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 CloudModel, LanguageModel, LanguageModelCacheConfiguration, LanguageModelId, LanguageModelName,
21 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
22 LanguageModelProviderTosView, LanguageModelRequest, RateLimiter, ZED_CLOUD_PROVIDER_ID,
23};
24use language_model::{
25 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider,
26};
27use proto::TypedEnvelope;
28use schemars::JsonSchema;
29use serde::{de::DeserializeOwned, Deserialize, Serialize};
30use serde_json::value::RawValue;
31use settings::{Settings, SettingsStore};
32use smol::{
33 io::{AsyncReadExt, BufReader},
34 lock::{RwLock, RwLockUpgradableReadGuard, RwLockWriteGuard},
35};
36use std::fmt;
37use std::{
38 future,
39 sync::{Arc, LazyLock},
40};
41use strum::IntoEnumIterator;
42use thiserror::Error;
43use ui::{prelude::*, TintColor};
44
45use crate::provider::anthropic::map_to_language_model_completion_events;
46use crate::AllLanguageModelSettings;
47
48use super::anthropic::count_anthropic_tokens;
49
50pub const PROVIDER_NAME: &str = "Zed";
51
52const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
53 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
54
55fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
56 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
57 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
58 .map(|json| serde_json::from_str(json).unwrap())
59 .unwrap_or_default()
60 });
61 ADDITIONAL_MODELS.as_slice()
62}
63
64#[derive(Default, Clone, Debug, PartialEq)]
65pub struct ZedDotDevSettings {
66 pub available_models: Vec<AvailableModel>,
67}
68
69#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
70#[serde(rename_all = "lowercase")]
71pub enum AvailableProvider {
72 Anthropic,
73 OpenAi,
74 Google,
75}
76
77#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
78pub struct AvailableModel {
79 /// The provider of the language model.
80 pub provider: AvailableProvider,
81 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
82 pub name: String,
83 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
84 pub display_name: Option<String>,
85 /// The size of the context window, indicating the maximum number of tokens the model can process.
86 pub max_tokens: usize,
87 /// The maximum number of output tokens allowed by the model.
88 pub max_output_tokens: Option<u32>,
89 /// The maximum number of completion tokens allowed by the model (o1-* only)
90 pub max_completion_tokens: Option<u32>,
91 /// Override this model with a different Anthropic model for tool calls.
92 pub tool_override: Option<String>,
93 /// Indicates whether this custom model supports caching.
94 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
95 /// The default temperature to use for this model.
96 pub default_temperature: Option<f32>,
97 /// Any extra beta headers to provide when using the model.
98 #[serde(default)]
99 pub extra_beta_headers: Vec<String>,
100}
101
102struct GlobalRefreshLlmTokenListener(Entity<RefreshLlmTokenListener>);
103
104impl Global for GlobalRefreshLlmTokenListener {}
105
106pub struct RefreshLlmTokenEvent;
107
108pub struct RefreshLlmTokenListener {
109 _llm_token_subscription: client::Subscription,
110}
111
112impl EventEmitter<RefreshLlmTokenEvent> for RefreshLlmTokenListener {}
113
114impl RefreshLlmTokenListener {
115 pub fn register(client: Arc<Client>, cx: &mut App) {
116 let listener = cx.new(|cx| RefreshLlmTokenListener::new(client, cx));
117 cx.set_global(GlobalRefreshLlmTokenListener(listener));
118 }
119
120 pub fn global(cx: &App) -> Entity<Self> {
121 GlobalRefreshLlmTokenListener::global(cx).0.clone()
122 }
123
124 fn new(client: Arc<Client>, cx: &mut Context<Self>) -> Self {
125 Self {
126 _llm_token_subscription: client
127 .add_message_handler(cx.weak_entity(), Self::handle_refresh_llm_token),
128 }
129 }
130
131 async fn handle_refresh_llm_token(
132 this: Entity<Self>,
133 _: TypedEnvelope<proto::RefreshLlmToken>,
134 mut cx: AsyncApp,
135 ) -> Result<()> {
136 this.update(&mut cx, |_this, cx| cx.emit(RefreshLlmTokenEvent))
137 }
138}
139
140pub struct CloudLanguageModelProvider {
141 client: Arc<Client>,
142 state: gpui::Entity<State>,
143 _maintain_client_status: Task<()>,
144}
145
146pub struct State {
147 client: Arc<Client>,
148 llm_api_token: LlmApiToken,
149 user_store: Entity<UserStore>,
150 status: client::Status,
151 accept_terms: Option<Task<Result<()>>>,
152 _settings_subscription: Subscription,
153 _llm_token_subscription: Subscription,
154}
155
156impl State {
157 fn new(
158 client: Arc<Client>,
159 user_store: Entity<UserStore>,
160 status: client::Status,
161 cx: &mut Context<Self>,
162 ) -> Self {
163 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
164
165 Self {
166 client: client.clone(),
167 llm_api_token: LlmApiToken::default(),
168 user_store,
169 status,
170 accept_terms: None,
171 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
172 cx.notify();
173 }),
174 _llm_token_subscription: cx.subscribe(
175 &refresh_llm_token_listener,
176 |this, _listener, _event, cx| {
177 let client = this.client.clone();
178 let llm_api_token = this.llm_api_token.clone();
179 cx.spawn(|_this, _cx| async move {
180 llm_api_token.refresh(&client).await?;
181 anyhow::Ok(())
182 })
183 .detach_and_log_err(cx);
184 },
185 ),
186 }
187 }
188
189 fn is_signed_out(&self) -> bool {
190 self.status.is_signed_out()
191 }
192
193 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
194 let client = self.client.clone();
195 cx.spawn(move |this, mut cx| async move {
196 client.authenticate_and_connect(true, &cx).await?;
197 this.update(&mut cx, |_, cx| cx.notify())
198 })
199 }
200
201 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
202 self.user_store
203 .read(cx)
204 .current_user_has_accepted_terms()
205 .unwrap_or(false)
206 }
207
208 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
209 let user_store = self.user_store.clone();
210 self.accept_terms = Some(cx.spawn(move |this, mut cx| async move {
211 let _ = user_store
212 .update(&mut cx, |store, cx| store.accept_terms_of_service(cx))?
213 .await;
214 this.update(&mut cx, |this, cx| {
215 this.accept_terms = None;
216 cx.notify()
217 })
218 }));
219 }
220}
221
222impl CloudLanguageModelProvider {
223 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
224 let mut status_rx = client.status();
225 let status = *status_rx.borrow();
226
227 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
228
229 let state_ref = state.downgrade();
230 let maintain_client_status = cx.spawn(|mut cx| async move {
231 while let Some(status) = status_rx.next().await {
232 if let Some(this) = state_ref.upgrade() {
233 _ = this.update(&mut cx, |this, cx| {
234 if this.status != status {
235 this.status = status;
236 cx.notify();
237 }
238 });
239 } else {
240 break;
241 }
242 }
243 });
244
245 Self {
246 client,
247 state: state.clone(),
248 _maintain_client_status: maintain_client_status,
249 }
250 }
251}
252
253impl LanguageModelProviderState for CloudLanguageModelProvider {
254 type ObservableEntity = State;
255
256 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
257 Some(self.state.clone())
258 }
259}
260
261impl LanguageModelProvider for CloudLanguageModelProvider {
262 fn id(&self) -> LanguageModelProviderId {
263 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
264 }
265
266 fn name(&self) -> LanguageModelProviderName {
267 LanguageModelProviderName(PROVIDER_NAME.into())
268 }
269
270 fn icon(&self) -> IconName {
271 IconName::AiZed
272 }
273
274 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
275 let mut models = BTreeMap::default();
276
277 if cx.is_staff() {
278 for model in anthropic::Model::iter() {
279 if !matches!(model, anthropic::Model::Custom { .. }) {
280 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
281 }
282 }
283 for model in open_ai::Model::iter() {
284 if !matches!(model, open_ai::Model::Custom { .. }) {
285 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
286 }
287 }
288 for model in google_ai::Model::iter() {
289 if !matches!(model, google_ai::Model::Custom { .. }) {
290 models.insert(model.id().to_string(), CloudModel::Google(model));
291 }
292 }
293 } else {
294 models.insert(
295 anthropic::Model::Claude3_5Sonnet.id().to_string(),
296 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
297 );
298 }
299
300 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBeta>() {
301 zed_cloud_provider_additional_models()
302 } else {
303 &[]
304 };
305
306 // Override with available models from settings
307 for model in AllLanguageModelSettings::get_global(cx)
308 .zed_dot_dev
309 .available_models
310 .iter()
311 .chain(llm_closed_beta_models)
312 .cloned()
313 {
314 let model = match model.provider {
315 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
316 name: model.name.clone(),
317 display_name: model.display_name.clone(),
318 max_tokens: model.max_tokens,
319 tool_override: model.tool_override.clone(),
320 cache_configuration: model.cache_configuration.as_ref().map(|config| {
321 anthropic::AnthropicModelCacheConfiguration {
322 max_cache_anchors: config.max_cache_anchors,
323 should_speculate: config.should_speculate,
324 min_total_token: config.min_total_token,
325 }
326 }),
327 default_temperature: model.default_temperature,
328 max_output_tokens: model.max_output_tokens,
329 extra_beta_headers: model.extra_beta_headers.clone(),
330 }),
331 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
332 name: model.name.clone(),
333 display_name: model.display_name.clone(),
334 max_tokens: model.max_tokens,
335 max_output_tokens: model.max_output_tokens,
336 max_completion_tokens: model.max_completion_tokens,
337 }),
338 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
339 name: model.name.clone(),
340 display_name: model.display_name.clone(),
341 max_tokens: model.max_tokens,
342 }),
343 };
344 models.insert(model.id().to_string(), model.clone());
345 }
346
347 let llm_api_token = self.state.read(cx).llm_api_token.clone();
348 models
349 .into_values()
350 .map(|model| {
351 Arc::new(CloudLanguageModel {
352 id: LanguageModelId::from(model.id().to_string()),
353 model,
354 llm_api_token: llm_api_token.clone(),
355 client: self.client.clone(),
356 request_limiter: RateLimiter::new(4),
357 }) as Arc<dyn LanguageModel>
358 })
359 .collect()
360 }
361
362 fn is_authenticated(&self, cx: &App) -> bool {
363 !self.state.read(cx).is_signed_out()
364 }
365
366 fn authenticate(&self, _cx: &mut App) -> Task<Result<()>> {
367 Task::ready(Ok(()))
368 }
369
370 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
371 cx.new(|_| ConfigurationView {
372 state: self.state.clone(),
373 })
374 .into()
375 }
376
377 fn must_accept_terms(&self, cx: &App) -> bool {
378 !self.state.read(cx).has_accepted_terms_of_service(cx)
379 }
380
381 fn render_accept_terms(
382 &self,
383 view: LanguageModelProviderTosView,
384 cx: &mut App,
385 ) -> Option<AnyElement> {
386 render_accept_terms(self.state.clone(), view, cx)
387 }
388
389 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
390 Task::ready(Ok(()))
391 }
392}
393
394fn render_accept_terms(
395 state: Entity<State>,
396 view_kind: LanguageModelProviderTosView,
397 cx: &mut App,
398) -> Option<AnyElement> {
399 if state.read(cx).has_accepted_terms_of_service(cx) {
400 return None;
401 }
402
403 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
404
405 let terms_button = Button::new("terms_of_service", "Terms of Service")
406 .style(ButtonStyle::Subtle)
407 .icon(IconName::ArrowUpRight)
408 .icon_color(Color::Muted)
409 .icon_size(IconSize::XSmall)
410 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
411
412 let text = "To start using Zed AI, please read and accept the";
413
414 let form = v_flex()
415 .w_full()
416 .gap_2()
417 .when(
418 view_kind == LanguageModelProviderTosView::ThreadEmptyState,
419 |form| form.items_center(),
420 )
421 .child(
422 h_flex()
423 .flex_wrap()
424 .when(
425 view_kind == LanguageModelProviderTosView::ThreadEmptyState,
426 |form| form.justify_center(),
427 )
428 .child(Label::new(text))
429 .child(terms_button),
430 )
431 .child({
432 let button_container = h_flex().w_full().child(
433 Button::new("accept_terms", "I accept the Terms of Service")
434 .style(ButtonStyle::Tinted(TintColor::Accent))
435 .disabled(accept_terms_disabled)
436 .on_click({
437 let state = state.downgrade();
438 move |_, _window, cx| {
439 state
440 .update(cx, |state, cx| state.accept_terms_of_service(cx))
441 .ok();
442 }
443 }),
444 );
445
446 match view_kind {
447 LanguageModelProviderTosView::ThreadEmptyState => button_container.justify_center(),
448 LanguageModelProviderTosView::PromptEditorPopup => button_container.justify_end(),
449 LanguageModelProviderTosView::Configuration => button_container.justify_start(),
450 }
451 });
452
453 Some(form.into_any())
454}
455
456pub struct CloudLanguageModel {
457 id: LanguageModelId,
458 model: CloudModel,
459 llm_api_token: LlmApiToken,
460 client: Arc<Client>,
461 request_limiter: RateLimiter,
462}
463
464#[derive(Clone, Default)]
465pub struct LlmApiToken(Arc<RwLock<Option<String>>>);
466
467#[derive(Error, Debug)]
468pub struct PaymentRequiredError;
469
470impl fmt::Display for PaymentRequiredError {
471 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
472 write!(
473 f,
474 "Payment required to use this language model. Please upgrade your account."
475 )
476 }
477}
478
479#[derive(Error, Debug)]
480pub struct MaxMonthlySpendReachedError;
481
482impl fmt::Display for MaxMonthlySpendReachedError {
483 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
484 write!(
485 f,
486 "Maximum spending limit reached for this month. For more usage, increase your spending limit."
487 )
488 }
489}
490
491impl CloudLanguageModel {
492 async fn perform_llm_completion(
493 client: Arc<Client>,
494 llm_api_token: LlmApiToken,
495 body: PerformCompletionParams,
496 ) -> Result<Response<AsyncBody>> {
497 let http_client = &client.http_client();
498
499 let mut token = llm_api_token.acquire(&client).await?;
500 let mut did_retry = false;
501
502 let response = loop {
503 let request_builder = http_client::Request::builder();
504 let request = request_builder
505 .method(Method::POST)
506 .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
507 .header("Content-Type", "application/json")
508 .header("Authorization", format!("Bearer {token}"))
509 .body(serde_json::to_string(&body)?.into())?;
510 let mut response = http_client.send(request).await?;
511 if response.status().is_success() {
512 break response;
513 } else if !did_retry
514 && response
515 .headers()
516 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
517 .is_some()
518 {
519 did_retry = true;
520 token = llm_api_token.refresh(&client).await?;
521 } else if response.status() == StatusCode::FORBIDDEN
522 && response
523 .headers()
524 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
525 .is_some()
526 {
527 break Err(anyhow!(MaxMonthlySpendReachedError))?;
528 } else if response.status() == StatusCode::PAYMENT_REQUIRED {
529 break Err(anyhow!(PaymentRequiredError))?;
530 } else {
531 let mut body = String::new();
532 response.body_mut().read_to_string(&mut body).await?;
533 break Err(anyhow!(
534 "cloud language model completion failed with status {}: {body}",
535 response.status()
536 ))?;
537 }
538 };
539
540 Ok(response)
541 }
542}
543
544impl LanguageModel for CloudLanguageModel {
545 fn id(&self) -> LanguageModelId {
546 self.id.clone()
547 }
548
549 fn name(&self) -> LanguageModelName {
550 LanguageModelName::from(self.model.display_name().to_string())
551 }
552
553 fn icon(&self) -> Option<IconName> {
554 self.model.icon()
555 }
556
557 fn provider_id(&self) -> LanguageModelProviderId {
558 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
559 }
560
561 fn provider_name(&self) -> LanguageModelProviderName {
562 LanguageModelProviderName(PROVIDER_NAME.into())
563 }
564
565 fn telemetry_id(&self) -> String {
566 format!("zed.dev/{}", self.model.id())
567 }
568
569 fn availability(&self) -> LanguageModelAvailability {
570 self.model.availability()
571 }
572
573 fn max_token_count(&self) -> usize {
574 self.model.max_token_count()
575 }
576
577 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
578 match &self.model {
579 CloudModel::Anthropic(model) => {
580 model
581 .cache_configuration()
582 .map(|cache| LanguageModelCacheConfiguration {
583 max_cache_anchors: cache.max_cache_anchors,
584 should_speculate: cache.should_speculate,
585 min_total_token: cache.min_total_token,
586 })
587 }
588 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
589 }
590 }
591
592 fn count_tokens(
593 &self,
594 request: LanguageModelRequest,
595 cx: &App,
596 ) -> BoxFuture<'static, Result<usize>> {
597 match self.model.clone() {
598 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
599 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
600 CloudModel::Google(model) => {
601 let client = self.client.clone();
602 let request = request.into_google(model.id().into());
603 let request = google_ai::CountTokensRequest {
604 contents: request.contents,
605 };
606 async move {
607 let request = serde_json::to_string(&request)?;
608 let response = client
609 .request(proto::CountLanguageModelTokens {
610 provider: proto::LanguageModelProvider::Google as i32,
611 request,
612 })
613 .await?;
614 Ok(response.token_count as usize)
615 }
616 .boxed()
617 }
618 }
619 }
620
621 fn stream_completion(
622 &self,
623 request: LanguageModelRequest,
624 _cx: &AsyncApp,
625 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
626 match &self.model {
627 CloudModel::Anthropic(model) => {
628 let request = request.into_anthropic(
629 model.id().into(),
630 model.default_temperature(),
631 model.max_output_tokens(),
632 );
633 let client = self.client.clone();
634 let llm_api_token = self.llm_api_token.clone();
635 let future = self.request_limiter.stream(async move {
636 let response = Self::perform_llm_completion(
637 client.clone(),
638 llm_api_token,
639 PerformCompletionParams {
640 provider: client::LanguageModelProvider::Anthropic,
641 model: request.model.clone(),
642 provider_request: RawValue::from_string(serde_json::to_string(
643 &request,
644 )?)?,
645 },
646 )
647 .await?;
648 Ok(map_to_language_model_completion_events(Box::pin(
649 response_lines(response).map_err(AnthropicError::Other),
650 )))
651 });
652 async move { Ok(future.await?.boxed()) }.boxed()
653 }
654 CloudModel::OpenAi(model) => {
655 let client = self.client.clone();
656 let request = request.into_open_ai(model.id().into(), model.max_output_tokens());
657 let llm_api_token = self.llm_api_token.clone();
658 let future = self.request_limiter.stream(async move {
659 let response = Self::perform_llm_completion(
660 client.clone(),
661 llm_api_token,
662 PerformCompletionParams {
663 provider: client::LanguageModelProvider::OpenAi,
664 model: request.model.clone(),
665 provider_request: RawValue::from_string(serde_json::to_string(
666 &request,
667 )?)?,
668 },
669 )
670 .await?;
671 Ok(open_ai::extract_text_from_events(response_lines(response)))
672 });
673 async move {
674 Ok(future
675 .await?
676 .map(|result| result.map(LanguageModelCompletionEvent::Text))
677 .boxed())
678 }
679 .boxed()
680 }
681 CloudModel::Google(model) => {
682 let client = self.client.clone();
683 let request = request.into_google(model.id().into());
684 let llm_api_token = self.llm_api_token.clone();
685 let future = self.request_limiter.stream(async move {
686 let response = Self::perform_llm_completion(
687 client.clone(),
688 llm_api_token,
689 PerformCompletionParams {
690 provider: client::LanguageModelProvider::Google,
691 model: request.model.clone(),
692 provider_request: RawValue::from_string(serde_json::to_string(
693 &request,
694 )?)?,
695 },
696 )
697 .await?;
698 Ok(google_ai::extract_text_from_events(response_lines(
699 response,
700 )))
701 });
702 async move {
703 Ok(future
704 .await?
705 .map(|result| result.map(LanguageModelCompletionEvent::Text))
706 .boxed())
707 }
708 .boxed()
709 }
710 }
711 }
712
713 fn use_any_tool(
714 &self,
715 request: LanguageModelRequest,
716 tool_name: String,
717 tool_description: String,
718 input_schema: serde_json::Value,
719 _cx: &AsyncApp,
720 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
721 let client = self.client.clone();
722 let llm_api_token = self.llm_api_token.clone();
723
724 match &self.model {
725 CloudModel::Anthropic(model) => {
726 let mut request = request.into_anthropic(
727 model.tool_model_id().into(),
728 model.default_temperature(),
729 model.max_output_tokens(),
730 );
731 request.tool_choice = Some(anthropic::ToolChoice::Tool {
732 name: tool_name.clone(),
733 });
734 request.tools = vec![anthropic::Tool {
735 name: tool_name.clone(),
736 description: tool_description,
737 input_schema,
738 }];
739
740 self.request_limiter
741 .run(async move {
742 let response = Self::perform_llm_completion(
743 client.clone(),
744 llm_api_token,
745 PerformCompletionParams {
746 provider: client::LanguageModelProvider::Anthropic,
747 model: request.model.clone(),
748 provider_request: RawValue::from_string(serde_json::to_string(
749 &request,
750 )?)?,
751 },
752 )
753 .await?;
754
755 Ok(anthropic::extract_tool_args_from_events(
756 tool_name,
757 Box::pin(response_lines(response)),
758 )
759 .await?
760 .boxed())
761 })
762 .boxed()
763 }
764 CloudModel::OpenAi(model) => {
765 let mut request =
766 request.into_open_ai(model.id().into(), model.max_output_tokens());
767 request.tool_choice = Some(open_ai::ToolChoice::Other(
768 open_ai::ToolDefinition::Function {
769 function: open_ai::FunctionDefinition {
770 name: tool_name.clone(),
771 description: None,
772 parameters: None,
773 },
774 },
775 ));
776 request.tools = vec![open_ai::ToolDefinition::Function {
777 function: open_ai::FunctionDefinition {
778 name: tool_name.clone(),
779 description: Some(tool_description),
780 parameters: Some(input_schema),
781 },
782 }];
783
784 self.request_limiter
785 .run(async move {
786 let response = Self::perform_llm_completion(
787 client.clone(),
788 llm_api_token,
789 PerformCompletionParams {
790 provider: client::LanguageModelProvider::OpenAi,
791 model: request.model.clone(),
792 provider_request: RawValue::from_string(serde_json::to_string(
793 &request,
794 )?)?,
795 },
796 )
797 .await?;
798
799 Ok(open_ai::extract_tool_args_from_events(
800 tool_name,
801 Box::pin(response_lines(response)),
802 )
803 .await?
804 .boxed())
805 })
806 .boxed()
807 }
808 CloudModel::Google(_) => {
809 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
810 }
811 }
812 }
813}
814
815fn response_lines<T: DeserializeOwned>(
816 response: Response<AsyncBody>,
817) -> impl Stream<Item = Result<T>> {
818 futures::stream::try_unfold(
819 (String::new(), BufReader::new(response.into_body())),
820 move |(mut line, mut body)| async {
821 match body.read_line(&mut line).await {
822 Ok(0) => Ok(None),
823 Ok(_) => {
824 let event: T = serde_json::from_str(&line)?;
825 line.clear();
826 Ok(Some((event, (line, body))))
827 }
828 Err(e) => Err(e.into()),
829 }
830 },
831 )
832}
833
834impl LlmApiToken {
835 pub async fn acquire(&self, client: &Arc<Client>) -> Result<String> {
836 let lock = self.0.upgradable_read().await;
837 if let Some(token) = lock.as_ref() {
838 Ok(token.to_string())
839 } else {
840 Self::fetch(RwLockUpgradableReadGuard::upgrade(lock).await, client).await
841 }
842 }
843
844 pub async fn refresh(&self, client: &Arc<Client>) -> Result<String> {
845 Self::fetch(self.0.write().await, client).await
846 }
847
848 async fn fetch<'a>(
849 mut lock: RwLockWriteGuard<'a, Option<String>>,
850 client: &Arc<Client>,
851 ) -> Result<String> {
852 let response = client.request(proto::GetLlmToken {}).await?;
853 *lock = Some(response.token.clone());
854 Ok(response.token.clone())
855 }
856}
857
858struct ConfigurationView {
859 state: gpui::Entity<State>,
860}
861
862impl ConfigurationView {
863 fn authenticate(&mut self, cx: &mut Context<Self>) {
864 self.state.update(cx, |state, cx| {
865 state.authenticate(cx).detach_and_log_err(cx);
866 });
867 cx.notify();
868 }
869}
870
871impl Render for ConfigurationView {
872 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
873 const ZED_AI_URL: &str = "https://zed.dev/ai";
874
875 let is_connected = !self.state.read(cx).is_signed_out();
876 let plan = self.state.read(cx).user_store.read(cx).current_plan();
877 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
878
879 let is_pro = plan == Some(proto::Plan::ZedPro);
880 let subscription_text = Label::new(if is_pro {
881 "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."
882 } else {
883 "You have basic access to models from Anthropic through the Zed AI Free plan."
884 });
885 let manage_subscription_button = if is_pro {
886 Some(
887 h_flex().child(
888 Button::new("manage_settings", "Manage Subscription")
889 .style(ButtonStyle::Tinted(TintColor::Accent))
890 .on_click(
891 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
892 ),
893 ),
894 )
895 } else if cx.has_flag::<ZedPro>() {
896 Some(
897 h_flex()
898 .gap_2()
899 .child(
900 Button::new("learn_more", "Learn more")
901 .style(ButtonStyle::Subtle)
902 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
903 )
904 .child(
905 Button::new("upgrade", "Upgrade")
906 .style(ButtonStyle::Subtle)
907 .color(Color::Accent)
908 .on_click(
909 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
910 ),
911 ),
912 )
913 } else {
914 None
915 };
916
917 if is_connected {
918 v_flex()
919 .gap_3()
920 .w_full()
921 .children(render_accept_terms(
922 self.state.clone(),
923 LanguageModelProviderTosView::Configuration,
924 cx,
925 ))
926 .when(has_accepted_terms, |this| {
927 this.child(subscription_text)
928 .children(manage_subscription_button)
929 })
930 } else {
931 v_flex()
932 .gap_2()
933 .child(Label::new("Use Zed AI to access hosted language models."))
934 .child(
935 Button::new("sign_in", "Sign In")
936 .icon_color(Color::Muted)
937 .icon(IconName::Github)
938 .icon_position(IconPosition::Start)
939 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
940 )
941 }
942 }
943}