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 thread_view = match view_kind {
414 LanguageModelProviderTosView::ThreadEmptyState => true,
415 LanguageModelProviderTosView::PromptEditorPopup => false,
416 LanguageModelProviderTosView::Configuration => false,
417 };
418
419 let form = v_flex()
420 .w_full()
421 .gap_2()
422 .child(
423 h_flex()
424 .flex_wrap()
425 .when(thread_view, |this| this.justify_center())
426 .child(Label::new(
427 "To start using Zed AI, please read and accept the",
428 ))
429 .child(terms_button),
430 )
431 .child({
432 let button_container = h_flex().w_full().child(
433 Button::new("accept_terms", "I accept the Terms of Service")
434 .style(ButtonStyle::Tinted(TintColor::Accent))
435 .icon(IconName::Check)
436 .icon_position(IconPosition::Start)
437 .icon_size(IconSize::Small)
438 .full_width()
439 .disabled(accept_terms_disabled)
440 .on_click({
441 let state = state.downgrade();
442 move |_, _window, cx| {
443 state
444 .update(cx, |state, cx| state.accept_terms_of_service(cx))
445 .ok();
446 }
447 }),
448 );
449
450 match view_kind {
451 LanguageModelProviderTosView::PromptEditorPopup => button_container.justify_end(),
452 LanguageModelProviderTosView::Configuration => button_container.justify_start(),
453 LanguageModelProviderTosView::ThreadEmptyState => button_container.justify_center(),
454 }
455 });
456
457 Some(form.into_any())
458}
459
460pub struct CloudLanguageModel {
461 id: LanguageModelId,
462 model: CloudModel,
463 llm_api_token: LlmApiToken,
464 client: Arc<Client>,
465 request_limiter: RateLimiter,
466}
467
468impl CloudLanguageModel {
469 const MAX_RETRIES: usize = 3;
470
471 async fn perform_llm_completion(
472 client: Arc<Client>,
473 llm_api_token: LlmApiToken,
474 body: PerformCompletionParams,
475 ) -> Result<Response<AsyncBody>> {
476 let http_client = &client.http_client();
477
478 let mut token = llm_api_token.acquire(&client).await?;
479 let mut retries_remaining = Self::MAX_RETRIES;
480 let mut retry_delay = Duration::from_secs(1);
481
482 loop {
483 let request_builder = http_client::Request::builder();
484 let request = request_builder
485 .method(Method::POST)
486 .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
487 .header("Content-Type", "application/json")
488 .header("Authorization", format!("Bearer {token}"))
489 .body(serde_json::to_string(&body)?.into())?;
490 let mut response = http_client.send(request).await?;
491 let status = response.status();
492 if status.is_success() {
493 return Ok(response);
494 } else if response
495 .headers()
496 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
497 .is_some()
498 {
499 retries_remaining -= 1;
500 token = llm_api_token.refresh(&client).await?;
501 } else if status == StatusCode::FORBIDDEN
502 && response
503 .headers()
504 .get(MAX_LLM_MONTHLY_SPEND_REACHED_HEADER_NAME)
505 .is_some()
506 {
507 return Err(anyhow!(MaxMonthlySpendReachedError));
508 } else if status.as_u16() >= 500 && status.as_u16() < 600 {
509 // If we encounter an error in the 500 range, retry after a delay.
510 // We've seen at least these in the wild from API providers:
511 // * 500 Internal Server Error
512 // * 502 Bad Gateway
513 // * 529 Service Overloaded
514
515 if retries_remaining == 0 {
516 let mut body = String::new();
517 response.body_mut().read_to_string(&mut body).await?;
518 return Err(anyhow!(
519 "cloud language model completion failed after {} retries with status {status}: {body}",
520 Self::MAX_RETRIES
521 ));
522 }
523
524 Timer::after(retry_delay).await;
525
526 retries_remaining -= 1;
527 retry_delay *= 2; // If it fails again, wait longer.
528 } else if status == StatusCode::PAYMENT_REQUIRED {
529 return Err(anyhow!(PaymentRequiredError));
530 } else {
531 let mut body = String::new();
532 response.body_mut().read_to_string(&mut body).await?;
533 return Err(anyhow!(
534 "cloud language model completion failed with status {status}: {body}",
535 ));
536 }
537 }
538 }
539}
540
541impl LanguageModel for CloudLanguageModel {
542 fn id(&self) -> LanguageModelId {
543 self.id.clone()
544 }
545
546 fn name(&self) -> LanguageModelName {
547 LanguageModelName::from(self.model.display_name().to_string())
548 }
549
550 fn icon(&self) -> Option<IconName> {
551 self.model.icon()
552 }
553
554 fn provider_id(&self) -> LanguageModelProviderId {
555 LanguageModelProviderId(ZED_CLOUD_PROVIDER_ID.into())
556 }
557
558 fn provider_name(&self) -> LanguageModelProviderName {
559 LanguageModelProviderName(PROVIDER_NAME.into())
560 }
561
562 fn telemetry_id(&self) -> String {
563 format!("zed.dev/{}", self.model.id())
564 }
565
566 fn availability(&self) -> LanguageModelAvailability {
567 self.model.availability()
568 }
569
570 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
571 self.model.tool_input_format()
572 }
573
574 fn max_token_count(&self) -> usize {
575 self.model.max_token_count()
576 }
577
578 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
579 match &self.model {
580 CloudModel::Anthropic(model) => {
581 model
582 .cache_configuration()
583 .map(|cache| LanguageModelCacheConfiguration {
584 max_cache_anchors: cache.max_cache_anchors,
585 should_speculate: cache.should_speculate,
586 min_total_token: cache.min_total_token,
587 })
588 }
589 CloudModel::OpenAi(_) | CloudModel::Google(_) => None,
590 }
591 }
592
593 fn count_tokens(
594 &self,
595 request: LanguageModelRequest,
596 cx: &App,
597 ) -> BoxFuture<'static, Result<usize>> {
598 match self.model.clone() {
599 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
600 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
601 CloudModel::Google(model) => {
602 let client = self.client.clone();
603 let request = into_google(request, model.id().into());
604 let request = google_ai::CountTokensRequest {
605 contents: request.contents,
606 };
607 async move {
608 let request = serde_json::to_string(&request)?;
609 let response = client
610 .request(proto::CountLanguageModelTokens {
611 provider: proto::LanguageModelProvider::Google as i32,
612 request,
613 })
614 .await?;
615 Ok(response.token_count as usize)
616 }
617 .boxed()
618 }
619 }
620 }
621
622 fn stream_completion(
623 &self,
624 request: LanguageModelRequest,
625 _cx: &AsyncApp,
626 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
627 match &self.model {
628 CloudModel::Anthropic(model) => {
629 let request = into_anthropic(
630 request,
631 model.request_id().into(),
632 model.default_temperature(),
633 model.max_output_tokens(),
634 model.mode(),
635 );
636 let client = self.client.clone();
637 let llm_api_token = self.llm_api_token.clone();
638 let future = self.request_limiter.stream(async move {
639 let response = Self::perform_llm_completion(
640 client.clone(),
641 llm_api_token,
642 PerformCompletionParams {
643 provider: client::LanguageModelProvider::Anthropic,
644 model: request.model.clone(),
645 provider_request: RawValue::from_string(serde_json::to_string(
646 &request,
647 )?)?,
648 },
649 )
650 .await?;
651 Ok(map_to_language_model_completion_events(Box::pin(
652 response_lines(response).map_err(AnthropicError::Other),
653 )))
654 });
655 async move { Ok(future.await?.boxed()) }.boxed()
656 }
657 CloudModel::OpenAi(model) => {
658 let client = self.client.clone();
659 let request = into_open_ai(request, model.id().into(), model.max_output_tokens());
660 let llm_api_token = self.llm_api_token.clone();
661 let future = self.request_limiter.stream(async move {
662 let response = Self::perform_llm_completion(
663 client.clone(),
664 llm_api_token,
665 PerformCompletionParams {
666 provider: client::LanguageModelProvider::OpenAi,
667 model: request.model.clone(),
668 provider_request: RawValue::from_string(serde_json::to_string(
669 &request,
670 )?)?,
671 },
672 )
673 .await?;
674 Ok(open_ai::extract_text_from_events(response_lines(response)))
675 });
676 async move {
677 Ok(future
678 .await?
679 .map(|result| result.map(LanguageModelCompletionEvent::Text))
680 .boxed())
681 }
682 .boxed()
683 }
684 CloudModel::Google(model) => {
685 let client = self.client.clone();
686 let request = into_google(request, model.id().into());
687 let llm_api_token = self.llm_api_token.clone();
688 let future = self.request_limiter.stream(async move {
689 let response = Self::perform_llm_completion(
690 client.clone(),
691 llm_api_token,
692 PerformCompletionParams {
693 provider: client::LanguageModelProvider::Google,
694 model: request.model.clone(),
695 provider_request: RawValue::from_string(serde_json::to_string(
696 &request,
697 )?)?,
698 },
699 )
700 .await?;
701 Ok(google_ai::extract_text_from_events(response_lines(
702 response,
703 )))
704 });
705 async move {
706 Ok(future
707 .await?
708 .map(|result| result.map(LanguageModelCompletionEvent::Text))
709 .boxed())
710 }
711 .boxed()
712 }
713 }
714 }
715
716 fn use_any_tool(
717 &self,
718 request: LanguageModelRequest,
719 tool_name: String,
720 tool_description: String,
721 input_schema: serde_json::Value,
722 _cx: &AsyncApp,
723 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
724 let client = self.client.clone();
725 let llm_api_token = self.llm_api_token.clone();
726
727 match &self.model {
728 CloudModel::Anthropic(model) => {
729 let mut request = into_anthropic(
730 request,
731 model.tool_model_id().into(),
732 model.default_temperature(),
733 model.max_output_tokens(),
734 model.mode(),
735 );
736 request.tool_choice = Some(anthropic::ToolChoice::Tool {
737 name: tool_name.clone(),
738 });
739 request.tools = vec![anthropic::Tool {
740 name: tool_name.clone(),
741 description: tool_description,
742 input_schema,
743 }];
744
745 self.request_limiter
746 .run(async move {
747 let response = Self::perform_llm_completion(
748 client.clone(),
749 llm_api_token,
750 PerformCompletionParams {
751 provider: client::LanguageModelProvider::Anthropic,
752 model: request.model.clone(),
753 provider_request: RawValue::from_string(serde_json::to_string(
754 &request,
755 )?)?,
756 },
757 )
758 .await?;
759
760 Ok(anthropic::extract_tool_args_from_events(
761 tool_name,
762 Box::pin(response_lines(response)),
763 )
764 .await?
765 .boxed())
766 })
767 .boxed()
768 }
769 CloudModel::OpenAi(model) => {
770 let mut request =
771 into_open_ai(request, model.id().into(), model.max_output_tokens());
772 request.tool_choice = Some(open_ai::ToolChoice::Other(
773 open_ai::ToolDefinition::Function {
774 function: open_ai::FunctionDefinition {
775 name: tool_name.clone(),
776 description: None,
777 parameters: None,
778 },
779 },
780 ));
781 request.tools = vec![open_ai::ToolDefinition::Function {
782 function: open_ai::FunctionDefinition {
783 name: tool_name.clone(),
784 description: Some(tool_description),
785 parameters: Some(input_schema),
786 },
787 }];
788
789 self.request_limiter
790 .run(async move {
791 let response = Self::perform_llm_completion(
792 client.clone(),
793 llm_api_token,
794 PerformCompletionParams {
795 provider: client::LanguageModelProvider::OpenAi,
796 model: request.model.clone(),
797 provider_request: RawValue::from_string(serde_json::to_string(
798 &request,
799 )?)?,
800 },
801 )
802 .await?;
803
804 Ok(open_ai::extract_tool_args_from_events(
805 tool_name,
806 Box::pin(response_lines(response)),
807 )
808 .await?
809 .boxed())
810 })
811 .boxed()
812 }
813 CloudModel::Google(_) => {
814 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
815 }
816 }
817 }
818}
819
820fn response_lines<T: DeserializeOwned>(
821 response: Response<AsyncBody>,
822) -> impl Stream<Item = Result<T>> {
823 futures::stream::try_unfold(
824 (String::new(), BufReader::new(response.into_body())),
825 move |(mut line, mut body)| async {
826 match body.read_line(&mut line).await {
827 Ok(0) => Ok(None),
828 Ok(_) => {
829 let event: T = serde_json::from_str(&line)?;
830 line.clear();
831 Ok(Some((event, (line, body))))
832 }
833 Err(e) => Err(e.into()),
834 }
835 },
836 )
837}
838
839struct ConfigurationView {
840 state: gpui::Entity<State>,
841}
842
843impl ConfigurationView {
844 fn authenticate(&mut self, cx: &mut Context<Self>) {
845 self.state.update(cx, |state, cx| {
846 state.authenticate(cx).detach_and_log_err(cx);
847 });
848 cx.notify();
849 }
850}
851
852impl Render for ConfigurationView {
853 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
854 const ZED_AI_URL: &str = "https://zed.dev/ai";
855
856 let is_connected = !self.state.read(cx).is_signed_out();
857 let plan = self.state.read(cx).user_store.read(cx).current_plan();
858 let has_accepted_terms = self.state.read(cx).has_accepted_terms_of_service(cx);
859
860 let is_pro = plan == Some(proto::Plan::ZedPro);
861 let subscription_text = Label::new(if is_pro {
862 "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."
863 } else {
864 "You have basic access to models from Anthropic through the Zed AI Free plan."
865 });
866 let manage_subscription_button = if is_pro {
867 Some(
868 h_flex().child(
869 Button::new("manage_settings", "Manage Subscription")
870 .style(ButtonStyle::Tinted(TintColor::Accent))
871 .on_click(
872 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
873 ),
874 ),
875 )
876 } else if cx.has_flag::<ZedPro>() {
877 Some(
878 h_flex()
879 .gap_2()
880 .child(
881 Button::new("learn_more", "Learn more")
882 .style(ButtonStyle::Subtle)
883 .on_click(cx.listener(|_, _, _, cx| cx.open_url(ZED_AI_URL))),
884 )
885 .child(
886 Button::new("upgrade", "Upgrade")
887 .style(ButtonStyle::Subtle)
888 .color(Color::Accent)
889 .on_click(
890 cx.listener(|_, _, _, cx| cx.open_url(&zed_urls::account_url(cx))),
891 ),
892 ),
893 )
894 } else {
895 None
896 };
897
898 if is_connected {
899 v_flex()
900 .gap_3()
901 .w_full()
902 .children(render_accept_terms(
903 self.state.clone(),
904 LanguageModelProviderTosView::Configuration,
905 cx,
906 ))
907 .when(has_accepted_terms, |this| {
908 this.child(subscription_text)
909 .children(manage_subscription_button)
910 })
911 } else {
912 v_flex()
913 .gap_2()
914 .child(Label::new("Use Zed AI to access hosted language models."))
915 .child(
916 Button::new("sign_in", "Sign In")
917 .icon_color(Color::Muted)
918 .icon(IconName::Github)
919 .icon_position(IconPosition::Start)
920 .on_click(cx.listener(move |this, _, _, cx| this.authenticate(cx))),
921 )
922 }
923 }
924}