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, AppContext, AsyncAppContext, EventEmitter, FontWeight, Global, Model,
16 ModelContext, ReadGlobal, 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 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(Model<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 AppContext) {
116 let listener = cx.new_model(|cx| RefreshLlmTokenListener::new(client, cx));
117 cx.set_global(GlobalRefreshLlmTokenListener(listener));
118 }
119
120 pub fn global(cx: &AppContext) -> Model<Self> {
121 GlobalRefreshLlmTokenListener::global(cx).0.clone()
122 }
123
124 fn new(client: Arc<Client>, cx: &mut ModelContext<Self>) -> Self {
125 Self {
126 _llm_token_subscription: client
127 .add_message_handler(cx.weak_model(), Self::handle_refresh_llm_token),
128 }
129 }
130
131 async fn handle_refresh_llm_token(
132 this: Model<Self>,
133 _: TypedEnvelope<proto::RefreshLlmToken>,
134 mut cx: AsyncAppContext,
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::Model<State>,
143 _maintain_client_status: Task<()>,
144}
145
146pub struct State {
147 client: Arc<Client>,
148 llm_api_token: LlmApiToken,
149 user_store: Model<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: Model<UserStore>,
160 status: client::Status,
161 cx: &mut ModelContext<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 ModelContext<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: &AppContext) -> 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 ModelContext<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: Model<UserStore>, client: Arc<Client>, cx: &mut AppContext) -> Self {
224 let mut status_rx = client.status();
225 let status = *status_rx.borrow();
226
227 let state = cx.new_model(|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::Model<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: &AppContext) -> 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: &AppContext) -> bool {
363 !self.state.read(cx).is_signed_out()
364 }
365
366 fn authenticate(&self, _cx: &mut AppContext) -> Task<Result<()>> {
367 Task::ready(Ok(()))
368 }
369
370 fn configuration_view(&self, cx: &mut WindowContext) -> AnyView {
371 cx.new_view(|_cx| ConfigurationView {
372 state: self.state.clone(),
373 })
374 .into()
375 }
376
377 fn must_accept_terms(&self, cx: &AppContext) -> bool {
378 !self.state.read(cx).has_accepted_terms_of_service(cx)
379 }
380
381 fn render_accept_terms(&self, cx: &mut WindowContext) -> Option<AnyElement> {
382 let state = self.state.read(cx);
383
384 let terms = [(
385 "terms_of_service",
386 "Terms of Service",
387 "https://zed.dev/terms-of-service",
388 )]
389 .map(|(id, label, url)| {
390 Button::new(id, label)
391 .style(ButtonStyle::Subtle)
392 .icon(IconName::ExternalLink)
393 .icon_size(IconSize::XSmall)
394 .icon_color(Color::Muted)
395 .on_click(move |_, cx| cx.open_url(url))
396 });
397
398 if state.has_accepted_terms_of_service(cx) {
399 None
400 } else {
401 let disabled = state.accept_terms.is_some();
402 Some(
403 v_flex()
404 .gap_2()
405 .child(
406 v_flex()
407 .child(Label::new("Terms and Conditions").weight(FontWeight::MEDIUM))
408 .child(
409 Label::new(
410 "Please read and accept our terms and conditions to continue.",
411 )
412 .size(LabelSize::Small),
413 ),
414 )
415 .child(v_flex().gap_1().children(terms))
416 .child(
417 h_flex().justify_end().child(
418 Button::new("accept_terms", "I've read it and accept it")
419 .disabled(disabled)
420 .on_click({
421 let state = self.state.downgrade();
422 move |_, cx| {
423 state
424 .update(cx, |state, cx| {
425 state.accept_terms_of_service(cx)
426 })
427 .ok();
428 }
429 }),
430 ),
431 )
432 .into_any(),
433 )
434 }
435 }
436
437 fn reset_credentials(&self, _cx: &mut AppContext) -> Task<Result<()>> {
438 Task::ready(Ok(()))
439 }
440}
441
442pub struct CloudLanguageModel {
443 id: LanguageModelId,
444 model: CloudModel,
445 llm_api_token: LlmApiToken,
446 client: Arc<Client>,
447 request_limiter: RateLimiter,
448}
449
450#[derive(Clone, Default)]
451pub struct LlmApiToken(Arc<RwLock<Option<String>>>);
452
453#[derive(Error, Debug)]
454pub struct PaymentRequiredError;
455
456impl fmt::Display for PaymentRequiredError {
457 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
458 write!(
459 f,
460 "Payment required to use this language model. Please upgrade your account."
461 )
462 }
463}
464
465#[derive(Error, Debug)]
466pub struct MaxMonthlySpendReachedError;
467
468impl fmt::Display for MaxMonthlySpendReachedError {
469 fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
470 write!(
471 f,
472 "Maximum spending limit reached for this month. For more usage, increase your spending limit."
473 )
474 }
475}
476
477impl CloudLanguageModel {
478 async fn perform_llm_completion(
479 client: Arc<Client>,
480 llm_api_token: LlmApiToken,
481 body: PerformCompletionParams,
482 ) -> Result<Response<AsyncBody>> {
483 let http_client = &client.http_client();
484
485 let mut token = llm_api_token.acquire(&client).await?;
486 let mut did_retry = false;
487
488 let response = loop {
489 let request_builder = http_client::Request::builder();
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(ZED_CLOUD_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 match &self.model {
613 CloudModel::Anthropic(model) => {
614 let request = request.into_anthropic(
615 model.id().into(),
616 model.default_temperature(),
617 model.max_output_tokens(),
618 );
619 let client = self.client.clone();
620 let llm_api_token = self.llm_api_token.clone();
621 let future = self.request_limiter.stream(async move {
622 let response = Self::perform_llm_completion(
623 client.clone(),
624 llm_api_token,
625 PerformCompletionParams {
626 provider: client::LanguageModelProvider::Anthropic,
627 model: request.model.clone(),
628 provider_request: RawValue::from_string(serde_json::to_string(
629 &request,
630 )?)?,
631 },
632 )
633 .await?;
634 Ok(map_to_language_model_completion_events(Box::pin(
635 response_lines(response).map_err(AnthropicError::Other),
636 )))
637 });
638 async move { Ok(future.await?.boxed()) }.boxed()
639 }
640 CloudModel::OpenAi(model) => {
641 let client = self.client.clone();
642 let request = request.into_open_ai(model.id().into(), model.max_output_tokens());
643 let llm_api_token = self.llm_api_token.clone();
644 let future = self.request_limiter.stream(async move {
645 let response = Self::perform_llm_completion(
646 client.clone(),
647 llm_api_token,
648 PerformCompletionParams {
649 provider: client::LanguageModelProvider::OpenAi,
650 model: request.model.clone(),
651 provider_request: RawValue::from_string(serde_json::to_string(
652 &request,
653 )?)?,
654 },
655 )
656 .await?;
657 Ok(open_ai::extract_text_from_events(response_lines(response)))
658 });
659 async move {
660 Ok(future
661 .await?
662 .map(|result| result.map(LanguageModelCompletionEvent::Text))
663 .boxed())
664 }
665 .boxed()
666 }
667 CloudModel::Google(model) => {
668 let client = self.client.clone();
669 let request = request.into_google(model.id().into());
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::Google,
677 model: request.model.clone(),
678 provider_request: RawValue::from_string(serde_json::to_string(
679 &request,
680 )?)?,
681 },
682 )
683 .await?;
684 Ok(google_ai::extract_text_from_events(response_lines(
685 response,
686 )))
687 });
688 async move {
689 Ok(future
690 .await?
691 .map(|result| result.map(LanguageModelCompletionEvent::Text))
692 .boxed())
693 }
694 .boxed()
695 }
696 }
697 }
698
699 fn use_any_tool(
700 &self,
701 request: LanguageModelRequest,
702 tool_name: String,
703 tool_description: String,
704 input_schema: serde_json::Value,
705 _cx: &AsyncAppContext,
706 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
707 let client = self.client.clone();
708 let llm_api_token = self.llm_api_token.clone();
709
710 match &self.model {
711 CloudModel::Anthropic(model) => {
712 let mut request = request.into_anthropic(
713 model.tool_model_id().into(),
714 model.default_temperature(),
715 model.max_output_tokens(),
716 );
717 request.tool_choice = Some(anthropic::ToolChoice::Tool {
718 name: tool_name.clone(),
719 });
720 request.tools = vec![anthropic::Tool {
721 name: tool_name.clone(),
722 description: tool_description,
723 input_schema,
724 }];
725
726 self.request_limiter
727 .run(async move {
728 let response = Self::perform_llm_completion(
729 client.clone(),
730 llm_api_token,
731 PerformCompletionParams {
732 provider: client::LanguageModelProvider::Anthropic,
733 model: request.model.clone(),
734 provider_request: RawValue::from_string(serde_json::to_string(
735 &request,
736 )?)?,
737 },
738 )
739 .await?;
740
741 Ok(anthropic::extract_tool_args_from_events(
742 tool_name,
743 Box::pin(response_lines(response)),
744 )
745 .await?
746 .boxed())
747 })
748 .boxed()
749 }
750 CloudModel::OpenAi(model) => {
751 let mut request =
752 request.into_open_ai(model.id().into(), model.max_output_tokens());
753 request.tool_choice = Some(open_ai::ToolChoice::Other(
754 open_ai::ToolDefinition::Function {
755 function: open_ai::FunctionDefinition {
756 name: tool_name.clone(),
757 description: None,
758 parameters: None,
759 },
760 },
761 ));
762 request.tools = vec![open_ai::ToolDefinition::Function {
763 function: open_ai::FunctionDefinition {
764 name: tool_name.clone(),
765 description: Some(tool_description),
766 parameters: Some(input_schema),
767 },
768 }];
769
770 self.request_limiter
771 .run(async move {
772 let response = Self::perform_llm_completion(
773 client.clone(),
774 llm_api_token,
775 PerformCompletionParams {
776 provider: client::LanguageModelProvider::OpenAi,
777 model: request.model.clone(),
778 provider_request: RawValue::from_string(serde_json::to_string(
779 &request,
780 )?)?,
781 },
782 )
783 .await?;
784
785 Ok(open_ai::extract_tool_args_from_events(
786 tool_name,
787 Box::pin(response_lines(response)),
788 )
789 .await?
790 .boxed())
791 })
792 .boxed()
793 }
794 CloudModel::Google(_) => {
795 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
796 }
797 }
798 }
799}
800
801fn response_lines<T: DeserializeOwned>(
802 response: Response<AsyncBody>,
803) -> impl Stream<Item = Result<T>> {
804 futures::stream::try_unfold(
805 (String::new(), BufReader::new(response.into_body())),
806 move |(mut line, mut body)| async {
807 match body.read_line(&mut line).await {
808 Ok(0) => Ok(None),
809 Ok(_) => {
810 let event: T = serde_json::from_str(&line)?;
811 line.clear();
812 Ok(Some((event, (line, body))))
813 }
814 Err(e) => Err(e.into()),
815 }
816 },
817 )
818}
819
820impl LlmApiToken {
821 pub async fn acquire(&self, client: &Arc<Client>) -> Result<String> {
822 let lock = self.0.upgradable_read().await;
823 if let Some(token) = lock.as_ref() {
824 Ok(token.to_string())
825 } else {
826 Self::fetch(RwLockUpgradableReadGuard::upgrade(lock).await, client).await
827 }
828 }
829
830 pub async fn refresh(&self, client: &Arc<Client>) -> Result<String> {
831 Self::fetch(self.0.write().await, client).await
832 }
833
834 async fn fetch<'a>(
835 mut lock: RwLockWriteGuard<'a, Option<String>>,
836 client: &Arc<Client>,
837 ) -> Result<String> {
838 let response = client.request(proto::GetLlmToken {}).await?;
839 *lock = Some(response.token.clone());
840 Ok(response.token.clone())
841 }
842}
843
844struct ConfigurationView {
845 state: gpui::Model<State>,
846}
847
848impl ConfigurationView {
849 fn authenticate(&mut self, cx: &mut ViewContext<Self>) {
850 self.state.update(cx, |state, cx| {
851 state.authenticate(cx).detach_and_log_err(cx);
852 });
853 cx.notify();
854 }
855
856 fn render_accept_terms(&mut self, cx: &mut ViewContext<Self>) -> Option<AnyElement> {
857 if self.state.read(cx).has_accepted_terms_of_service(cx) {
858 return None;
859 }
860
861 let accept_terms_disabled = self.state.read(cx).accept_terms.is_some();
862
863 let terms_button = Button::new("terms_of_service", "Terms of Service")
864 .style(ButtonStyle::Subtle)
865 .icon(IconName::ExternalLink)
866 .icon_color(Color::Muted)
867 .on_click(move |_, cx| cx.open_url("https://zed.dev/terms-of-service"));
868
869 let text =
870 "In order to use Zed AI, please read and accept our terms and conditions to continue:";
871
872 let form = v_flex()
873 .gap_2()
874 .child(Label::new("Terms and Conditions"))
875 .child(Label::new(text))
876 .child(h_flex().justify_center().child(terms_button))
877 .child(
878 h_flex().justify_center().child(
879 Button::new("accept_terms", "I've read and accept the terms of service")
880 .style(ButtonStyle::Tinted(TintColor::Accent))
881 .disabled(accept_terms_disabled)
882 .on_click({
883 let state = self.state.downgrade();
884 move |_, cx| {
885 state
886 .update(cx, |state, cx| state.accept_terms_of_service(cx))
887 .ok();
888 }
889 }),
890 ),
891 );
892
893 Some(form.into_any())
894 }
895}
896
897impl Render for ConfigurationView {
898 fn render(&mut self, cx: &mut ViewContext<Self>) -> impl IntoElement {
899 const ZED_AI_URL: &str = "https://zed.dev/ai";
900
901 let is_connected = !self.state.read(cx).is_signed_out();
902 let plan = self.state.read(cx).user_store.read(cx).current_plan();
903 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
904
905 let is_pro = plan == Some(proto::Plan::ZedPro);
906 let subscription_text = Label::new(if is_pro {
907 "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."
908 } else {
909 "You have basic access to models from Anthropic through the Zed AI Free plan."
910 });
911 let manage_subscription_button = if is_pro {
912 Some(
913 h_flex().child(
914 Button::new("manage_settings", "Manage Subscription")
915 .style(ButtonStyle::Tinted(TintColor::Accent))
916 .on_click(cx.listener(|_, _, cx| cx.open_url(&zed_urls::account_url(cx)))),
917 ),
918 )
919 } else if cx.has_flag::<ZedPro>() {
920 Some(
921 h_flex()
922 .gap_2()
923 .child(
924 Button::new("learn_more", "Learn more")
925 .style(ButtonStyle::Subtle)
926 .on_click(cx.listener(|_, _, cx| cx.open_url(ZED_AI_URL))),
927 )
928 .child(
929 Button::new("upgrade", "Upgrade")
930 .style(ButtonStyle::Subtle)
931 .color(Color::Accent)
932 .on_click(
933 cx.listener(|_, _, cx| cx.open_url(&zed_urls::account_url(cx))),
934 ),
935 ),
936 )
937 } else {
938 None
939 };
940
941 if is_connected {
942 v_flex()
943 .gap_3()
944 .max_w_4_5()
945 .children(self.render_accept_terms(cx))
946 .when(has_accepted_terms, |this| {
947 this.child(subscription_text)
948 .children(manage_subscription_button)
949 })
950 } else {
951 v_flex()
952 .gap_2()
953 .child(Label::new("Use Zed AI to access hosted language models."))
954 .child(
955 Button::new("sign_in", "Sign In")
956 .icon_color(Color::Muted)
957 .icon(IconName::Github)
958 .icon_position(IconPosition::Start)
959 .on_click(cx.listener(move |this, _, cx| this.authenticate(cx))),
960 )
961 }
962 }
963}