413
414fn render_accept_terms(
415 state: Entity<State>,
416 view_kind: LanguageModelProviderTosView,
417 cx: &mut App,
418) -> Option<AnyElement> {
419 if state.read(cx).has_accepted_terms_of_service(cx) {
420 return None;
421 }
422
423 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
424
425 let thread_fresh_start = matches!(view_kind, LanguageModelProviderTosView::ThreadFreshStart);
426 let thread_empty_state = matches!(view_kind, LanguageModelProviderTosView::ThreadtEmptyState);
427
428 let terms_button = Button::new("terms_of_service", "Terms of Service")
429 .style(ButtonStyle::Subtle)
430 .icon(IconName::ArrowUpRight)
431 .icon_color(Color::Muted)
432 .icon_size(IconSize::XSmall)
433 .when(thread_empty_state, |this| this.label_size(LabelSize::Small))
434 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
435
436 let button_container = h_flex().child(
437 Button::new("accept_terms", "I accept the Terms of Service")
438 .when(!thread_empty_state, |this| {
439 this.full_width()
440 .style(ButtonStyle::Tinted(TintColor::Accent))
441 .icon(IconName::Check)
442 .icon_position(IconPosition::Start)
443 .icon_size(IconSize::Small)
444 })
445 .when(thread_empty_state, |this| {
446 this.style(ButtonStyle::Tinted(TintColor::Warning))
447 .label_size(LabelSize::Small)
448 })
449 .disabled(accept_terms_disabled)
450 .on_click({
451 let state = state.downgrade();
452 move |_, _window, cx| {
453 state
454 .update(cx, |state, cx| state.accept_terms_of_service(cx))
455 .ok();
456 }
457 }),
458 );
459
460 let form = if thread_empty_state {
461 h_flex()
462 .w_full()
463 .flex_wrap()
464 .justify_between()
465 .child(
466 h_flex()
467 .child(
468 Label::new("To start using Zed AI, please read and accept the")
469 .size(LabelSize::Small),
470 )
471 .child(terms_button),
472 )
473 .child(button_container)
474 } else {
475 v_flex()
476 .w_full()
477 .gap_2()
478 .child(
479 h_flex()
480 .flex_wrap()
481 .when(thread_fresh_start, |this| this.justify_center())
482 .child(Label::new(
483 "To start using Zed AI, please read and accept the",
484 ))
485 .child(terms_button),
486 )
487 .child({
488 match view_kind {
489 LanguageModelProviderTosView::PromptEditorPopup => {
490 button_container.w_full().justify_end()
491 }
492 LanguageModelProviderTosView::Configuration => {
493 button_container.w_full().justify_start()
494 }
495 LanguageModelProviderTosView::ThreadFreshStart => {
496 button_container.w_full().justify_center()
497 }
498 LanguageModelProviderTosView::ThreadtEmptyState => div().w_0(),
499 }
500 })
501 };
502
503 Some(form.into_any())
504}
505
506pub struct CloudLanguageModel {
507 id: LanguageModelId,
508 model: CloudModel,
509 llm_api_token: LlmApiToken,
510 client: Arc<Client>,
511 request_limiter: RateLimiter,
512}
513
514impl CloudLanguageModel {
515 const MAX_RETRIES: usize = 3;
516
517 async fn perform_llm_completion(
518 client: Arc<Client>,
519 llm_api_token: LlmApiToken,
520 body: CompletionBody,
521 ) -> Result<(Response<AsyncBody>, Option<RequestUsage>, bool)> {
522 let http_client = &client.http_client();
523
524 let mut token = llm_api_token.acquire(&client).await?;
525 let mut retries_remaining = Self::MAX_RETRIES;
526 let mut retry_delay = Duration::from_secs(1);
527
528 loop {
529 let request_builder = http_client::Request::builder().method(Method::POST);
530 let request_builder = if let Ok(completions_url) = std::env::var("ZED_COMPLETIONS_URL")
531 {
532 request_builder.uri(completions_url)
533 } else {
534 request_builder.uri(http_client.build_zed_llm_url("/completions", &[])?.as_ref())
535 };
536 let request = request_builder
537 .header("Content-Type", "application/json")
538 .header("Authorization", format!("Bearer {token}"))
539 .header("x-zed-client-supports-queueing", "true")
540 .body(serde_json::to_string(&body)?.into())?;
541 let mut response = http_client.send(request).await?;
542 let status = response.status();
543 if status.is_success() {
544 let includes_queue_events = response
545 .headers()
546 .get("x-zed-server-supports-queueing")
547 .is_some();
548 let usage = RequestUsage::from_headers(response.headers()).ok();
549
550 return Ok((response, usage, includes_queue_events));
551 } else if response
552 .headers()
553 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
554 .is_some()
555 {
556 retries_remaining -= 1;
557 token = llm_api_token.refresh(&client).await?;
558 } else if status == StatusCode::FORBIDDEN
559 && response
560 .headers()
561 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
562 .is_some()
563 {
564 return Err(anyhow!(MaxMonthlySpendReachedError));
565 } else if status == StatusCode::FORBIDDEN
566 && response
567 .headers()
568 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
569 .is_some()
570 {
571 if let Some(MODEL_REQUESTS_RESOURCE_HEADER_VALUE) = response
572 .headers()
573 .get(SUBSCRIPTION_LIMIT_RESOURCE_HEADER_NAME)
574 .and_then(|resource| resource.to_str().ok())
575 {
576 if let Some(plan) = response
577 .headers()
578 .get(CURRENT_PLAN_HEADER_NAME)
579 .and_then(|plan| plan.to_str().ok())
580 .and_then(|plan| zed_llm_client::Plan::from_str(plan).ok())
581 {
582 let plan = match plan {
583 zed_llm_client::Plan::Free => Plan::Free,
584 zed_llm_client::Plan::ZedPro => Plan::ZedPro,
585 zed_llm_client::Plan::ZedProTrial => Plan::ZedProTrial,
586 };
587 return Err(anyhow!(ModelRequestLimitReachedError { plan }));
588 }
589 }
590
591 return Err(anyhow!("Forbidden"));
592 } else if status.as_u16() >= 500 && status.as_u16() < 600 {
593 // If we encounter an error in the 500 range, retry after a delay.
594 // We've seen at least these in the wild from API providers:
595 // * 500 Internal Server Error
596 // * 502 Bad Gateway
597 // * 529 Service Overloaded
598
599 if retries_remaining == 0 {
600 let mut body = String::new();
601 response.body_mut().read_to_string(&mut body).await?;
602 return Err(anyhow!(
603 "cloud language model completion failed after {} retries with status {status}: {body}",
604 Self::MAX_RETRIES
605 ));
606 }
607
608 Timer::after(retry_delay).await;
609
610 retries_remaining -= 1;
611 retry_delay *= 2; // If it fails again, wait longer.
612 } else if status == StatusCode::PAYMENT_REQUIRED {
613 return Err(anyhow!(PaymentRequiredError));
614 } else {
615 let mut body = String::new();
616 response.body_mut().read_to_string(&mut body).await?;
617 return Err(anyhow!(ApiError { status, body }));
618 }
619 }
620 }
621}
622
623#[derive(Debug, Error)]
624#[error("cloud language model completion failed with status {status}: {body}")]
625struct ApiError {
626 status: StatusCode,
627 body: String,
628}
629
630impl LanguageModel for CloudLanguageModel {
631 fn id(&self) -> LanguageModelId {
632 self.id.clone()
633 }
634
635 fn name(&self) -> LanguageModelName {
636 LanguageModelName::from(self.model.display_name().to_string())
637 }
638
639 fn provider_id(&self) -> LanguageModelProviderId {
640 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
641 }
642
643 fn provider_name(&self) -> LanguageModelProviderName {
644 LanguageModelProviderName(PROVIDER_NAME.into())
645 }
646
647 fn supports_tools(&self) -> bool {
648 match self.model {
649 CloudModel::Anthropic(_) => true,
650 CloudModel::Google(_) => true,
651 CloudModel::OpenAi(_) => true,
652 }
653 }
654
655 fn telemetry_id(&self) -> String {
656 format!("zed.dev/{}", self.model.id())
657 }
658
659 fn availability(&self) -> LanguageModelAvailability {
660 self.model.availability()
661 }
662
663 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
664 self.model.tool_input_format()
665 }
666
667 fn max_token_count(&self) -> usize {
668 self.model.max_token_count()
669 }
670
671 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
672 match &self.model {
673 CloudModel::Anthropic(model) => {
674 model
675 .cache_configuration()
676 .map(|cache| LanguageModelCacheConfiguration {
677 max_cache_anchors: cache.max_cache_anchors,
678 should_speculate: cache.should_speculate,
679 min_total_token: cache.min_total_token,
680 })
681 }
682 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
683 }
684 }
685
686 fn count_tokens(
687 &self,
688 request: LanguageModelRequest,
689 cx: &App,
690 ) -> BoxFuture<'static, Result<usize>> {
691 match self.model.clone() {
692 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
693 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
694 CloudModel::Google(model) => {
695 let client = self.client.clone();
696 let llm_api_token = self.llm_api_token.clone();
697 let request = into_google(request, model.id().into());
698 async move {
699 let http_client = &client.http_client();
700 let token = llm_api_token.acquire(&client).await?;
701
702 let request_builder = http_client::Request::builder().method(Method::POST);
703 let request_builder =
704 if let Ok(completions_url) = std::env::var("ZED_COUNT_TOKENS_URL") {
705 request_builder.uri(completions_url)
706 } else {
707 request_builder.uri(
708 http_client
709 .build_zed_llm_url("/count_tokens", &[])?
710 .as_ref(),
711 )
712 };
713 let request_body = CountTokensBody {
714 provider: zed_llm_client::LanguageModelProvider::Google,
715 model: model.id().into(),
716 provider_request: serde_json::to_value(&google_ai::CountTokensRequest {
717 contents: request.contents,
718 })?,
719 };
720 let request = request_builder
721 .header("Content-Type", "application/json")
722 .header("Authorization", format!("Bearer {token}"))
723 .body(serde_json::to_string(&request_body)?.into())?;
724 let mut response = http_client.send(request).await?;
725 let status = response.status();
726 let mut response_body = String::new();
727 response
728 .body_mut()
729 .read_to_string(&mut response_body)
730 .await?;
731
732 if status.is_success() {
733 let response_body: CountTokensResponse =
734 serde_json::from_str(&response_body)?;
735
736 Ok(response_body.tokens)
737 } else {
738 Err(anyhow!(ApiError {
739 status,
740 body: response_body
741 }))
742 }
743 }
744 .boxed()
745 }
746 }
747 }
748
749 fn stream_completion(
750 &self,
751 request: LanguageModelRequest,
752 cx: &AsyncApp,
753 ) -> BoxFuture<
754 'static,
755 Result<
756 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
757 >,
758 > {
759 self.stream_completion_with_usage(request, cx)
760 .map(|result| result.map(|(stream, _)| stream))
761 .boxed()
762 }
763
764 fn stream_completion_with_usage(
765 &self,
766 request: LanguageModelRequest,
767 _cx: &AsyncApp,
768 ) -> BoxFuture<
769 'static,
770 Result<(
771 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
772 Option<RequestUsage>,
773 )>,
774 > {
775 let thread_id = request.thread_id.clone();
776 let prompt_id = request.prompt_id.clone();
777 let mode = request.mode;
778 match &self.model {
779 CloudModel::Anthropic(model) => {
780 let request = into_anthropic(
781 request,
782 model.request_id().into(),
783 model.default_temperature(),
784 model.max_output_tokens(),
785 model.mode(),
786 );
787 let client = self.client.clone();
788 let llm_api_token = self.llm_api_token.clone();
789 let future = self.request_limiter.stream_with_usage(async move {
790 let (response, usage, includes_queue_events) = Self::perform_llm_completion(
791 client.clone(),
792 llm_api_token,
793 CompletionBody {
794 thread_id,
795 prompt_id,
796 mode,
797 provider: zed_llm_client::LanguageModelProvider::Anthropic,
798 model: request.model.clone(),
799 provider_request: serde_json::to_value(&request)?,
800 },
801 )
802 .await
803 .map_err(|err| match err.downcast::<ApiError>() {
804 Ok(api_err) => {
805 if api_err.status == StatusCode::BAD_REQUEST {
806 if let Some(tokens) = parse_prompt_too_long(&api_err.body) {
807 return anyhow!(
808 LanguageModelKnownError::ContextWindowLimitExceeded {
809 tokens
810 }
811 );
812 }
813 }
814 anyhow!(api_err)
815 }
816 Err(err) => anyhow!(err),
817 })?;
818
819 let mut mapper = AnthropicEventMapper::new();
820 Ok((
821 map_cloud_completion_events(
822 Box::pin(response_lines(response, includes_queue_events)),
823 move |event| mapper.map_event(event),
824 ),
825 usage,
826 ))
827 });
828 async move {
829 let (stream, usage) = future.await?;
830 Ok((stream.boxed(), usage))
831 }
832 .boxed()
833 }
834 CloudModel::OpenAi(model) => {
835 let client = self.client.clone();
836 let request = into_open_ai(request, model, model.max_output_tokens());
837 let llm_api_token = self.llm_api_token.clone();
838 let future = self.request_limiter.stream_with_usage(async move {
839 let (response, usage, includes_queue_events) = Self::perform_llm_completion(
840 client.clone(),
841 llm_api_token,
842 CompletionBody {
843 thread_id,
844 prompt_id,
845 mode,
846 provider: zed_llm_client::LanguageModelProvider::OpenAi,
847 model: request.model.clone(),
848 provider_request: serde_json::to_value(&request)?,
849 },
850 )
851 .await?;
852
853 let mut mapper = OpenAiEventMapper::new();
854 Ok((
855 map_cloud_completion_events(
856 Box::pin(response_lines(response, includes_queue_events)),
857 move |event| mapper.map_event(event),
858 ),
859 usage,
860 ))
861 });
862 async move {
863 let (stream, usage) = future.await?;
864 Ok((stream.boxed(), usage))
865 }
866 .boxed()
867 }
868 CloudModel::Google(model) => {
869 let client = self.client.clone();
870 let request = into_google(request, model.id().into());
871 let llm_api_token = self.llm_api_token.clone();
872 let future = self.request_limiter.stream_with_usage(async move {
873 let (response, usage, includes_queue_events) = Self::perform_llm_completion(
874 client.clone(),
875 llm_api_token,
876 CompletionBody {
877 thread_id,
878 prompt_id,
879 mode,
880 provider: zed_llm_client::LanguageModelProvider::Google,
881 model: request.model.clone(),
882 provider_request: serde_json::to_value(&request)?,
883 },
884 )
885 .await?;
886 let mut mapper = GoogleEventMapper::new();
887 Ok((
888 map_cloud_completion_events(
889 Box::pin(response_lines(response, includes_queue_events)),
890 move |event| mapper.map_event(event),
891 ),
892 usage,
893 ))
894 });
895 async move {
896 let (stream, usage) = future.await?;
897 Ok((stream.boxed(), usage))
898 }
899 .boxed()
900 }
901 }
902 }
903}
904
905#[derive(Serialize, Deserialize)]
906#[serde(rename_all = "snake_case")]
907pub enum CloudCompletionEvent<T> {
908 Queue(QueueState),
909 Event(T),
910}
911
912fn map_cloud_completion_events<T, F>(
913 stream: Pin<Box<dyn Stream<Item = Result<CloudCompletionEvent<T>>> + Send>>,
914 mut map_callback: F,
915) -> BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
916where
917 T: DeserializeOwned + 'static,
918 F: FnMut(T) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
919 + Send
920 + 'static,
921{
922 stream
923 .flat_map(move |event| {
924 futures::stream::iter(match event {
925 Err(error) => {
926 vec![Err(LanguageModelCompletionError::Other(error))]
927 }
928 Ok(CloudCompletionEvent::Queue(event)) => {
929 vec![Ok(LanguageModelCompletionEvent::QueueUpdate(event))]
930 }
931 Ok(CloudCompletionEvent::Event(event)) => map_callback(event),
932 })
933 })
934 .boxed()
935}
936
937fn response_lines<T: DeserializeOwned>(
938 response: Response<AsyncBody>,
939 includes_queue_events: bool,
940) -> impl Stream<Item = Result<CloudCompletionEvent<T>>> {
941 futures::stream::try_unfold(
942 (String::new(), BufReader::new(response.into_body())),
943 move |(mut line, mut body)| async move {
944 match body.read_line(&mut line).await {
945 Ok(0) => Ok(None),
946 Ok(_) => {
947 let event = if includes_queue_events {
948 serde_json::from_str::<CloudCompletionEvent<T>>(&line)?
949 } else {
950 CloudCompletionEvent::Event(serde_json::from_str::<T>(&line)?)
951 };
952
953 line.clear();
954 Ok(Some((event, (line, body))))
955 }
956 Err(e) => Err(e.into()),
957 }
958 },
959 )
960}
961
962struct ConfigurationView {
963 state: gpui::Entity<State>,
964}
965
966impl ConfigurationView {
967 fn authenticate(&mut self, cx: &mut Context<Self>) {
968 self.state.update(cx, |state, cx| {
969 state.authenticate(cx).detach_and_log_err(cx);
970 });
971 cx.notify();
972 }
973}
974
975impl Render for ConfigurationView {
976 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
977 const ZED_AI_URL: &str = "https://zed.dev/ai";
978
979 let is_connected = !self.state.read(cx).is_signed_out();
980 let plan = self.state.read(cx).user_store.read(cx).current_plan();
981 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
982
983 let is_pro = plan == Some(proto::Plan::ZedPro);
984 let subscription_text = Label::new(if is_pro {
985 "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."
986 } else {
987 "You have basic access to models from Anthropic through the Zed AI Free plan."
988 });
989 let manage_subscription_button = if is_pro {
990 Some(
991 h_flex().child(
992 Button::new("manage_settings", "Manage Subscription")
993 .style(ButtonStyle::Tinted(TintColor::Accent))
994 .on_click(
995 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
996 ),
997 ),
998 )
999 } else if cx.has_flag::<ZedProFeatureFlag>() {
1000 Some(
1001 h_flex()
1002 .gap_2()
1003 .child(
1004 Button::new("learn_more", "Learn more")
1005 .style(ButtonStyle::Subtle)
1006 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
1007 )
1008 .child(
1009 Button::new("upgrade", "Upgrade")
1010 .style(ButtonStyle::Subtle)
1011 .color(Color::Accent)
1012 .on_click(
1013 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
1014 ),
1015 ),
1016 )
1017 } else {
1018 None
1019 };
1020
1021 if is_connected {
1022 v_flex()
1023 .gap_3()
1024 .w_full()
1025 .children(render_accept_terms(
1026 self.state.clone(),
1027 LanguageModelProviderTosView::Configuration,
1028 cx,
1029 ))
1030 .when(has_accepted_terms, |this| {
1031 this.child(subscription_text)
1032 .children(manage_subscription_button)
1033 })
1034 } else {
1035 v_flex()
1036 .gap_2()
1037 .child(Label::new("Use Zed AI to access hosted language models."))
1038 .child(
1039 Button::new("sign_in", "Sign In")
1040 .icon_color(Color::Muted)
1041 .icon(IconName::Github)
1042 .icon_position(IconPosition::Start)
1043 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
1044 )
1045 }
1046 }
1047}