1use anthropic::{AnthropicError, AnthropicModelMode};
2use anyhow::{anyhow, Result};
3use client::{
4 zed_urls, Client, PerformCompletionParams, UserStore, EXPIRED_LLM_TOKEN_HEADER_NAME,
5 MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
6};
7use collections::BTreeMap;
8use feature_flags::{FeatureFlagAppExt, LlmClosedBeta, ZedPro};
9use futures::{
10 future::BoxFuture, stream::BoxStream, AsyncBufReadExt, FutureExt, Stream, StreamExt,
11 TryStreamExt as _,
12};
13use gpui::{AnyElement, AnyView, App, AsyncApp, Context, Entity, Subscription, Task};
14use http_client::{AsyncBody, HttpClient, Method, Response, StatusCode};
15use language_model::{
16 AuthenticateError, CloudModel, LanguageModel, LanguageModelCacheConfiguration, LanguageModelId,
17 LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,
18 LanguageModelProviderState, LanguageModelProviderTosView, LanguageModelRequest, RateLimiter,
19 ZED_CLOUD_PROVIDER_ID,
20};
21use language_model::{
22 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider, LlmApiToken,
23 MaxMonthlySpendReachedError, PaymentRequiredError, RefreshLlmTokenListener,
24};
25use schemars::JsonSchema;
26use serde::{de::DeserializeOwned, Deserialize, Serialize};
27use serde_json::value::RawValue;
28use settings::{Settings, SettingsStore};
29use smol::io::{AsyncReadExt, BufReader};
30use std::{
31 future,
32 sync::{Arc, LazyLock},
33};
34use strum::IntoEnumIterator;
35use ui::{prelude::*, TintColor};
36
37use crate::provider::anthropic::{
38 count_anthropic_tokens, into_anthropic, map_to_language_model_completion_events,
39};
40use crate::provider::google::into_google;
41use crate::provider::open_ai::{count_open_ai_tokens, into_open_ai};
42use crate::AllLanguageModelSettings;
43
44pub const PROVIDER_NAME: &str = "Zed";
45
46const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
47 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
48
49fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
50 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
51 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
52 .map(|json| serde_json::from_str(json).unwrap())
53 .unwrap_or_default()
54 });
55 ADDITIONAL_MODELS.as_slice()
56}
57
58#[derive(Default, Clone, Debug, PartialEq)]
59pub struct ZedDotDevSettings {
60 pub available_models: Vec<AvailableModel>,
61}
62
63#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
64#[serde(rename_all = "lowercase")]
65pub enum AvailableProvider {
66 Anthropic,
67 OpenAi,
68 Google,
69}
70
71#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
72pub struct AvailableModel {
73 /// The provider of the language model.
74 pub provider: AvailableProvider,
75 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
76 pub name: String,
77 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
78 pub display_name: Option<String>,
79 /// The size of the context window, indicating the maximum number of tokens the model can process.
80 pub max_tokens: usize,
81 /// The maximum number of output tokens allowed by the model.
82 pub max_output_tokens: Option<u32>,
83 /// The maximum number of completion tokens allowed by the model (o1-* only)
84 pub max_completion_tokens: Option<u32>,
85 /// Override this model with a different Anthropic model for tool calls.
86 pub tool_override: Option<String>,
87 /// Indicates whether this custom model supports caching.
88 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
89 /// The default temperature to use for this model.
90 pub default_temperature: Option<f32>,
91 /// Any extra beta headers to provide when using the model.
92 #[serde(default)]
93 pub extra_beta_headers: Vec<String>,
94 /// The model's mode (e.g. thinking)
95 pub mode: Option<ModelMode>,
96}
97
98#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
99#[serde(tag = "type", rename_all = "lowercase")]
100pub enum ModelMode {
101 #[default]
102 Default,
103 Thinking {
104 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
105 budget_tokens: Option<u32>,
106 },
107}
108
109impl From<ModelMode> for AnthropicModelMode {
110 fn from(value: ModelMode) -> Self {
111 match value {
112 ModelMode::Default => AnthropicModelMode::Default,
113 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
114 }
115 }
116}
117
118pub struct CloudLanguageModelProvider {
119 client: Arc<Client>,
120 state: gpui::Entity<State>,
121 _maintain_client_status: Task<()>,
122}
123
124pub struct State {
125 client: Arc<Client>,
126 llm_api_token: LlmApiToken,
127 user_store: Entity<UserStore>,
128 status: client::Status,
129 accept_terms: Option<Task<Result<()>>>,
130 _settings_subscription: Subscription,
131 _llm_token_subscription: Subscription,
132}
133
134impl State {
135 fn new(
136 client: Arc<Client>,
137 user_store: Entity<UserStore>,
138 status: client::Status,
139 cx: &mut Context<Self>,
140 ) -> Self {
141 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
142
143 Self {
144 client: client.clone(),
145 llm_api_token: LlmApiToken::default(),
146 user_store,
147 status,
148 accept_terms: None,
149 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
150 cx.notify();
151 }),
152 _llm_token_subscription: cx.subscribe(
153 &refresh_llm_token_listener,
154 |this, _listener, _event, cx| {
155 let client = this.client.clone();
156 let llm_api_token = this.llm_api_token.clone();
157 cx.spawn(async move |_this, _cx| {
158 llm_api_token.refresh(&client).await?;
159 anyhow::Ok(())
160 })
161 .detach_and_log_err(cx);
162 },
163 ),
164 }
165 }
166
167 fn is_signed_out(&self) -> bool {
168 self.status.is_signed_out()
169 }
170
171 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
172 let client = self.client.clone();
173 cx.spawn(async move |this, cx| {
174 client.authenticate_and_connect(true, &cx).await?;
175 this.update(cx, |_, cx| cx.notify())
176 })
177 }
178
179 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
180 self.user_store
181 .read(cx)
182 .current_user_has_accepted_terms()
183 .unwrap_or(false)
184 }
185
186 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
187 let user_store = self.user_store.clone();
188 self.accept_terms = Some(cx.spawn(async move |this, cx| {
189 let _ = user_store
190 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
191 .await;
192 this.update(cx, |this, cx| {
193 this.accept_terms = None;
194 cx.notify()
195 })
196 }));
197 }
198}
199
200impl CloudLanguageModelProvider {
201 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
202 let mut status_rx = client.status();
203 let status = *status_rx.borrow();
204
205 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
206
207 let state_ref = state.downgrade();
208 let maintain_client_status = cx.spawn(async move |cx| {
209 while let Some(status) = status_rx.next().await {
210 if let Some(this) = state_ref.upgrade() {
211 _ = this.update(cx, |this, cx| {
212 if this.status != status {
213 this.status = status;
214 cx.notify();
215 }
216 });
217 } else {
218 break;
219 }
220 }
221 });
222
223 Self {
224 client,
225 state: state.clone(),
226 _maintain_client_status: maintain_client_status,
227 }
228 }
229}
230
231impl LanguageModelProviderState for CloudLanguageModelProvider {
232 type ObservableEntity = State;
233
234 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
235 Some(self.state.clone())
236 }
237}
238
239impl LanguageModelProvider for CloudLanguageModelProvider {
240 fn id(&self) -> LanguageModelProviderId {
241 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
242 }
243
244 fn name(&self) -> LanguageModelProviderName {
245 LanguageModelProviderName(PROVIDER_NAME.into())
246 }
247
248 fn icon(&self) -> IconName {
249 IconName::AiZed
250 }
251
252 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
253 let llm_api_token = self.state.read(cx).llm_api_token.clone();
254 let model = CloudModel::Anthropic(anthropic::Model::default());
255 Some(Arc::new(CloudLanguageModel {
256 id: LanguageModelId::from(model.id().to_string()),
257 model,
258 llm_api_token: llm_api_token.clone(),
259 client: self.client.clone(),
260 request_limiter: RateLimiter::new(4),
261 }))
262 }
263
264 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
265 let mut models = BTreeMap::default();
266
267 if cx.is_staff() {
268 for model in anthropic::Model::iter() {
269 if !matches!(model, anthropic::Model::Custom { .. }) {
270 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
271 }
272 }
273 for model in open_ai::Model::iter() {
274 if !matches!(model, open_ai::Model::Custom { .. }) {
275 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
276 }
277 }
278 for model in google_ai::Model::iter() {
279 if !matches!(model, google_ai::Model::Custom { .. }) {
280 models.insert(model.id().to_string(), CloudModel::Google(model));
281 }
282 }
283 } else {
284 models.insert(
285 anthropic::Model::Claude3_5Sonnet.id().to_string(),
286 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
287 );
288 models.insert(
289 anthropic::Model::Claude3_7Sonnet.id().to_string(),
290 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
291 );
292 models.insert(
293 anthropic::Model::Claude3_7SonnetThinking.id().to_string(),
294 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
295 );
296 }
297
298 let llm_closed_beta_models = if cx.has_flag::<LlmClosedBeta>() {
299 zed_cloud_provider_additional_models()
300 } else {
301 &[]
302 };
303
304 // Override with available models from settings
305 for model in AllLanguageModelSettings::get_global(cx)
306 .zed_dot_dev
307 .available_models
308 .iter()
309 .chain(llm_closed_beta_models)
310 .cloned()
311 {
312 let model = match model.provider {
313 AvailableProvider::Anthropic => CloudModel::Anthropic(anthropic::Model::Custom {
314 name: model.name.clone(),
315 display_name: model.display_name.clone(),
316 max_tokens: model.max_tokens,
317 tool_override: model.tool_override.clone(),
318 cache_configuration: model.cache_configuration.as_ref().map(|config| {
319 anthropic::AnthropicModelCacheConfiguration {
320 max_cache_anchors: config.max_cache_anchors,
321 should_speculate: config.should_speculate,
322 min_total_token: config.min_total_token,
323 }
324 }),
325 default_temperature: model.default_temperature,
326 max_output_tokens: model.max_output_tokens,
327 extra_beta_headers: model.extra_beta_headers.clone(),
328 mode: model.mode.unwrap_or_default().into(),
329 }),
330 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
331 name: model.name.clone(),
332 display_name: model.display_name.clone(),
333 max_tokens: model.max_tokens,
334 max_output_tokens: model.max_output_tokens,
335 max_completion_tokens: model.max_completion_tokens,
336 }),
337 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
338 name: model.name.clone(),
339 display_name: model.display_name.clone(),
340 max_tokens: model.max_tokens,
341 }),
342 };
343 models.insert(model.id().to_string(), model.clone());
344 }
345
346 let llm_api_token = self.state.read(cx).llm_api_token.clone();
347 models
348 .into_values()
349 .map(|model| {
350 Arc::new(CloudLanguageModel {
351 id: LanguageModelId::from(model.id().to_string()),
352 model,
353 llm_api_token: llm_api_token.clone(),
354 client: self.client.clone(),
355 request_limiter: RateLimiter::new(4),
356 }) as Arc<dyn LanguageModel>
357 })
358 .collect()
359 }
360
361 fn is_authenticated(&self, cx: &App) -> bool {
362 !self.state.read(cx).is_signed_out()
363 }
364
365 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
366 Task::ready(Ok(()))
367 }
368
369 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
370 cx.new(|_| ConfigurationView {
371 state: self.state.clone(),
372 })
373 .into()
374 }
375
376 fn must_accept_terms(&self, cx: &App) -> bool {
377 !self.state.read(cx).has_accepted_terms_of_service(cx)
378 }
379
380 fn render_accept_terms(
381 &self,
382 view: LanguageModelProviderTosView,
383 cx: &mut App,
384 ) -> Option<AnyElement> {
385 render_accept_terms(self.state.clone(), view, cx)
386 }
387
388 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
389 Task::ready(Ok(()))
390 }
391}
392
393fn render_accept_terms(
394 state: Entity<State>,
395 view_kind: LanguageModelProviderTosView,
396 cx: &mut App,
397) -> Option<AnyElement> {
398 if state.read(cx).has_accepted_terms_of_service(cx) {
399 return None;
400 }
401
402 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
403
404 let terms_button = Button::new("terms_of_service", "Terms of Service")
405 .style(ButtonStyle::Subtle)
406 .icon(IconName::ArrowUpRight)
407 .icon_color(Color::Muted)
408 .icon_size(IconSize::XSmall)
409 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
410
411 let text = "To start using Zed AI, please read and accept the";
412
413 let form = v_flex()
414 .w_full()
415 .gap_2()
416 .child(
417 h_flex()
418 .flex_wrap()
419 .items_start()
420 .child(Label::new(text))
421 .child(terms_button),
422 )
423 .child({
424 let button_container = h_flex().w_full().child(
425 Button::new("accept_terms", "I accept the Terms of Service")
426 .style(ButtonStyle::Tinted(TintColor::Accent))
427 .disabled(accept_terms_disabled)
428 .on_click({
429 let state = state.downgrade();
430 move |_, _window, cx| {
431 state
432 .update(cx, |state, cx| state.accept_terms_of_service(cx))
433 .ok();
434 }
435 }),
436 );
437
438 match view_kind {
439 LanguageModelProviderTosView::PromptEditorPopup => button_container.justify_end(),
440 LanguageModelProviderTosView::Configuration
441 | LanguageModelProviderTosView::ThreadEmptyState => {
442 button_container.justify_start()
443 }
444 }
445 });
446
447 Some(form.into_any())
448}
449
450pub struct CloudLanguageModel {
451 id: LanguageModelId,
452 model: CloudModel,
453 llm_api_token: LlmApiToken,
454 client: Arc<Client>,
455 request_limiter: RateLimiter,
456}
457
458impl CloudLanguageModel {
459 async fn perform_llm_completion(
460 client: Arc<Client>,
461 llm_api_token: LlmApiToken,
462 body: PerformCompletionParams,
463 ) -> Result<Response<AsyncBody>> {
464 let http_client = &client.http_client();
465
466 let mut token = llm_api_token.acquire(&client).await?;
467 let mut did_retry = false;
468
469 let response = loop {
470 let request_builder = http_client::Request::builder();
471 let request = request_builder
472 .method(Method::POST)
473 .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
474 .header("Content-Type", "application/json")
475 .header("Authorization", format!("Bearer {token}"))
476 .body(serde_json::to_string(&body)?.into())?;
477 let mut response = http_client.send(request).await?;
478 if response.status().is_success() {
479 break response;
480 } else if !did_retry
481 && response
482 .headers()
483 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
484 .is_some()
485 {
486 did_retry = true;
487 token = llm_api_token.refresh(&client).await?;
488 } else if response.status() == StatusCode::FORBIDDEN
489 && response
490 .headers()
491 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
492 .is_some()
493 {
494 break Err(anyhow!(MaxMonthlySpendReachedError))?;
495 } else if response.status() == StatusCode::PAYMENT_REQUIRED {
496 break Err(anyhow!(PaymentRequiredError))?;
497 } else {
498 let mut body = String::new();
499 response.body_mut().read_to_string(&mut body).await?;
500 break Err(anyhow!(
501 "cloud language model completion failed with status {}: {body}",
502 response.status()
503 ))?;
504 }
505 };
506
507 Ok(response)
508 }
509}
510
511impl LanguageModel for CloudLanguageModel {
512 fn id(&self) -> LanguageModelId {
513 self.id.clone()
514 }
515
516 fn name(&self) -> LanguageModelName {
517 LanguageModelName::from(self.model.display_name().to_string())
518 }
519
520 fn icon(&self) -> Option<IconName> {
521 self.model.icon()
522 }
523
524 fn provider_id(&self) -> LanguageModelProviderId {
525 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
526 }
527
528 fn provider_name(&self) -> LanguageModelProviderName {
529 LanguageModelProviderName(PROVIDER_NAME.into())
530 }
531
532 fn telemetry_id(&self) -> String {
533 format!("zed.dev/{}", self.model.id())
534 }
535
536 fn availability(&self) -> LanguageModelAvailability {
537 self.model.availability()
538 }
539
540 fn max_token_count(&self) -> usize {
541 self.model.max_token_count()
542 }
543
544 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
545 match &self.model {
546 CloudModel::Anthropic(model) => {
547 model
548 .cache_configuration()
549 .map(|cache| LanguageModelCacheConfiguration {
550 max_cache_anchors: cache.max_cache_anchors,
551 should_speculate: cache.should_speculate,
552 min_total_token: cache.min_total_token,
553 })
554 }
555 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
556 }
557 }
558
559 fn count_tokens(
560 &self,
561 request: LanguageModelRequest,
562 cx: &App,
563 ) -> BoxFuture<'static, Result<usize>> {
564 match self.model.clone() {
565 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
566 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
567 CloudModel::Google(model) => {
568 let client = self.client.clone();
569 let request = into_google(request, model.id().into());
570 let request = google_ai::CountTokensRequest {
571 contents: request.contents,
572 };
573 async move {
574 let request = serde_json::to_string(&request)?;
575 let response = client
576 .request(proto::CountLanguageModelTokens {
577 provider: proto::LanguageModelProvider::Google as i32,
578 request,
579 })
580 .await?;
581 Ok(response.token_count as usize)
582 }
583 .boxed()
584 }
585 }
586 }
587
588 fn stream_completion(
589 &self,
590 request: LanguageModelRequest,
591 _cx: &AsyncApp,
592 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
593 match &self.model {
594 CloudModel::Anthropic(model) => {
595 let request = into_anthropic(
596 request,
597 model.request_id().into(),
598 model.default_temperature(),
599 model.max_output_tokens(),
600 model.mode(),
601 );
602 let client = self.client.clone();
603 let llm_api_token = self.llm_api_token.clone();
604 let future = self.request_limiter.stream(async move {
605 let response = Self::perform_llm_completion(
606 client.clone(),
607 llm_api_token,
608 PerformCompletionParams {
609 provider: client::LanguageModelProvider::Anthropic,
610 model: request.model.clone(),
611 provider_request: RawValue::from_string(serde_json::to_string(
612 &request,
613 )?)?,
614 },
615 )
616 .await?;
617 Ok(map_to_language_model_completion_events(Box::pin(
618 response_lines(response).map_err(AnthropicError::Other),
619 )))
620 });
621 async move { Ok(future.await?.boxed()) }.boxed()
622 }
623 CloudModel::OpenAi(model) => {
624 let client = self.client.clone();
625 let request = into_open_ai(request, model.id().into(), model.max_output_tokens());
626 let llm_api_token = self.llm_api_token.clone();
627 let future = self.request_limiter.stream(async move {
628 let response = Self::perform_llm_completion(
629 client.clone(),
630 llm_api_token,
631 PerformCompletionParams {
632 provider: client::LanguageModelProvider::OpenAi,
633 model: request.model.clone(),
634 provider_request: RawValue::from_string(serde_json::to_string(
635 &request,
636 )?)?,
637 },
638 )
639 .await?;
640 Ok(open_ai::extract_text_from_events(response_lines(response)))
641 });
642 async move {
643 Ok(future
644 .await?
645 .map(|result| result.map(LanguageModelCompletionEvent::Text))
646 .boxed())
647 }
648 .boxed()
649 }
650 CloudModel::Google(model) => {
651 let client = self.client.clone();
652 let request = into_google(request, model.id().into());
653 let llm_api_token = self.llm_api_token.clone();
654 let future = self.request_limiter.stream(async move {
655 let response = Self::perform_llm_completion(
656 client.clone(),
657 llm_api_token,
658 PerformCompletionParams {
659 provider: client::LanguageModelProvider::Google,
660 model: request.model.clone(),
661 provider_request: RawValue::from_string(serde_json::to_string(
662 &request,
663 )?)?,
664 },
665 )
666 .await?;
667 Ok(google_ai::extract_text_from_events(response_lines(
668 response,
669 )))
670 });
671 async move {
672 Ok(future
673 .await?
674 .map(|result| result.map(LanguageModelCompletionEvent::Text))
675 .boxed())
676 }
677 .boxed()
678 }
679 }
680 }
681
682 fn use_any_tool(
683 &self,
684 request: LanguageModelRequest,
685 tool_name: String,
686 tool_description: String,
687 input_schema: serde_json::Value,
688 _cx: &AsyncApp,
689 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
690 let client = self.client.clone();
691 let llm_api_token = self.llm_api_token.clone();
692
693 match &self.model {
694 CloudModel::Anthropic(model) => {
695 let mut request = into_anthropic(
696 request,
697 model.tool_model_id().into(),
698 model.default_temperature(),
699 model.max_output_tokens(),
700 model.mode(),
701 );
702 request.tool_choice = Some(anthropic::ToolChoice::Tool {
703 name: tool_name.clone(),
704 });
705 request.tools = vec![anthropic::Tool {
706 name: tool_name.clone(),
707 description: tool_description,
708 input_schema,
709 }];
710
711 self.request_limiter
712 .run(async move {
713 let response = Self::perform_llm_completion(
714 client.clone(),
715 llm_api_token,
716 PerformCompletionParams {
717 provider: client::LanguageModelProvider::Anthropic,
718 model: request.model.clone(),
719 provider_request: RawValue::from_string(serde_json::to_string(
720 &request,
721 )?)?,
722 },
723 )
724 .await?;
725
726 Ok(anthropic::extract_tool_args_from_events(
727 tool_name,
728 Box::pin(response_lines(response)),
729 )
730 .await?
731 .boxed())
732 })
733 .boxed()
734 }
735 CloudModel::OpenAi(model) => {
736 let mut request =
737 into_open_ai(request, model.id().into(), model.max_output_tokens());
738 request.tool_choice = Some(open_ai::ToolChoice::Other(
739 open_ai::ToolDefinition::Function {
740 function: open_ai::FunctionDefinition {
741 name: tool_name.clone(),
742 description: None,
743 parameters: None,
744 },
745 },
746 ));
747 request.tools = vec![open_ai::ToolDefinition::Function {
748 function: open_ai::FunctionDefinition {
749 name: tool_name.clone(),
750 description: Some(tool_description),
751 parameters: Some(input_schema),
752 },
753 }];
754
755 self.request_limiter
756 .run(async move {
757 let response = Self::perform_llm_completion(
758 client.clone(),
759 llm_api_token,
760 PerformCompletionParams {
761 provider: client::LanguageModelProvider::OpenAi,
762 model: request.model.clone(),
763 provider_request: RawValue::from_string(serde_json::to_string(
764 &request,
765 )?)?,
766 },
767 )
768 .await?;
769
770 Ok(open_ai::extract_tool_args_from_events(
771 tool_name,
772 Box::pin(response_lines(response)),
773 )
774 .await?
775 .boxed())
776 })
777 .boxed()
778 }
779 CloudModel::Google(_) => {
780 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
781 }
782 }
783 }
784}
785
786fn response_lines<T: DeserializeOwned>(
787 response: Response<AsyncBody>,
788) -> impl Stream<Item = Result<T>> {
789 futures::stream::try_unfold(
790 (String::new(), BufReader::new(response.into_body())),
791 move |(mut line, mut body)| async {
792 match body.read_line(&mut line).await {
793 Ok(0) => Ok(None),
794 Ok(_) => {
795 let event: T = serde_json::from_str(&line)?;
796 line.clear();
797 Ok(Some((event, (line, body))))
798 }
799 Err(e) => Err(e.into()),
800 }
801 },
802 )
803}
804
805struct ConfigurationView {
806 state: gpui::Entity<State>,
807}
808
809impl ConfigurationView {
810 fn authenticate(&mut self, cx: &mut Context<Self>) {
811 self.state.update(cx, |state, cx| {
812 state.authenticate(cx).detach_and_log_err(cx);
813 });
814 cx.notify();
815 }
816}
817
818impl Render for ConfigurationView {
819 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
820 const ZED_AI_URL: &str = "https://zed.dev/ai";
821
822 let is_connected = !self.state.read(cx).is_signed_out();
823 let plan = self.state.read(cx).user_store.read(cx).current_plan();
824 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
825
826 let is_pro = plan == Some(proto::Plan::ZedPro);
827 let subscription_text = Label::new(if is_pro {
828 "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."
829 } else {
830 "You have basic access to models from Anthropic through the Zed AI Free plan."
831 });
832 let manage_subscription_button = if is_pro {
833 Some(
834 h_flex().child(
835 Button::new("manage_settings", "Manage Subscription")
836 .style(ButtonStyle::Tinted(TintColor::Accent))
837 .on_click(
838 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
839 ),
840 ),
841 )
842 } else if cx.has_flag::<ZedPro>() {
843 Some(
844 h_flex()
845 .gap_2()
846 .child(
847 Button::new("learn_more", "Learn more")
848 .style(ButtonStyle::Subtle)
849 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
850 )
851 .child(
852 Button::new("upgrade", "Upgrade")
853 .style(ButtonStyle::Subtle)
854 .color(Color::Accent)
855 .on_click(
856 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
857 ),
858 ),
859 )
860 } else {
861 None
862 };
863
864 if is_connected {
865 v_flex()
866 .gap_3()
867 .w_full()
868 .children(render_accept_terms(
869 self.state.clone(),
870 LanguageModelProviderTosView::Configuration,
871 cx,
872 ))
873 .when(has_accepted_terms, |this| {
874 this.child(subscription_text)
875 .children(manage_subscription_button)
876 })
877 } else {
878 v_flex()
879 .gap_2()
880 .child(Label::new("Use Zed AI to access hosted language models."))
881 .child(
882 Button::new("sign_in", "Sign In")
883 .icon_color(Color::Muted)
884 .icon(IconName::Github)
885 .icon_position(IconPosition::Start)
886 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
887 )
888 }
889 }
890}