1use super::open_ai::count_open_ai_tokens;
2use crate::provider::anthropic::map_to_language_model_completion_events;
3use crate::{
4 settings::AllLanguageModelSettings, CloudModel, LanguageModel, LanguageModelCacheConfiguration,
5 LanguageModelId, LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,
6 LanguageModelProviderState, LanguageModelRequest, RateLimiter,
7};
8use anthropic::AnthropicError;
9use anyhow::{anyhow, Result};
10use client::{
11 zed_urls, Client, PerformCompletionParams, UserStore, EXPIRED_LLM_TOKEN_HEADER_NAME,
12 MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
13};
14use collections::BTreeMap;
15use feature_flags::{FeatureFlagAppExt, LlmClosedBeta, ZedPro};
16use futures::{
17 future::BoxFuture, stream::BoxStream, AsyncBufReadExt, FutureExt, Stream, StreamExt,
18 TryStreamExt as _,
19};
20use gpui::{
21 AnyElement, AnyView, AppContext, AsyncAppContext, EventEmitter, FontWeight, Global, Model,
22 ModelContext, ReadGlobal, Subscription, Task,
23};
24use http_client::{AsyncBody, HttpClient, HttpRequestExt, Method, Response, StatusCode};
25use proto::TypedEnvelope;
26use schemars::JsonSchema;
27use serde::{de::DeserializeOwned, Deserialize, Serialize};
28use serde_json::value::RawValue;
29use settings::{Settings, SettingsStore};
30use smol::{
31 io::{AsyncReadExt, BufReader},
32 lock::{RwLock, RwLockUpgradableReadGuard, RwLockWriteGuard},
33};
34use std::fmt;
35use std::time::Duration;
36use std::{
37 future,
38 sync::{Arc, LazyLock},
39};
40use strum::IntoEnumIterator;
41use thiserror::Error;
42use ui::{prelude::*, TintColor};
43
44use crate::{LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider};
45
46use super::anthropic::count_anthropic_tokens;
47
48pub const PROVIDER_ID: &str = "zed.dev";
49pub const PROVIDER_NAME: &str = "Zed";
50
51const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
52 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
53
54fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
55 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
56 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
57 .map(|json| serde_json::from_str(json).unwrap())
58 .unwrap_or_default()
59 });
60 ADDITIONAL_MODELS.as_slice()
61}
62
63#[derive(Default, Clone, Debug, PartialEq)]
64pub struct ZedDotDevSettings {
65 pub available_models: Vec<AvailableModel>,
66 pub low_speed_timeout: Option<Duration>,
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}
98
99struct GlobalRefreshLlmTokenListener(Model<RefreshLlmTokenListener>);
100
101impl Global for GlobalRefreshLlmTokenListener {}
102
103pub struct RefreshLlmTokenEvent;
104
105pub struct RefreshLlmTokenListener {
106 _llm_token_subscription: client::Subscription,
107}
108
109impl EventEmitter<RefreshLlmTokenEvent> for RefreshLlmTokenListener {}
110
111impl RefreshLlmTokenListener {
112 pub fn register(client: Arc<Client>, cx: &mut AppContext) {
113 let listener = cx.new_model(|cx| RefreshLlmTokenListener::new(client, cx));
114 cx.set_global(GlobalRefreshLlmTokenListener(listener));
115 }
116
117 pub fn global(cx: &AppContext) -> Model<Self> {
118 GlobalRefreshLlmTokenListener::global(cx).0.clone()
119 }
120
121 fn new(client: Arc<Client>, cx: &mut ModelContext<Self>) -> Self {
122 Self {
123 _llm_token_subscription: client
124 .add_message_handler(cx.weak_model(), Self::handle_refresh_llm_token),
125 }
126 }
127
128 async fn handle_refresh_llm_token(
129 this: Model<Self>,
130 _: TypedEnvelope<proto::RefreshLlmToken>,
131 mut cx: AsyncAppContext,
132 ) -> Result<()> {
133 this.update(&mut cx, |_this, cx| cx.emit(RefreshLlmTokenEvent))
134 }
135}
136
137pub struct CloudLanguageModelProvider {
138 client: Arc<Client>,
139 state: gpui::Model<State>,
140 _maintain_client_status: Task<()>,
141}
142
143pub struct State {
144 client: Arc<Client>,
145 llm_api_token: LlmApiToken,
146 user_store: Model<UserStore>,
147 status: client::Status,
148 accept_terms: Option<Task<Result<()>>>,
149 _settings_subscription: Subscription,
150 _llm_token_subscription: Subscription,
151}
152
153impl State {
154 fn new(
155 client: Arc<Client>,
156 user_store: Model<UserStore>,
157 status: client::Status,
158 cx: &mut ModelContext<Self>,
159 ) -> Self {
160 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
161
162 Self {
163 client: client.clone(),
164 llm_api_token: LlmApiToken::default(),
165 user_store,
166 status,
167 accept_terms: None,
168 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
169 cx.notify();
170 }),
171 _llm_token_subscription: cx.subscribe(
172 &refresh_llm_token_listener,
173 |this, _listener, _event, cx| {
174 let client = this.client.clone();
175 let llm_api_token = this.llm_api_token.clone();
176 cx.spawn(|_this, _cx| async move {
177 llm_api_token.refresh(&client).await?;
178 anyhow::Ok(())
179 })
180 .detach_and_log_err(cx);
181 },
182 ),
183 }
184 }
185
186 fn is_signed_out(&self) -> bool {
187 self.status.is_signed_out()
188 }
189
190 fn authenticate(&self, cx: &mut ModelContext<Self>) -> Task<Result<()>> {
191 let client = self.client.clone();
192 cx.spawn(move |this, mut cx| async move {
193 client.authenticate_and_connect(true, &cx).await?;
194 this.update(&mut cx, |_, cx| cx.notify())
195 })
196 }
197
198 fn has_accepted_terms_of_service(&self, cx: &AppContext) -> bool {
199 self.user_store
200 .read(cx)
201 .current_user_has_accepted_terms()
202 .unwrap_or(false)
203 }
204
205 fn accept_terms_of_service(&mut self, cx: &mut ModelContext<Self>) {
206 let user_store = self.user_store.clone();
207 self.accept_terms = Some(cx.spawn(move |this, mut cx| async move {
208 let _ = user_store
209 .update(&mut cx, |store, cx| store.accept_terms_of_service(cx))?
210 .await;
211 this.update(&mut cx, |this, cx| {
212 this.accept_terms = None;
213 cx.notify()
214 })
215 }));
216 }
217}
218
219impl CloudLanguageModelProvider {
220 pub fn new(user_store: Model<UserStore>, client: Arc<Client>, cx: &mut AppContext) -> Self {
221 let mut status_rx = client.status();
222 let status = *status_rx.borrow();
223
224 let state = cx.new_model(|cx| State::new(client.clone(), user_store.clone(), status, cx));
225
226 let state_ref = state.downgrade();
227 let maintain_client_status = cx.spawn(|mut cx| async move {
228 while let Some(status) = status_rx.next().await {
229 if let Some(this) = state_ref.upgrade() {
230 _ = this.update(&mut cx, |this, cx| {
231 if this.status != status {
232 this.status = status;
233 cx.notify();
234 }
235 });
236 } else {
237 break;
238 }
239 }
240 });
241
242 Self {
243 client,
244 state: state.clone(),
245 _maintain_client_status: maintain_client_status,
246 }
247 }
248}
249
250impl LanguageModelProviderState for CloudLanguageModelProvider {
251 type ObservableEntity = State;
252
253 fn observable_entity(&self) -> Option<gpui::Model<Self::ObservableEntity>> {
254 Some(self.state.clone())
255 }
256}
257
258impl LanguageModelProvider for CloudLanguageModelProvider {
259 fn id(&self) -> LanguageModelProviderId {
260 LanguageModelProviderId(PROVIDER_ID.into())
261 }
262
263 fn name(&self) -> LanguageModelProviderName {
264 LanguageModelProviderName(PROVIDER_NAME.into())
265 }
266
267 fn icon(&self) -> IconName {
268 IconName::AiZed
269 }
270
271 fn provided_models(&self, cx: &AppContext) -> Vec<Arc<dyn LanguageModel>> {
272 let mut models = BTreeMap::default();
273
274 if cx.is_staff() {
275 for model in anthropic::Model::iter() {
276 if !matches!(model, anthropic::Model::Custom { .. }) {
277 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
278 }
279 }
280 for model in open_ai::Model::iter() {
281 if !matches!(model, open_ai::Model::Custom { .. }) {
282 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
283 }
284 }
285 for model in google_ai::Model::iter() {
286 if !matches!(model, google_ai::Model::Custom { .. }) {
287 models.insert(model.id().to_string(), CloudModel::Google(model));
288 }
289 }
290 } else {
291 models.insert(
292 anthropic::Model::Claude3_5Sonnet.id().to_string(),
293 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
294 );
295 }
296
297 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBeta>() {
298 zed_cloud_provider_additional_models()
299 } else {
300 &[]
301 };
302
303 // Override with available models from settings
304 for model in AllLanguageModelSettings::get_global(cx)
305 .zed_dot_dev
306 .available_models
307 .iter()
308 .chain(llm_closed_beta_models)
309 .cloned()
310 {
311 let model = match model.provider {
312 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
313 name: model.name.clone(),
314 display_name: model.display_name.clone(),
315 max_tokens: model.max_tokens,
316 tool_override: model.tool_override.clone(),
317 cache_configuration: model.cache_configuration.as_ref().map(|config| {
318 anthropic::AnthropicModelCacheConfiguration {
319 max_cache_anchors: config.max_cache_anchors,
320 should_speculate: config.should_speculate,
321 min_total_token: config.min_total_token,
322 }
323 }),
324 default_temperature: model.default_temperature,
325 max_output_tokens: model.max_output_tokens,
326 }),
327 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
328 name: model.name.clone(),
329 display_name: model.display_name.clone(),
330 max_tokens: model.max_tokens,
331 max_output_tokens: model.max_output_tokens,
332 max_completion_tokens: model.max_completion_tokens,
333 }),
334 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
335 name: model.name.clone(),
336 display_name: model.display_name.clone(),
337 max_tokens: model.max_tokens,
338 }),
339 };
340 models.insert(model.id().to_string(), model.clone());
341 }
342
343 let llm_api_token = self.state.read(cx).llm_api_token.clone();
344 models
345 .into_values()
346 .map(|model| {
347 Arc::new(CloudLanguageModel {
348 id: LanguageModelId::from(model.id().to_string()),
349 model,
350 llm_api_token: llm_api_token.clone(),
351 client: self.client.clone(),
352 request_limiter: RateLimiter::new(4),
353 }) as Arc<dyn LanguageModel>
354 })
355 .collect()
356 }
357
358 fn is_authenticated(&self, cx: &AppContext) -> bool {
359 !self.state.read(cx).is_signed_out()
360 }
361
362 fn authenticate(&self, _cx: &mut AppContext) -> Task<Result<()>> {
363 Task::ready(Ok(()))
364 }
365
366 fn configuration_view(&self, cx: &mut WindowContext) -> AnyView {
367 cx.new_view(|_cx| ConfigurationView {
368 state: self.state.clone(),
369 })
370 .into()
371 }
372
373 fn must_accept_terms(&self, cx: &AppContext) -> bool {
374 !self.state.read(cx).has_accepted_terms_of_service(cx)
375 }
376
377 fn render_accept_terms(&self, cx: &mut WindowContext) -> Option<AnyElement> {
378 let state = self.state.read(cx);
379
380 let terms = [(
381 "terms_of_service",
382 "Terms of Service",
383 "https://zed.dev/terms-of-service",
384 )]
385 .map(|(id, label, url)| {
386 Button::new(id, label)
387 .style(ButtonStyle::Subtle)
388 .icon(IconName::ExternalLink)
389 .icon_size(IconSize::XSmall)
390 .icon_color(Color::Muted)
391 .on_click(move |_, cx| cx.open_url(url))
392 });
393
394 if state.has_accepted_terms_of_service(cx) {
395 None
396 } else {
397 let disabled = state.accept_terms.is_some();
398 Some(
399 v_flex()
400 .gap_2()
401 .child(
402 v_flex()
403 .child(Label::new("Terms and Conditions").weight(FontWeight::MEDIUM))
404 .child(
405 Label::new(
406 "Please read and accept our terms and conditions to continue.",
407 )
408 .size(LabelSize::Small),
409 ),
410 )
411 .child(v_flex().gap_1().children(terms))
412 .child(
413 h_flex().justify_end().child(
414 Button::new("accept_terms", "I've read it and accept it")
415 .disabled(disabled)
416 .on_click({
417 let state = self.state.downgrade();
418 move |_, cx| {
419 state
420 .update(cx, |state, cx| {
421 state.accept_terms_of_service(cx)
422 })
423 .ok();
424 }
425 }),
426 ),
427 )
428 .into_any(),
429 )
430 }
431 }
432
433 fn reset_credentials(&self, _cx: &mut AppContext) -> Task<Result<()>> {
434 Task::ready(Ok(()))
435 }
436}
437
438pub struct CloudLanguageModel {
439 id: LanguageModelId,
440 model: CloudModel,
441 llm_api_token: LlmApiToken,
442 client: Arc<Client>,
443 request_limiter: RateLimiter,
444}
445
446#[derive(Clone, Default)]
447struct LlmApiToken(Arc<RwLock<Option<String>>>);
448
449#[derive(Error, Debug)]
450pub struct PaymentRequiredError;
451
452impl fmt::Display for PaymentRequiredError {
453 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
454 write!(
455 f,
456 "Payment required to use this language model. Please upgrade your account."
457 )
458 }
459}
460
461#[derive(Error, Debug)]
462pub struct MaxMonthlySpendReachedError;
463
464impl fmt::Display for MaxMonthlySpendReachedError {
465 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
466 write!(
467 f,
468 "Maximum spending limit reached for this month. For more usage, increase your spending limit."
469 )
470 }
471}
472
473impl CloudLanguageModel {
474 async fn perform_llm_completion(
475 client: Arc<Client>,
476 llm_api_token: LlmApiToken,
477 body: PerformCompletionParams,
478 low_speed_timeout: Option<Duration>,
479 ) -> Result<Response<AsyncBody>> {
480 let http_client = &client.http_client();
481
482 let mut token = llm_api_token.acquire(&client).await?;
483 let mut did_retry = false;
484
485 let response = loop {
486 let mut request_builder = http_client::Request::builder();
487 if let Some(low_speed_timeout) = low_speed_timeout {
488 request_builder = request_builder.read_timeout(low_speed_timeout);
489 };
490 let request = request_builder
491 .method(Method::POST)
492 .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
493 .header("Content-Type", "application/json")
494 .header("Authorization", format!("Bearer {token}"))
495 .body(serde_json::to_string(&body)?.into())?;
496 let mut response = http_client.send(request).await?;
497 if response.status().is_success() {
498 break response;
499 } else if !did_retry
500 && response
501 .headers()
502 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
503 .is_some()
504 {
505 did_retry = true;
506 token = llm_api_token.refresh(&client).await?;
507 } else if response.status() == StatusCode::FORBIDDEN
508 && response
509 .headers()
510 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
511 .is_some()
512 {
513 break Err(anyhow!(MaxMonthlySpendReachedError))?;
514 } else if response.status() == StatusCode::PAYMENT_REQUIRED {
515 break Err(anyhow!(PaymentRequiredError))?;
516 } else {
517 let mut body = String::new();
518 response.body_mut().read_to_string(&mut body).await?;
519 break Err(anyhow!(
520 "cloud language model completion failed with status {}: {body}",
521 response.status()
522 ))?;
523 }
524 };
525
526 Ok(response)
527 }
528}
529
530impl LanguageModel for CloudLanguageModel {
531 fn id(&self) -> LanguageModelId {
532 self.id.clone()
533 }
534
535 fn name(&self) -> LanguageModelName {
536 LanguageModelName::from(self.model.display_name().to_string())
537 }
538
539 fn icon(&self) -> Option<IconName> {
540 self.model.icon()
541 }
542
543 fn provider_id(&self) -> LanguageModelProviderId {
544 LanguageModelProviderId(PROVIDER_ID.into())
545 }
546
547 fn provider_name(&self) -> LanguageModelProviderName {
548 LanguageModelProviderName(PROVIDER_NAME.into())
549 }
550
551 fn telemetry_id(&self) -> String {
552 format!("zed.dev/{}", self.model.id())
553 }
554
555 fn availability(&self) -> LanguageModelAvailability {
556 self.model.availability()
557 }
558
559 fn max_token_count(&self) -> usize {
560 self.model.max_token_count()
561 }
562
563 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
564 match &self.model {
565 CloudModel::Anthropic(model) => {
566 model
567 .cache_configuration()
568 .map(|cache| LanguageModelCacheConfiguration {
569 max_cache_anchors: cache.max_cache_anchors,
570 should_speculate: cache.should_speculate,
571 min_total_token: cache.min_total_token,
572 })
573 }
574 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
575 }
576 }
577
578 fn count_tokens(
579 &self,
580 request: LanguageModelRequest,
581 cx: &AppContext,
582 ) -> BoxFuture<'static, Result<usize>> {
583 match self.model.clone() {
584 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
585 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
586 CloudModel::Google(model) => {
587 let client = self.client.clone();
588 let request = request.into_google(model.id().into());
589 let request = google_ai::CountTokensRequest {
590 contents: request.contents,
591 };
592 async move {
593 let request = serde_json::to_string(&request)?;
594 let response = client
595 .request(proto::CountLanguageModelTokens {
596 provider: proto::LanguageModelProvider::Google as i32,
597 request,
598 })
599 .await?;
600 Ok(response.token_count as usize)
601 }
602 .boxed()
603 }
604 }
605 }
606
607 fn stream_completion(
608 &self,
609 request: LanguageModelRequest,
610 cx: &AsyncAppContext,
611 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
612 let openai_low_speed_timeout =
613 AllLanguageModelSettings::try_read_global(cx, |s| s.openai.low_speed_timeout.unwrap());
614
615 match &self.model {
616 CloudModel::Anthropic(model) => {
617 let request = request.into_anthropic(
618 model.id().into(),
619 model.default_temperature(),
620 model.max_output_tokens(),
621 );
622 let client = self.client.clone();
623 let llm_api_token = self.llm_api_token.clone();
624 let future = self.request_limiter.stream(async move {
625 let response = Self::perform_llm_completion(
626 client.clone(),
627 llm_api_token,
628 PerformCompletionParams {
629 provider: client::LanguageModelProvider::Anthropic,
630 model: request.model.clone(),
631 provider_request: RawValue::from_string(serde_json::to_string(
632 &request,
633 )?)?,
634 },
635 None,
636 )
637 .await?;
638 Ok(map_to_language_model_completion_events(Box::pin(
639 response_lines(response).map_err(AnthropicError::Other),
640 )))
641 });
642 async move { Ok(future.await?.boxed()) }.boxed()
643 }
644 CloudModel::OpenAi(model) => {
645 let client = self.client.clone();
646 let request = request.into_open_ai(model.id().into(), model.max_output_tokens());
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::OpenAi,
654 model: request.model.clone(),
655 provider_request: RawValue::from_string(serde_json::to_string(
656 &request,
657 )?)?,
658 },
659 openai_low_speed_timeout,
660 )
661 .await?;
662 Ok(open_ai::extract_text_from_events(response_lines(response)))
663 });
664 async move {
665 Ok(future
666 .await?
667 .map(|result| result.map(LanguageModelCompletionEvent::Text))
668 .boxed())
669 }
670 .boxed()
671 }
672 CloudModel::Google(model) => {
673 let client = self.client.clone();
674 let request = request.into_google(model.id().into());
675 let llm_api_token = self.llm_api_token.clone();
676 let future = self.request_limiter.stream(async move {
677 let response = Self::perform_llm_completion(
678 client.clone(),
679 llm_api_token,
680 PerformCompletionParams {
681 provider: client::LanguageModelProvider::Google,
682 model: request.model.clone(),
683 provider_request: RawValue::from_string(serde_json::to_string(
684 &request,
685 )?)?,
686 },
687 None,
688 )
689 .await?;
690 Ok(google_ai::extract_text_from_events(response_lines(
691 response,
692 )))
693 });
694 async move {
695 Ok(future
696 .await?
697 .map(|result| result.map(LanguageModelCompletionEvent::Text))
698 .boxed())
699 }
700 .boxed()
701 }
702 }
703 }
704
705 fn use_any_tool(
706 &self,
707 request: LanguageModelRequest,
708 tool_name: String,
709 tool_description: String,
710 input_schema: serde_json::Value,
711 _cx: &AsyncAppContext,
712 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
713 let client = self.client.clone();
714 let llm_api_token = self.llm_api_token.clone();
715
716 match &self.model {
717 CloudModel::Anthropic(model) => {
718 let mut request = request.into_anthropic(
719 model.tool_model_id().into(),
720 model.default_temperature(),
721 model.max_output_tokens(),
722 );
723 request.tool_choice = Some(anthropic::ToolChoice::Tool {
724 name: tool_name.clone(),
725 });
726 request.tools = vec![anthropic::Tool {
727 name: tool_name.clone(),
728 description: tool_description,
729 input_schema,
730 }];
731
732 self.request_limiter
733 .run(async move {
734 let response = Self::perform_llm_completion(
735 client.clone(),
736 llm_api_token,
737 PerformCompletionParams {
738 provider: client::LanguageModelProvider::Anthropic,
739 model: request.model.clone(),
740 provider_request: RawValue::from_string(serde_json::to_string(
741 &request,
742 )?)?,
743 },
744 None,
745 )
746 .await?;
747
748 Ok(anthropic::extract_tool_args_from_events(
749 tool_name,
750 Box::pin(response_lines(response)),
751 )
752 .await?
753 .boxed())
754 })
755 .boxed()
756 }
757 CloudModel::OpenAi(model) => {
758 let mut request =
759 request.into_open_ai(model.id().into(), model.max_output_tokens());
760 request.tool_choice = Some(open_ai::ToolChoice::Other(
761 open_ai::ToolDefinition::Function {
762 function: open_ai::FunctionDefinition {
763 name: tool_name.clone(),
764 description: None,
765 parameters: None,
766 },
767 },
768 ));
769 request.tools = vec![open_ai::ToolDefinition::Function {
770 function: open_ai::FunctionDefinition {
771 name: tool_name.clone(),
772 description: Some(tool_description),
773 parameters: Some(input_schema),
774 },
775 }];
776
777 self.request_limiter
778 .run(async move {
779 let response = Self::perform_llm_completion(
780 client.clone(),
781 llm_api_token,
782 PerformCompletionParams {
783 provider: client::LanguageModelProvider::OpenAi,
784 model: request.model.clone(),
785 provider_request: RawValue::from_string(serde_json::to_string(
786 &request,
787 )?)?,
788 },
789 None,
790 )
791 .await?;
792
793 Ok(open_ai::extract_tool_args_from_events(
794 tool_name,
795 Box::pin(response_lines(response)),
796 )
797 .await?
798 .boxed())
799 })
800 .boxed()
801 }
802 CloudModel::Google(_) => {
803 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
804 }
805 }
806 }
807}
808
809fn response_lines<T: DeserializeOwned>(
810 response: Response<AsyncBody>,
811) -> impl Stream<Item = Result<T>> {
812 futures::stream::try_unfold(
813 (String::new(), BufReader::new(response.into_body())),
814 move |(mut line, mut body)| async {
815 match body.read_line(&mut line).await {
816 Ok(0) => Ok(None),
817 Ok(_) => {
818 let event: T = serde_json::from_str(&line)?;
819 line.clear();
820 Ok(Some((event, (line, body))))
821 }
822 Err(e) => Err(e.into()),
823 }
824 },
825 )
826}
827
828impl LlmApiToken {
829 async fn acquire(&self, client: &Arc<Client>) -> Result<String> {
830 let lock = self.0.upgradable_read().await;
831 if let Some(token) = lock.as_ref() {
832 Ok(token.to_string())
833 } else {
834 Self::fetch(RwLockUpgradableReadGuard::upgrade(lock).await, client).await
835 }
836 }
837
838 async fn refresh(&self, client: &Arc<Client>) -> Result<String> {
839 Self::fetch(self.0.write().await, client).await
840 }
841
842 async fn fetch<'a>(
843 mut lock: RwLockWriteGuard<'a, Option<String>>,
844 client: &Arc<Client>,
845 ) -> Result<String> {
846 let response = client.request(proto::GetLlmToken {}).await?;
847 *lock = Some(response.token.clone());
848 Ok(response.token.clone())
849 }
850}
851
852struct ConfigurationView {
853 state: gpui::Model<State>,
854}
855
856impl ConfigurationView {
857 fn authenticate(&mut self, cx: &mut ViewContext<Self>) {
858 self.state.update(cx, |state, cx| {
859 state.authenticate(cx).detach_and_log_err(cx);
860 });
861 cx.notify();
862 }
863
864 fn render_accept_terms(&mut self, cx: &mut ViewContext<Self>) -> Option<AnyElement> {
865 if self.state.read(cx).has_accepted_terms_of_service(cx) {
866 return None;
867 }
868
869 let accept_terms_disabled = self.state.read(cx).accept_terms.is_some();
870
871 let terms_button = Button::new("terms_of_service", "Terms of Service")
872 .style(ButtonStyle::Subtle)
873 .icon(IconName::ExternalLink)
874 .icon_color(Color::Muted)
875 .on_click(move |_, cx| cx.open_url("https://zed.dev/terms-of-service"));
876
877 let text =
878 "In order to use Zed AI, please read and accept our terms and conditions to continue:";
879
880 let form = v_flex()
881 .gap_2()
882 .child(Label::new("Terms and Conditions"))
883 .child(Label::new(text))
884 .child(h_flex().justify_center().child(terms_button))
885 .child(
886 h_flex().justify_center().child(
887 Button::new("accept_terms", "I've read and accept the terms of service")
888 .style(ButtonStyle::Tinted(TintColor::Accent))
889 .disabled(accept_terms_disabled)
890 .on_click({
891 let state = self.state.downgrade();
892 move |_, cx| {
893 state
894 .update(cx, |state, cx| state.accept_terms_of_service(cx))
895 .ok();
896 }
897 }),
898 ),
899 );
900
901 Some(form.into_any())
902 }
903}
904
905impl Render for ConfigurationView {
906 fn render(&mut self, cx: &mut ViewContext<Self>) -> impl IntoElement {
907 const ZED_AI_URL: &str = "https://zed.dev/ai";
908
909 let is_connected = !self.state.read(cx).is_signed_out();
910 let plan = self.state.read(cx).user_store.read(cx).current_plan();
911 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
912
913 let is_pro = plan == Some(proto::Plan::ZedPro);
914 let subscription_text = Label::new(if is_pro {
915 "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."
916 } else {
917 "You have basic access to models from Anthropic through the Zed AI Free plan."
918 });
919 let manage_subscription_button = if is_pro {
920 Some(
921 h_flex().child(
922 Button::new("manage_settings", "Manage Subscription")
923 .style(ButtonStyle::Tinted(TintColor::Accent))
924 .on_click(cx.listener(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))),
925 ),
926 )
927 } else if cx.has_flag::<ZedPro>() {
928 Some(
929 h_flex()
930 .gap_2()
931 .child(
932 Button::new("learn_more", "Learn more")
933 .style(ButtonStyle::Subtle)
934 .on_click(cx.listener(|_, _, cx| cx.open_url(ZED_AI_URL))),
935 )
936 .child(
937 Button::new("upgrade", "Upgrade")
938 .style(ButtonStyle::Subtle)
939 .color(Color::Accent)
940 .on_click(
941 cx.listener(|_, _, cx| cx.open_url(&zed_urls::account_url(cx))),
942 ),
943 ),
944 )
945 } else {
946 None
947 };
948
949 if is_connected {
950 v_flex()
951 .gap_3()
952 .max_w_4_5()
953 .children(self.render_accept_terms(cx))
954 .when(has_accepted_terms, |this| {
955 this.child(subscription_text)
956 .children(manage_subscription_button)
957 })
958 } else {
959 v_flex()
960 .gap_2()
961 .child(Label::new("Use Zed AI to access hosted language models."))
962 .child(
963 Button::new("sign_in", "Sign In")
964 .icon_color(Color::Muted)
965 .icon(IconName::Github)
966 .icon_position(IconPosition::Start)
967 .on_click(cx.listener(move |this, _, cx| this.authenticate(cx))),
968 )
969 }
970 }
971}