1use super::open_ai::count_open_ai_tokens;
2use anthropic::AnthropicError;
3use anyhow::{anyhow, Result};
4use client::{
5 zed_urls, Client, PerformCompletionParams, UserStore, EXPIRED_LLM_TOKEN_HEADER_NAME,
6 MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
7};
8use collections::BTreeMap;
9use feature_flags::{FeatureFlagAppExt, LlmClosedBeta, ZedPro};
10use futures::{
11 future::BoxFuture, stream::BoxStream, AsyncBufReadExt, FutureExt, Stream, StreamExt,
12 TryStreamExt as _,
13};
14use gpui::{
15 AnyElement, AnyView, App, AsyncApp, Context, Entity, EventEmitter, Global, ReadGlobal,
16 Subscription, Task,
17};
18use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
19use language_model::{
20 AuthenticateError, CloudModel, LanguageModel, LanguageModelCacheConfiguration, LanguageModelId,
21 LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,
22 LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest, RateLimiter,
23 ZED_CLOUD_PROVIDER_ID,
24};
25use language_model::{
26 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider,
27};
28use proto::TypedEnvelope;
29use schemars::JsonSchema;
30use serde::{de::DeserializeOwned, Deserialize, Serialize};
31use serde_json::value::RawValue;
32use settings::{Settings, SettingsStore};
33use smol::{
34 io::{AsyncReadExt, BufReader},
35 lock::{RwLock, RwLockUpgradableReadGuard, RwLockWriteGuard},
36};
37use std::fmt;
38use std::{
39 future,
40 sync::{Arc, LazyLock},
41};
42use strum::IntoEnumIterator;
43use thiserror::Error;
44use ui::{prelude::*, TintColor};
45
46use crate::provider::anthropic::map_to_language_model_completion_events;
47use crate::AllLanguageModelSettings;
48
49use super::anthropic::count_anthropic_tokens;
50
51pub const PROVIDER_NAME: &str = "Zed";
52
53const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
54 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
55
56fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
57 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
58 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
59 .map(|json| serde_json::from_str(json).unwrap())
60 .unwrap_or_default()
61 });
62 ADDITIONAL_MODELS.as_slice()
63}
64
65#[derive(Default, Clone, Debug, PartialEq)]
66pub struct ZedDotDevSettings {
67 pub available_models: Vec<AvailableModel>,
68}
69
70#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
71#[serde(rename_all = "lowercase")]
72pub enum AvailableProvider {
73 Anthropic,
74 OpenAi,
75 Google,
76}
77
78#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
79pub struct AvailableModel {
80 /// The provider of the language model.
81 pub provider: AvailableProvider,
82 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
83 pub name: String,
84 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
85 pub display_name: Option<String>,
86 /// The size of the context window, indicating the maximum number of tokens the model can process.
87 pub max_tokens: usize,
88 /// The maximum number of output tokens allowed by the model.
89 pub max_output_tokens: Option<u32>,
90 /// The maximum number of completion tokens allowed by the model (o1-* only)
91 pub max_completion_tokens: Option<u32>,
92 /// Override this model with a different Anthropic model for tool calls.
93 pub tool_override: Option<String>,
94 /// Indicates whether this custom model supports caching.
95 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
96 /// The default temperature to use for this model.
97 pub default_temperature: Option<f32>,
98 /// Any extra beta headers to provide when using the model.
99 #[serde(default)]
100 pub extra_beta_headers: Vec<String>,
101}
102
103struct GlobalRefreshLlmTokenListener(Entity<RefreshLlmTokenListener>);
104
105impl Global for GlobalRefreshLlmTokenListener {}
106
107pub struct RefreshLlmTokenEvent;
108
109pub struct RefreshLlmTokenListener {
110 _llm_token_subscription: client::Subscription,
111}
112
113impl EventEmitter<RefreshLlmTokenEvent> for RefreshLlmTokenListener {}
114
115impl RefreshLlmTokenListener {
116 pub fn register(client: Arc<Client>, cx: &mut App) {
117 let listener = cx.new(|cx| RefreshLlmTokenListener::new(client, cx));
118 cx.set_global(GlobalRefreshLlmTokenListener(listener));
119 }
120
121 pub fn global(cx: &App) -> Entity<Self> {
122 GlobalRefreshLlmTokenListener::global(cx).0.clone()
123 }
124
125 fn new(client: Arc<Client>, cx: &mut Context<Self>) -> Self {
126 Self {
127 _llm_token_subscription: client
128 .add_message_handler(cx.weak_entity(), Self::handle_refresh_llm_token),
129 }
130 }
131
132 async fn handle_refresh_llm_token(
133 this: Entity<Self>,
134 _: TypedEnvelope<proto::RefreshLlmToken>,
135 mut cx: AsyncApp,
136 ) -> Result<()> {
137 this.update(&mut cx, |_this, cx| cx.emit(RefreshLlmTokenEvent))
138 }
139}
140
141pub struct CloudLanguageModelProvider {
142 client: Arc<Client>,
143 state: gpui::Entity<State>,
144 _maintain_client_status: Task<()>,
145}
146
147pub struct State {
148 client: Arc<Client>,
149 llm_api_token: LlmApiToken,
150 user_store: Entity<UserStore>,
151 status: client::Status,
152 accept_terms: Option<Task<Result<()>>>,
153 _settings_subscription: Subscription,
154 _llm_token_subscription: Subscription,
155}
156
157impl State {
158 fn new(
159 client: Arc<Client>,
160 user_store: Entity<UserStore>,
161 status: client::Status,
162 cx: &mut Context<Self>,
163 ) -> Self {
164 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
165
166 Self {
167 client: client.clone(),
168 llm_api_token: LlmApiToken::default(),
169 user_store,
170 status,
171 accept_terms: None,
172 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
173 cx.notify();
174 }),
175 _llm_token_subscription: cx.subscribe(
176 &refresh_llm_token_listener,
177 |this, _listener, _event, cx| {
178 let client = this.client.clone();
179 let llm_api_token = this.llm_api_token.clone();
180 cx.spawn(|_this, _cx| async move {
181 llm_api_token.refresh(&client).await?;
182 anyhow::Ok(())
183 })
184 .detach_and_log_err(cx);
185 },
186 ),
187 }
188 }
189
190 fn is_signed_out(&self) -> bool {
191 self.status.is_signed_out()
192 }
193
194 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
195 let client = self.client.clone();
196 cx.spawn(move |this, mut cx| async move {
197 client.authenticate_and_connect(true, &cx).await?;
198 this.update(&mut cx, |_, cx| cx.notify())
199 })
200 }
201
202 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
203 self.user_store
204 .read(cx)
205 .current_user_has_accepted_terms()
206 .unwrap_or(false)
207 }
208
209 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
210 let user_store = self.user_store.clone();
211 self.accept_terms = Some(cx.spawn(move |this, mut cx| async move {
212 let _ = user_store
213 .update(&mut cx, |store, cx| store.accept_terms_of_service(cx))?
214 .await;
215 this.update(&mut cx, |this, cx| {
216 this.accept_terms = None;
217 cx.notify()
218 })
219 }));
220 }
221}
222
223impl CloudLanguageModelProvider {
224 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
225 let mut status_rx = client.status();
226 let status = *status_rx.borrow();
227
228 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
229
230 let state_ref = state.downgrade();
231 let maintain_client_status = cx.spawn(|mut cx| async move {
232 while let Some(status) = status_rx.next().await {
233 if let Some(this) = state_ref.upgrade() {
234 _ = this.update(&mut cx, |this, cx| {
235 if this.status != status {
236 this.status = status;
237 cx.notify();
238 }
239 });
240 } else {
241 break;
242 }
243 }
244 });
245
246 Self {
247 client,
248 state: state.clone(),
249 _maintain_client_status: maintain_client_status,
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::default());
278 Some(Arc::new(CloudLanguageModel {
279 id: LanguageModelId::from(model.id().to_string()),
280 model,
281 llm_api_token: llm_api_token.clone(),
282 client: self.client.clone(),
283 request_limiter: RateLimiter::new(4),
284 }))
285 }
286
287 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
288 let mut models = BTreeMap::default();
289
290 if cx.is_staff() {
291 for model in anthropic::Model::iter() {
292 if !matches!(model, anthropic::Model::Custom { .. }) {
293 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
294 }
295 }
296 for model in open_ai::Model::iter() {
297 if !matches!(model, open_ai::Model::Custom { .. }) {
298 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
299 }
300 }
301 for model in google_ai::Model::iter() {
302 if !matches!(model, google_ai::Model::Custom { .. }) {
303 models.insert(model.id().to_string(), CloudModel::Google(model));
304 }
305 }
306 } else {
307 models.insert(
308 anthropic::Model::Claude3_5Sonnet.id().to_string(),
309 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
310 );
311 }
312
313 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBeta>() {
314 zed_cloud_provider_additional_models()
315 } else {
316 &[]
317 };
318
319 // Override with available models from settings
320 for model in AllLanguageModelSettings::get_global(cx)
321 .zed_dot_dev
322 .available_models
323 .iter()
324 .chain(llm_closed_beta_models)
325 .cloned()
326 {
327 let model = match model.provider {
328 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
329 name: model.name.clone(),
330 display_name: model.display_name.clone(),
331 max_tokens: model.max_tokens,
332 tool_override: model.tool_override.clone(),
333 cache_configuration: model.cache_configuration.as_ref().map(|config| {
334 anthropic::AnthropicModelCacheConfiguration {
335 max_cache_anchors: config.max_cache_anchors,
336 should_speculate: config.should_speculate,
337 min_total_token: config.min_total_token,
338 }
339 }),
340 default_temperature: model.default_temperature,
341 max_output_tokens: model.max_output_tokens,
342 extra_beta_headers: model.extra_beta_headers.clone(),
343 }),
344 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
345 name: model.name.clone(),
346 display_name: model.display_name.clone(),
347 max_tokens: model.max_tokens,
348 max_output_tokens: model.max_output_tokens,
349 max_completion_tokens: model.max_completion_tokens,
350 }),
351 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
352 name: model.name.clone(),
353 display_name: model.display_name.clone(),
354 max_tokens: model.max_tokens,
355 }),
356 };
357 models.insert(model.id().to_string(), model.clone());
358 }
359
360 let llm_api_token = self.state.read(cx).llm_api_token.clone();
361 models
362 .into_values()
363 .map(|model| {
364 Arc::new(CloudLanguageModel {
365 id: LanguageModelId::from(model.id().to_string()),
366 model,
367 llm_api_token: llm_api_token.clone(),
368 client: self.client.clone(),
369 request_limiter: RateLimiter::new(4),
370 }) as Arc<dyn LanguageModel>
371 })
372 .collect()
373 }
374
375 fn is_authenticated(&self, cx: &App) -> bool {
376 !self.state.read(cx).is_signed_out()
377 }
378
379 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
380 Task::ready(Ok(()))
381 }
382
383 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
384 cx.new(|_| ConfigurationView {
385 state: self.state.clone(),
386 })
387 .into()
388 }
389
390 fn must_accept_terms(&self, cx: &App) -> bool {
391 !self.state.read(cx).has_accepted_terms_of_service(cx)
392 }
393
394 fn render_accept_terms(
395 &self,
396 view: LanguageModelProviderTosView,
397 cx: &mut App,
398 ) -> Option<AnyElement> {
399 render_accept_terms(self.state.clone(), view, cx)
400 }
401
402 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
403 Task::ready(Ok(()))
404 }
405}
406
407fn render_accept_terms(
408 state: Entity<State>,
409 view_kind: LanguageModelProviderTosView,
410 cx: &mut App,
411) -> Option<AnyElement> {
412 if state.read(cx).has_accepted_terms_of_service(cx) {
413 return None;
414 }
415
416 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
417
418 let terms_button = Button::new("terms_of_service", "Terms of Service")
419 .style(ButtonStyle::Subtle)
420 .icon(IconName::ArrowUpRight)
421 .icon_color(Color::Muted)
422 .icon_size(IconSize::XSmall)
423 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
424
425 let text = "To start using Zed AI, please read and accept the";
426
427 let form = v_flex()
428 .w_full()
429 .gap_2()
430 .when(
431 view_kind == LanguageModelProviderTosView::ThreadEmptyState,
432 |form| form.items_center(),
433 )
434 .child(
435 h_flex()
436 .flex_wrap()
437 .when(
438 view_kind == LanguageModelProviderTosView::ThreadEmptyState,
439 |form| form.justify_center(),
440 )
441 .child(Label::new(text))
442 .child(terms_button),
443 )
444 .child({
445 let button_container = h_flex().w_full().child(
446 Button::new("accept_terms", "I accept the Terms of Service")
447 .style(ButtonStyle::Tinted(TintColor::Accent))
448 .disabled(accept_terms_disabled)
449 .on_click({
450 let state = state.downgrade();
451 move |_, _window, cx| {
452 state
453 .update(cx, |state, cx| state.accept_terms_of_service(cx))
454 .ok();
455 }
456 }),
457 );
458
459 match view_kind {
460 LanguageModelProviderTosView::ThreadEmptyState => button_container.justify_center(),
461 LanguageModelProviderTosView::PromptEditorPopup => button_container.justify_end(),
462 LanguageModelProviderTosView::Configuration => button_container.justify_start(),
463 }
464 });
465
466 Some(form.into_any())
467}
468
469pub struct CloudLanguageModel {
470 id: LanguageModelId,
471 model: CloudModel,
472 llm_api_token: LlmApiToken,
473 client: Arc<Client>,
474 request_limiter: RateLimiter,
475}
476
477#[derive(Clone, Default)]
478pub struct LlmApiToken(Arc<RwLock<Option<String>>>);
479
480#[derive(Error, Debug)]
481pub struct PaymentRequiredError;
482
483impl fmt::Display for PaymentRequiredError {
484 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
485 write!(
486 f,
487 "Payment required to use this language model. Please upgrade your account."
488 )
489 }
490}
491
492#[derive(Error, Debug)]
493pub struct MaxMonthlySpendReachedError;
494
495impl fmt::Display for MaxMonthlySpendReachedError {
496 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
497 write!(
498 f,
499 "Maximum spending limit reached for this month. For more usage, increase your spending limit."
500 )
501 }
502}
503
504impl CloudLanguageModel {
505 async fn perform_llm_completion(
506 client: Arc<Client>,
507 llm_api_token: LlmApiToken,
508 body: PerformCompletionParams,
509 ) -> Result<Response<AsyncBody>> {
510 let http_client = &client.http_client();
511
512 let mut token = llm_api_token.acquire(&client).await?;
513 let mut did_retry = false;
514
515 let response = loop {
516 let request_builder = http_client::Request::builder();
517 let request = request_builder
518 .method(Method::POST)
519 .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
520 .header("Content-Type", "application/json")
521 .header("Authorization", format!("Bearer {token}"))
522 .body(serde_json::to_string(&body)?.into())?;
523 let mut response = http_client.send(request).await?;
524 if response.status().is_success() {
525 break response;
526 } else if !did_retry
527 && response
528 .headers()
529 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
530 .is_some()
531 {
532 did_retry = true;
533 token = llm_api_token.refresh(&client).await?;
534 } else if response.status() == StatusCode::FORBIDDEN
535 && response
536 .headers()
537 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
538 .is_some()
539 {
540 break Err(anyhow!(MaxMonthlySpendReachedError))?;
541 } else if response.status() == StatusCode::PAYMENT_REQUIRED {
542 break Err(anyhow!(PaymentRequiredError))?;
543 } else {
544 let mut body = String::new();
545 response.body_mut().read_to_string(&mut body).await?;
546 break Err(anyhow!(
547 "cloud language model completion failed with status {}: {body}",
548 response.status()
549 ))?;
550 }
551 };
552
553 Ok(response)
554 }
555}
556
557impl LanguageModel for CloudLanguageModel {
558 fn id(&self) -> LanguageModelId {
559 self.id.clone()
560 }
561
562 fn name(&self) -> LanguageModelName {
563 LanguageModelName::from(self.model.display_name().to_string())
564 }
565
566 fn icon(&self) -> Option<IconName> {
567 self.model.icon()
568 }
569
570 fn provider_id(&self) -> LanguageModelProviderId {
571 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
572 }
573
574 fn provider_name(&self) -> LanguageModelProviderName {
575 LanguageModelProviderName(PROVIDER_NAME.into())
576 }
577
578 fn telemetry_id(&self) -> String {
579 format!("zed.dev/{}", self.model.id())
580 }
581
582 fn availability(&self) -> LanguageModelAvailability {
583 self.model.availability()
584 }
585
586 fn max_token_count(&self) -> usize {
587 self.model.max_token_count()
588 }
589
590 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
591 match &self.model {
592 CloudModel::Anthropic(model) => {
593 model
594 .cache_configuration()
595 .map(|cache| LanguageModelCacheConfiguration {
596 max_cache_anchors: cache.max_cache_anchors,
597 should_speculate: cache.should_speculate,
598 min_total_token: cache.min_total_token,
599 })
600 }
601 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
602 }
603 }
604
605 fn count_tokens(
606 &self,
607 request: LanguageModelRequest,
608 cx: &App,
609 ) -> BoxFuture<'static, Result<usize>> {
610 match self.model.clone() {
611 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
612 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
613 CloudModel::Google(model) => {
614 let client = self.client.clone();
615 let request = request.into_google(model.id().into());
616 let request = google_ai::CountTokensRequest {
617 contents: request.contents,
618 };
619 async move {
620 let request = serde_json::to_string(&request)?;
621 let response = client
622 .request(proto::CountLanguageModelTokens {
623 provider: proto::LanguageModelProvider::Google as i32,
624 request,
625 })
626 .await?;
627 Ok(response.token_count as usize)
628 }
629 .boxed()
630 }
631 }
632 }
633
634 fn stream_completion(
635 &self,
636 request: LanguageModelRequest,
637 _cx: &AsyncApp,
638 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
639 match &self.model {
640 CloudModel::Anthropic(model) => {
641 let request = request.into_anthropic(
642 model.id().into(),
643 model.default_temperature(),
644 model.max_output_tokens(),
645 );
646 let client = self.client.clone();
647 let llm_api_token = self.llm_api_token.clone();
648 let future = self.request_limiter.stream(async move {
649 let response = Self::perform_llm_completion(
650 client.clone(),
651 llm_api_token,
652 PerformCompletionParams {
653 provider: client::LanguageModelProvider::Anthropic,
654 model: request.model.clone(),
655 provider_request: RawValue::from_string(serde_json::to_string(
656 &request,
657 )?)?,
658 },
659 )
660 .await?;
661 Ok(map_to_language_model_completion_events(Box::pin(
662 response_lines(response).map_err(AnthropicError::Other),
663 )))
664 });
665 async move { Ok(future.await?.boxed()) }.boxed()
666 }
667 CloudModel::OpenAi(model) => {
668 let client = self.client.clone();
669 let request = request.into_open_ai(model.id().into(), model.max_output_tokens());
670 let llm_api_token = self.llm_api_token.clone();
671 let future = self.request_limiter.stream(async move {
672 let response = Self::perform_llm_completion(
673 client.clone(),
674 llm_api_token,
675 PerformCompletionParams {
676 provider: client::LanguageModelProvider::OpenAi,
677 model: request.model.clone(),
678 provider_request: RawValue::from_string(serde_json::to_string(
679 &request,
680 )?)?,
681 },
682 )
683 .await?;
684 Ok(open_ai::extract_text_from_events(response_lines(response)))
685 });
686 async move {
687 Ok(future
688 .await?
689 .map(|result| result.map(LanguageModelCompletionEvent::Text))
690 .boxed())
691 }
692 .boxed()
693 }
694 CloudModel::Google(model) => {
695 let client = self.client.clone();
696 let request = request.into_google(model.id().into());
697 let llm_api_token = self.llm_api_token.clone();
698 let future = self.request_limiter.stream(async move {
699 let response = Self::perform_llm_completion(
700 client.clone(),
701 llm_api_token,
702 PerformCompletionParams {
703 provider: client::LanguageModelProvider::Google,
704 model: request.model.clone(),
705 provider_request: RawValue::from_string(serde_json::to_string(
706 &request,
707 )?)?,
708 },
709 )
710 .await?;
711 Ok(google_ai::extract_text_from_events(response_lines(
712 response,
713 )))
714 });
715 async move {
716 Ok(future
717 .await?
718 .map(|result| result.map(LanguageModelCompletionEvent::Text))
719 .boxed())
720 }
721 .boxed()
722 }
723 }
724 }
725
726 fn use_any_tool(
727 &self,
728 request: LanguageModelRequest,
729 tool_name: String,
730 tool_description: String,
731 input_schema: serde_json::Value,
732 _cx: &AsyncApp,
733 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
734 let client = self.client.clone();
735 let llm_api_token = self.llm_api_token.clone();
736
737 match &self.model {
738 CloudModel::Anthropic(model) => {
739 let mut request = request.into_anthropic(
740 model.tool_model_id().into(),
741 model.default_temperature(),
742 model.max_output_tokens(),
743 );
744 request.tool_choice = Some(anthropic::ToolChoice::Tool {
745 name: tool_name.clone(),
746 });
747 request.tools = vec![anthropic::Tool {
748 name: tool_name.clone(),
749 description: tool_description,
750 input_schema,
751 }];
752
753 self.request_limiter
754 .run(async move {
755 let response = Self::perform_llm_completion(
756 client.clone(),
757 llm_api_token,
758 PerformCompletionParams {
759 provider: client::LanguageModelProvider::Anthropic,
760 model: request.model.clone(),
761 provider_request: RawValue::from_string(serde_json::to_string(
762 &request,
763 )?)?,
764 },
765 )
766 .await?;
767
768 Ok(anthropic::extract_tool_args_from_events(
769 tool_name,
770 Box::pin(response_lines(response)),
771 )
772 .await?
773 .boxed())
774 })
775 .boxed()
776 }
777 CloudModel::OpenAi(model) => {
778 let mut request =
779 request.into_open_ai(model.id().into(), model.max_output_tokens());
780 request.tool_choice = Some(open_ai::ToolChoice::Other(
781 open_ai::ToolDefinition::Function {
782 function: open_ai::FunctionDefinition {
783 name: tool_name.clone(),
784 description: None,
785 parameters: None,
786 },
787 },
788 ));
789 request.tools = vec![open_ai::ToolDefinition::Function {
790 function: open_ai::FunctionDefinition {
791 name: tool_name.clone(),
792 description: Some(tool_description),
793 parameters: Some(input_schema),
794 },
795 }];
796
797 self.request_limiter
798 .run(async move {
799 let response = Self::perform_llm_completion(
800 client.clone(),
801 llm_api_token,
802 PerformCompletionParams {
803 provider: client::LanguageModelProvider::OpenAi,
804 model: request.model.clone(),
805 provider_request: RawValue::from_string(serde_json::to_string(
806 &request,
807 )?)?,
808 },
809 )
810 .await?;
811
812 Ok(open_ai::extract_tool_args_from_events(
813 tool_name,
814 Box::pin(response_lines(response)),
815 )
816 .await?
817 .boxed())
818 })
819 .boxed()
820 }
821 CloudModel::Google(_) => {
822 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
823 }
824 }
825 }
826}
827
828fn response_lines<T: DeserializeOwned>(
829 response: Response<AsyncBody>,
830) -> impl Stream<Item = Result<T>> {
831 futures::stream::try_unfold(
832 (String::new(), BufReader::new(response.into_body())),
833 move |(mut line, mut body)| async {
834 match body.read_line(&mut line).await {
835 Ok(0) => Ok(None),
836 Ok(_) => {
837 let event: T = serde_json::from_str(&line)?;
838 line.clear();
839 Ok(Some((event, (line, body))))
840 }
841 Err(e) => Err(e.into()),
842 }
843 },
844 )
845}
846
847impl LlmApiToken {
848 pub async fn acquire(&self, client: &Arc<Client>) -> Result<String> {
849 let lock = self.0.upgradable_read().await;
850 if let Some(token) = lock.as_ref() {
851 Ok(token.to_string())
852 } else {
853 Self::fetch(RwLockUpgradableReadGuard::upgrade(lock).await, client).await
854 }
855 }
856
857 pub async fn refresh(&self, client: &Arc<Client>) -> Result<String> {
858 Self::fetch(self.0.write().await, client).await
859 }
860
861 async fn fetch<'a>(
862 mut lock: RwLockWriteGuard<'a, Option<String>>,
863 client: &Arc<Client>,
864 ) -> Result<String> {
865 let response = client.request(proto::GetLlmToken {}).await?;
866 *lock = Some(response.token.clone());
867 Ok(response.token.clone())
868 }
869}
870
871struct ConfigurationView {
872 state: gpui::Entity<State>,
873}
874
875impl ConfigurationView {
876 fn authenticate(&mut self, cx: &mut Context<Self>) {
877 self.state.update(cx, |state, cx| {
878 state.authenticate(cx).detach_and_log_err(cx);
879 });
880 cx.notify();
881 }
882}
883
884impl Render for ConfigurationView {
885 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
886 const ZED_AI_URL: &str = "https://zed.dev/ai";
887
888 let is_connected = !self.state.read(cx).is_signed_out();
889 let plan = self.state.read(cx).user_store.read(cx).current_plan();
890 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
891
892 let is_pro = plan == Some(proto::Plan::ZedPro);
893 let subscription_text = Label::new(if is_pro {
894 "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."
895 } else {
896 "You have basic access to models from Anthropic through the Zed AI Free plan."
897 });
898 let manage_subscription_button = if is_pro {
899 Some(
900 h_flex().child(
901 Button::new("manage_settings", "Manage Subscription")
902 .style(ButtonStyle::Tinted(TintColor::Accent))
903 .on_click(
904 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
905 ),
906 ),
907 )
908 } else if cx.has_flag::<ZedPro>() {
909 Some(
910 h_flex()
911 .gap_2()
912 .child(
913 Button::new("learn_more", "Learn more")
914 .style(ButtonStyle::Subtle)
915 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
916 )
917 .child(
918 Button::new("upgrade", "Upgrade")
919 .style(ButtonStyle::Subtle)
920 .color(Color::Accent)
921 .on_click(
922 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
923 ),
924 ),
925 )
926 } else {
927 None
928 };
929
930 if is_connected {
931 v_flex()
932 .gap_3()
933 .w_full()
934 .children(render_accept_terms(
935 self.state.clone(),
936 LanguageModelProviderTosView::Configuration,
937 cx,
938 ))
939 .when(has_accepted_terms, |this| {
940 this.child(subscription_text)
941 .children(manage_subscription_button)
942 })
943 } else {
944 v_flex()
945 .gap_2()
946 .child(Label::new("Use Zed AI to access hosted language models."))
947 .child(
948 Button::new("sign_in", "Sign In")
949 .icon_color(Color::Muted)
950 .icon(IconName::Github)
951 .icon_position(IconPosition::Start)
952 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
953 )
954 }
955 }
956}