1use anthropic::{AnthropicError, AnthropicModelMode};
2use anyhow::{Result, anyhow};
3use client::{
4 Client, EXPIRED_LLM_TOKEN_HEADER_NAME, MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME,
5 PerformCompletionParams, UserStore, zed_urls,
6};
7use collections::BTreeMap;
8use feature_flags::{FeatureFlagAppExt, LlmClosedBeta, ZedPro};
9use futures::{
10 AsyncBufReadExt, FutureExt, Stream, StreamExt, TryStreamExt as _, future::BoxFuture,
11 stream::BoxStream,
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,
19 LanguageModelToolSchemaFormat, RateLimiter, ZED_CLOUD_PROVIDER_ID,
20};
21use language_model::{
22 LanguageModelAvailability, LanguageModelCompletionEvent, LanguageModelProvider, LlmApiToken,
23 MaxMonthlySpendReachedError, PaymentRequiredError, RefreshLlmTokenListener,
24};
25use schemars::JsonSchema;
26use serde::{Deserialize, Serialize, de::DeserializeOwned};
27use serde_json::value::RawValue;
28use settings::{Settings, SettingsStore};
29use smol::Timer;
30use smol::io::{AsyncReadExt, BufReader};
31use std::{
32 future,
33 sync::{Arc, LazyLock},
34 time::Duration,
35};
36use strum::IntoEnumIterator;
37use ui::{TintColor, prelude::*};
38
39use crate::AllLanguageModelSettings;
40use crate::provider::anthropic::{
41 count_anthropic_tokens, into_anthropic, map_to_language_model_completion_events,
42};
43use crate::provider::google::into_google;
44use crate::provider::open_ai::{count_open_ai_tokens, into_open_ai};
45
46pub const PROVIDER_NAME: &str = "Zed";
47
48const ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON: Option<&str> =
49 option_env!("ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON");
50
51fn zed_cloud_provider_additional_models() -> &'static [AvailableModel] {
52 static ADDITIONAL_MODELS: LazyLock<Vec<AvailableModel>> = LazyLock::new(|| {
53 ZED_CLOUD_PROVIDER_ADDITIONAL_MODELS_JSON
54 .map(|json| serde_json::from_str(json).unwrap())
55 .unwrap_or_default()
56 });
57 ADDITIONAL_MODELS.as_slice()
58}
59
60#[derive(Default, Clone, Debug, PartialEq)]
61pub struct ZedDotDevSettings {
62 pub available_models: Vec<AvailableModel>,
63}
64
65#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
66#[serde(rename_all = "lowercase")]
67pub enum AvailableProvider {
68 Anthropic,
69 OpenAi,
70 Google,
71}
72
73#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
74pub struct AvailableModel {
75 /// The provider of the language model.
76 pub provider: AvailableProvider,
77 /// The model's name in the provider's API. e.g. claude-3-5-sonnet-20240620
78 pub name: String,
79 /// The name displayed in the UI, such as in the assistant panel model dropdown menu.
80 pub display_name: Option<String>,
81 /// The size of the context window, indicating the maximum number of tokens the model can process.
82 pub max_tokens: usize,
83 /// The maximum number of output tokens allowed by the model.
84 pub max_output_tokens: Option<u32>,
85 /// The maximum number of completion tokens allowed by the model (o1-* only)
86 pub max_completion_tokens: Option<u32>,
87 /// Override this model with a different Anthropic model for tool calls.
88 pub tool_override: Option<String>,
89 /// Indicates whether this custom model supports caching.
90 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
91 /// The default temperature to use for this model.
92 pub default_temperature: Option<f32>,
93 /// Any extra beta headers to provide when using the model.
94 #[serde(default)]
95 pub extra_beta_headers: Vec<String>,
96 /// The model's mode (e.g. thinking)
97 pub mode: Option<ModelMode>,
98}
99
100#[derive(Default, Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
101#[serde(tag = "type", rename_all = "lowercase")]
102pub enum ModelMode {
103 #[default]
104 Default,
105 Thinking {
106 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
107 budget_tokens: Option<u32>,
108 },
109}
110
111impl From<ModelMode> for AnthropicModelMode {
112 fn from(value: ModelMode) -> Self {
113 match value {
114 ModelMode::Default => AnthropicModelMode::Default,
115 ModelMode::Thinking { budget_tokens } => AnthropicModelMode::Thinking { budget_tokens },
116 }
117 }
118}
119
120pub struct CloudLanguageModelProvider {
121 client: Arc<Client>,
122 state: gpui::Entity<State>,
123 _maintain_client_status: Task<()>,
124}
125
126pub struct State {
127 client: Arc<Client>,
128 llm_api_token: LlmApiToken,
129 user_store: Entity<UserStore>,
130 status: client::Status,
131 accept_terms: Option<Task<Result<()>>>,
132 _settings_subscription: Subscription,
133 _llm_token_subscription: Subscription,
134}
135
136impl State {
137 fn new(
138 client: Arc<Client>,
139 user_store: Entity<UserStore>,
140 status: client::Status,
141 cx: &mut Context<Self>,
142 ) -> Self {
143 let refresh_llm_token_listener = RefreshLlmTokenListener::global(cx);
144
145 Self {
146 client: client.clone(),
147 llm_api_token: LlmApiToken::default(),
148 user_store,
149 status,
150 accept_terms: None,
151 _settings_subscription: cx.observe_global::<SettingsStore>(|_, cx| {
152 cx.notify();
153 }),
154 _llm_token_subscription: cx.subscribe(
155 &refresh_llm_token_listener,
156 |this, _listener, _event, cx| {
157 let client = this.client.clone();
158 let llm_api_token = this.llm_api_token.clone();
159 cx.spawn(async move |_this, _cx| {
160 llm_api_token.refresh(&client).await?;
161 anyhow::Ok(())
162 })
163 .detach_and_log_err(cx);
164 },
165 ),
166 }
167 }
168
169 fn is_signed_out(&self) -> bool {
170 self.status.is_signed_out()
171 }
172
173 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
174 let client = self.client.clone();
175 cx.spawn(async move |this, cx| {
176 client.authenticate_and_connect(true, &cx).await?;
177 this.update(cx, |_, cx| cx.notify())
178 })
179 }
180
181 fn has_accepted_terms_of_service(&self, cx: &App) -> bool {
182 self.user_store
183 .read(cx)
184 .current_user_has_accepted_terms()
185 .unwrap_or(false)
186 }
187
188 fn accept_terms_of_service(&mut self, cx: &mut Context<Self>) {
189 let user_store = self.user_store.clone();
190 self.accept_terms = Some(cx.spawn(async move |this, cx| {
191 let _ = user_store
192 .update(cx, |store, cx| store.accept_terms_of_service(cx))?
193 .await;
194 this.update(cx, |this, cx| {
195 this.accept_terms = None;
196 cx.notify()
197 })
198 }));
199 }
200}
201
202impl CloudLanguageModelProvider {
203 pub fn new(user_store: Entity<UserStore>, client: Arc<Client>, cx: &mut App) -> Self {
204 let mut status_rx = client.status();
205 let status = *status_rx.borrow();
206
207 let state = cx.new(|cx| State::new(client.clone(), user_store.clone(), status, cx));
208
209 let state_ref = state.downgrade();
210 let maintain_client_status = cx.spawn(async move |cx| {
211 while let Some(status) = status_rx.next().await {
212 if let Some(this) = state_ref.upgrade() {
213 _ = this.update(cx, |this, cx| {
214 if this.status != status {
215 this.status = status;
216 cx.notify();
217 }
218 });
219 } else {
220 break;
221 }
222 }
223 });
224
225 Self {
226 client,
227 state: state.clone(),
228 _maintain_client_status: maintain_client_status,
229 }
230 }
231}
232
233impl LanguageModelProviderState for CloudLanguageModelProvider {
234 type ObservableEntity = State;
235
236 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
237 Some(self.state.clone())
238 }
239}
240
241impl LanguageModelProvider for CloudLanguageModelProvider {
242 fn id(&self) -> LanguageModelProviderId {
243 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
244 }
245
246 fn name(&self) -> LanguageModelProviderName {
247 LanguageModelProviderName(PROVIDER_NAME.into())
248 }
249
250 fn icon(&self) -> IconName {
251 IconName::AiZed
252 }
253
254 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
255 let llm_api_token = self.state.read(cx).llm_api_token.clone();
256 let model = CloudModel::Anthropic(anthropic::Model::default());
257 Some(Arc::new(CloudLanguageModel {
258 id: LanguageModelId::from(model.id().to_string()),
259 model,
260 llm_api_token: llm_api_token.clone(),
261 client: self.client.clone(),
262 request_limiter: RateLimiter::new(4),
263 }))
264 }
265
266 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
267 let mut models = BTreeMap::default();
268
269 if cx.is_staff() {
270 for model in anthropic::Model::iter() {
271 if !matches!(model, anthropic::Model::Custom { .. }) {
272 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
273 }
274 }
275 for model in open_ai::Model::iter() {
276 if !matches!(model, open_ai::Model::Custom { .. }) {
277 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
278 }
279 }
280 for model in google_ai::Model::iter() {
281 if !matches!(model, google_ai::Model::Custom { .. }) {
282 models.insert(model.id().to_string(), CloudModel::Google(model));
283 }
284 }
285 } else {
286 models.insert(
287 anthropic::Model::Claude3_5Sonnet.id().to_string(),
288 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
289 );
290 models.insert(
291 anthropic::Model::Claude3_7Sonnet.id().to_string(),
292 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
293 );
294 models.insert(
295 anthropic::Model::Claude3_7SonnetThinking.id().to_string(),
296 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
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 mode: model.mode.unwrap_or_default().into(),
331 }),
332 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
333 name: model.name.clone(),
334 display_name: model.display_name.clone(),
335 max_tokens: model.max_tokens,
336 max_output_tokens: model.max_output_tokens,
337 max_completion_tokens: model.max_completion_tokens,
338 }),
339 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
340 name: model.name.clone(),
341 display_name: model.display_name.clone(),
342 max_tokens: model.max_tokens,
343 }),
344 };
345 models.insert(model.id().to_string(), model.clone());
346 }
347
348 let llm_api_token = self.state.read(cx).llm_api_token.clone();
349 models
350 .into_values()
351 .map(|model| {
352 Arc::new(CloudLanguageModel {
353 id: LanguageModelId::from(model.id().to_string()),
354 model,
355 llm_api_token: llm_api_token.clone(),
356 client: self.client.clone(),
357 request_limiter: RateLimiter::new(4),
358 }) as Arc<dyn LanguageModel>
359 })
360 .collect()
361 }
362
363 fn is_authenticated(&self, cx: &App) -> bool {
364 !self.state.read(cx).is_signed_out()
365 }
366
367 fn authenticate(&self, _cx: &mut App) -> Task<Result<(), AuthenticateError>> {
368 Task::ready(Ok(()))
369 }
370
371 fn configuration_view(&self, _: &mut Window, cx: &mut App) -> AnyView {
372 cx.new(|_| ConfigurationView {
373 state: self.state.clone(),
374 })
375 .into()
376 }
377
378 fn must_accept_terms(&self, cx: &App) -> bool {
379 !self.state.read(cx).has_accepted_terms_of_service(cx)
380 }
381
382 fn render_accept_terms(
383 &self,
384 view: LanguageModelProviderTosView,
385 cx: &mut App,
386 ) -> Option<AnyElement> {
387 render_accept_terms(self.state.clone(), view, cx)
388 }
389
390 fn reset_credentials(&self, _cx: &mut App) -> Task<Result<()>> {
391 Task::ready(Ok(()))
392 }
393}
394
395fn render_accept_terms(
396 state: Entity<State>,
397 view_kind: LanguageModelProviderTosView,
398 cx: &mut App,
399) -> Option<AnyElement> {
400 if state.read(cx).has_accepted_terms_of_service(cx) {
401 return None;
402 }
403
404 let accept_terms_disabled = state.read(cx).accept_terms.is_some();
405
406 let terms_button = Button::new("terms_of_service", "Terms of Service")
407 .style(ButtonStyle::Subtle)
408 .icon(IconName::ArrowUpRight)
409 .icon_color(Color::Muted)
410 .icon_size(IconSize::XSmall)
411 .on_click(move |_, _window, cx| cx.open_url("https://zed.dev/terms-of-service"));
412
413 let text = "To start using Zed AI, please read and accept the";
414
415 let form = v_flex()
416 .w_full()
417 .gap_2()
418 .child(
419 h_flex()
420 .flex_wrap()
421 .items_start()
422 .child(Label::new(text))
423 .child(terms_button),
424 )
425 .child({
426 let button_container = h_flex().w_full().child(
427 Button::new("accept_terms", "I accept the Terms of Service")
428 .style(ButtonStyle::Tinted(TintColor::Accent))
429 .disabled(accept_terms_disabled)
430 .on_click({
431 let state = state.downgrade();
432 move |_, _window, cx| {
433 state
434 .update(cx, |state, cx| state.accept_terms_of_service(cx))
435 .ok();
436 }
437 }),
438 );
439
440 match view_kind {
441 LanguageModelProviderTosView::PromptEditorPopup => button_container.justify_end(),
442 LanguageModelProviderTosView::Configuration
443 | LanguageModelProviderTosView::ThreadEmptyState => {
444 button_container.justify_start()
445 }
446 }
447 });
448
449 Some(form.into_any())
450}
451
452pub struct CloudLanguageModel {
453 id: LanguageModelId,
454 model: CloudModel,
455 llm_api_token: LlmApiToken,
456 client: Arc<Client>,
457 request_limiter: RateLimiter,
458}
459
460impl CloudLanguageModel {
461 const MAX_RETRIES: usize = 3;
462
463 async fn perform_llm_completion(
464 client: Arc<Client>,
465 llm_api_token: LlmApiToken,
466 body: PerformCompletionParams,
467 ) -> Result<Response<AsyncBody>> {
468 let http_client = &client.http_client();
469
470 let mut token = llm_api_token.acquire(&client).await?;
471 let mut retries_remaining = Self::MAX_RETRIES;
472 let mut retry_delay = Duration::from_secs(1);
473
474 loop {
475 let request_builder = http_client::Request::builder();
476 let request = request_builder
477 .method(Method::POST)
478 .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
479 .header("Content-Type", "application/json")
480 .header("Authorization", format!("Bearer {token}"))
481 .body(serde_json::to_string(&body)?.into())?;
482 let mut response = http_client.send(request).await?;
483 let status = response.status();
484 if status.is_success() {
485 return Ok(response);
486 } else if response
487 .headers()
488 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
489 .is_some()
490 {
491 retries_remaining -= 1;
492 token = llm_api_token.refresh(&client).await?;
493 } else if status == StatusCode::FORBIDDEN
494 && response
495 .headers()
496 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
497 .is_some()
498 {
499 return Err(anyhow!(MaxMonthlySpendReachedError));
500 } else if status.as_u16() >= 500 && status.as_u16() < 600 {
501 // If we encounter an error in the 500 range, retry after a delay.
502 // We've seen at least these in the wild from API providers:
503 // * 500 Internal Server Error
504 // * 502 Bad Gateway
505 // * 529 Service Overloaded
506
507 if retries_remaining == 0 {
508 let mut body = String::new();
509 response.body_mut().read_to_string(&mut body).await?;
510 return Err(anyhow!(
511 "cloud language model completion failed after {} retries with status {status}: {body}",
512 Self::MAX_RETRIES
513 ));
514 }
515
516 Timer::after(retry_delay).await;
517
518 retries_remaining -= 1;
519 retry_delay *= 2; // If it fails again, wait longer.
520 } else if status == StatusCode::PAYMENT_REQUIRED {
521 return Err(anyhow!(PaymentRequiredError));
522 } else {
523 let mut body = String::new();
524 response.body_mut().read_to_string(&mut body).await?;
525 return Err(anyhow!(
526 "cloud language model completion failed with status {status}: {body}",
527 ));
528 }
529 }
530 }
531}
532
533impl LanguageModel for CloudLanguageModel {
534 fn id(&self) -> LanguageModelId {
535 self.id.clone()
536 }
537
538 fn name(&self) -> LanguageModelName {
539 LanguageModelName::from(self.model.display_name().to_string())
540 }
541
542 fn icon(&self) -> Option<IconName> {
543 self.model.icon()
544 }
545
546 fn provider_id(&self) -> LanguageModelProviderId {
547 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
548 }
549
550 fn provider_name(&self) -> LanguageModelProviderName {
551 LanguageModelProviderName(PROVIDER_NAME.into())
552 }
553
554 fn telemetry_id(&self) -> String {
555 format!("zed.dev/{}", self.model.id())
556 }
557
558 fn availability(&self) -> LanguageModelAvailability {
559 self.model.availability()
560 }
561
562 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
563 self.model.tool_input_format()
564 }
565
566 fn max_token_count(&self) -> usize {
567 self.model.max_token_count()
568 }
569
570 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
571 match &self.model {
572 CloudModel::Anthropic(model) => {
573 model
574 .cache_configuration()
575 .map(|cache| LanguageModelCacheConfiguration {
576 max_cache_anchors: cache.max_cache_anchors,
577 should_speculate: cache.should_speculate,
578 min_total_token: cache.min_total_token,
579 })
580 }
581 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
582 }
583 }
584
585 fn count_tokens(
586 &self,
587 request: LanguageModelRequest,
588 cx: &App,
589 ) -> BoxFuture<'static, Result<usize>> {
590 match self.model.clone() {
591 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
592 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
593 CloudModel::Google(model) => {
594 let client = self.client.clone();
595 let request = into_google(request, model.id().into());
596 let request = google_ai::CountTokensRequest {
597 contents: request.contents,
598 };
599 async move {
600 let request = serde_json::to_string(&request)?;
601 let response = client
602 .request(proto::CountLanguageModelTokens {
603 provider: proto::LanguageModelProvider::Google as i32,
604 request,
605 })
606 .await?;
607 Ok(response.token_count as usize)
608 }
609 .boxed()
610 }
611 }
612 }
613
614 fn stream_completion(
615 &self,
616 request: LanguageModelRequest,
617 _cx: &AsyncApp,
618 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
619 match &self.model {
620 CloudModel::Anthropic(model) => {
621 let request = into_anthropic(
622 request,
623 model.request_id().into(),
624 model.default_temperature(),
625 model.max_output_tokens(),
626 model.mode(),
627 );
628 let client = self.client.clone();
629 let llm_api_token = self.llm_api_token.clone();
630 let future = self.request_limiter.stream(async move {
631 let response = Self::perform_llm_completion(
632 client.clone(),
633 llm_api_token,
634 PerformCompletionParams {
635 provider: client::LanguageModelProvider::Anthropic,
636 model: request.model.clone(),
637 provider_request: RawValue::from_string(serde_json::to_string(
638 &request,
639 )?)?,
640 },
641 )
642 .await?;
643 Ok(map_to_language_model_completion_events(Box::pin(
644 response_lines(response).map_err(AnthropicError::Other),
645 )))
646 });
647 async move { Ok(future.await?.boxed()) }.boxed()
648 }
649 CloudModel::OpenAi(model) => {
650 let client = self.client.clone();
651 let request = into_open_ai(request, model.id().into(), model.max_output_tokens());
652 let llm_api_token = self.llm_api_token.clone();
653 let future = self.request_limiter.stream(async move {
654 let response = Self::perform_llm_completion(
655 client.clone(),
656 llm_api_token,
657 PerformCompletionParams {
658 provider: client::LanguageModelProvider::OpenAi,
659 model: request.model.clone(),
660 provider_request: RawValue::from_string(serde_json::to_string(
661 &request,
662 )?)?,
663 },
664 )
665 .await?;
666 Ok(open_ai::extract_text_from_events(response_lines(response)))
667 });
668 async move {
669 Ok(future
670 .await?
671 .map(|result| result.map(LanguageModelCompletionEvent::Text))
672 .boxed())
673 }
674 .boxed()
675 }
676 CloudModel::Google(model) => {
677 let client = self.client.clone();
678 let request = into_google(request, model.id().into());
679 let llm_api_token = self.llm_api_token.clone();
680 let future = self.request_limiter.stream(async move {
681 let response = Self::perform_llm_completion(
682 client.clone(),
683 llm_api_token,
684 PerformCompletionParams {
685 provider: client::LanguageModelProvider::Google,
686 model: request.model.clone(),
687 provider_request: RawValue::from_string(serde_json::to_string(
688 &request,
689 )?)?,
690 },
691 )
692 .await?;
693 Ok(google_ai::extract_text_from_events(response_lines(
694 response,
695 )))
696 });
697 async move {
698 Ok(future
699 .await?
700 .map(|result| result.map(LanguageModelCompletionEvent::Text))
701 .boxed())
702 }
703 .boxed()
704 }
705 }
706 }
707
708 fn use_any_tool(
709 &self,
710 request: LanguageModelRequest,
711 tool_name: String,
712 tool_description: String,
713 input_schema: serde_json::Value,
714 _cx: &AsyncApp,
715 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
716 let client = self.client.clone();
717 let llm_api_token = self.llm_api_token.clone();
718
719 match &self.model {
720 CloudModel::Anthropic(model) => {
721 let mut request = into_anthropic(
722 request,
723 model.tool_model_id().into(),
724 model.default_temperature(),
725 model.max_output_tokens(),
726 model.mode(),
727 );
728 request.tool_choice = Some(anthropic::ToolChoice::Tool {
729 name: tool_name.clone(),
730 });
731 request.tools = vec![anthropic::Tool {
732 name: tool_name.clone(),
733 description: tool_description,
734 input_schema,
735 }];
736
737 self.request_limiter
738 .run(async move {
739 let response = Self::perform_llm_completion(
740 client.clone(),
741 llm_api_token,
742 PerformCompletionParams {
743 provider: client::LanguageModelProvider::Anthropic,
744 model: request.model.clone(),
745 provider_request: RawValue::from_string(serde_json::to_string(
746 &request,
747 )?)?,
748 },
749 )
750 .await?;
751
752 Ok(anthropic::extract_tool_args_from_events(
753 tool_name,
754 Box::pin(response_lines(response)),
755 )
756 .await?
757 .boxed())
758 })
759 .boxed()
760 }
761 CloudModel::OpenAi(model) => {
762 let mut request =
763 into_open_ai(request, model.id().into(), model.max_output_tokens());
764 request.tool_choice = Some(open_ai::ToolChoice::Other(
765 open_ai::ToolDefinition::Function {
766 function: open_ai::FunctionDefinition {
767 name: tool_name.clone(),
768 description: None,
769 parameters: None,
770 },
771 },
772 ));
773 request.tools = vec![open_ai::ToolDefinition::Function {
774 function: open_ai::FunctionDefinition {
775 name: tool_name.clone(),
776 description: Some(tool_description),
777 parameters: Some(input_schema),
778 },
779 }];
780
781 self.request_limiter
782 .run(async move {
783 let response = Self::perform_llm_completion(
784 client.clone(),
785 llm_api_token,
786 PerformCompletionParams {
787 provider: client::LanguageModelProvider::OpenAi,
788 model: request.model.clone(),
789 provider_request: RawValue::from_string(serde_json::to_string(
790 &request,
791 )?)?,
792 },
793 )
794 .await?;
795
796 Ok(open_ai::extract_tool_args_from_events(
797 tool_name,
798 Box::pin(response_lines(response)),
799 )
800 .await?
801 .boxed())
802 })
803 .boxed()
804 }
805 CloudModel::Google(_) => {
806 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
807 }
808 }
809 }
810}
811
812fn response_lines<T: DeserializeOwned>(
813 response: Response<AsyncBody>,
814) -> impl Stream<Item = Result<T>> {
815 futures::stream::try_unfold(
816 (String::new(), BufReader::new(response.into_body())),
817 move |(mut line, mut body)| async {
818 match body.read_line(&mut line).await {
819 Ok(0) => Ok(None),
820 Ok(_) => {
821 let event: T = serde_json::from_str(&line)?;
822 line.clear();
823 Ok(Some((event, (line, body))))
824 }
825 Err(e) => Err(e.into()),
826 }
827 },
828 )
829}
830
831struct ConfigurationView {
832 state: gpui::Entity<State>,
833}
834
835impl ConfigurationView {
836 fn authenticate(&mut self, cx: &mut Context<Self>) {
837 self.state.update(cx, |state, cx| {
838 state.authenticate(cx).detach_and_log_err(cx);
839 });
840 cx.notify();
841 }
842}
843
844impl Render for ConfigurationView {
845 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
846 const ZED_AI_URL: &str = "https://zed.dev/ai";
847
848 let is_connected = !self.state.read(cx).is_signed_out();
849 let plan = self.state.read(cx).user_store.read(cx).current_plan();
850 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
851
852 let is_pro = plan == Some(proto::Plan::ZedPro);
853 let subscription_text = Label::new(if is_pro {
854 "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."
855 } else {
856 "You have basic access to models from Anthropic through the Zed AI Free plan."
857 });
858 let manage_subscription_button = if is_pro {
859 Some(
860 h_flex().child(
861 Button::new("manage_settings", "Manage Subscription")
862 .style(ButtonStyle::Tinted(TintColor::Accent))
863 .on_click(
864 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
865 ),
866 ),
867 )
868 } else if cx.has_flag::<ZedPro>() {
869 Some(
870 h_flex()
871 .gap_2()
872 .child(
873 Button::new("learn_more", "Learn more")
874 .style(ButtonStyle::Subtle)
875 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
876 )
877 .child(
878 Button::new("upgrade", "Upgrade")
879 .style(ButtonStyle::Subtle)
880 .color(Color::Accent)
881 .on_click(
882 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
883 ),
884 ),
885 )
886 } else {
887 None
888 };
889
890 if is_connected {
891 v_flex()
892 .gap_3()
893 .w_full()
894 .children(render_accept_terms(
895 self.state.clone(),
896 LanguageModelProviderTosView::Configuration,
897 cx,
898 ))
899 .when(has_accepted_terms, |this| {
900 this.child(subscription_text)
901 .children(manage_subscription_button)
902 })
903 } else {
904 v_flex()
905 .gap_2()
906 .child(Label::new("Use Zed AI to access hosted language models."))
907 .child(
908 Button::new("sign_in", "Sign In")
909 .icon_color(Color::Muted)
910 .icon(IconName::Github)
911 .icon_position(IconPosition::Start)
912 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
913 )
914 }
915 }
916}