1use super::open_ai::count_open_ai_tokens;
2use crate::{
3 settings::AllLanguageModelSettings, CloudModel, LanguageModel, LanguageModelId,
4 LanguageModelName, LanguageModelProviderId, LanguageModelProviderName,
5 LanguageModelProviderState, LanguageModelRequest, RateLimiter, ZedModel,
6};
7use anyhow::{anyhow, Context as _, Result};
8use client::{Client, PerformCompletionParams, UserStore, EXPIRED_LLM_TOKEN_HEADER_NAME};
9use collections::BTreeMap;
10use feature_flags::{FeatureFlag, FeatureFlagAppExt, LanguageModels};
11use futures::{future::BoxFuture, stream::BoxStream, AsyncBufReadExt, FutureExt, StreamExt};
12use gpui::{AnyView, AppContext, AsyncAppContext, Model, ModelContext, Subscription, Task};
13use http_client::{AsyncBody, HttpClient, Method, Response};
14use schemars::JsonSchema;
15use serde::{Deserialize, Serialize};
16use serde_json::value::RawValue;
17use settings::{Settings, SettingsStore};
18use smol::{
19 io::BufReader,
20 lock::{RwLock, RwLockUpgradableReadGuard, RwLockWriteGuard},
21};
22use std::{future, sync::Arc};
23use strum::IntoEnumIterator;
24use ui::prelude::*;
25
26use crate::{LanguageModelAvailability, LanguageModelProvider};
27
28use super::anthropic::count_anthropic_tokens;
29
30pub const PROVIDER_ID: &str = "zed.dev";
31pub const PROVIDER_NAME: &str = "Zed";
32
33#[derive(Default, Clone, Debug, PartialEq)]
34pub struct ZedDotDevSettings {
35 pub available_models: Vec<AvailableModel>,
36}
37
38#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
39#[serde(rename_all = "lowercase")]
40pub enum AvailableProvider {
41 Anthropic,
42 OpenAi,
43 Google,
44}
45
46#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
47pub struct AvailableModel {
48 provider: AvailableProvider,
49 name: String,
50 max_tokens: usize,
51 tool_override: Option<String>,
52}
53
54pub struct CloudLanguageModelProvider {
55 client: Arc<Client>,
56 llm_api_token: LlmApiToken,
57 state: gpui::Model<State>,
58 _maintain_client_status: Task<()>,
59}
60
61pub struct State {
62 client: Arc<Client>,
63 user_store: Model<UserStore>,
64 status: client::Status,
65 _subscription: Subscription,
66}
67
68impl State {
69 fn is_signed_out(&self) -> bool {
70 self.status.is_signed_out()
71 }
72
73 fn authenticate(&self, cx: &mut ModelContext<Self>) -> Task<Result<()>> {
74 let client = self.client.clone();
75 cx.spawn(move |this, mut cx| async move {
76 client.authenticate_and_connect(true, &cx).await?;
77 this.update(&mut cx, |_, cx| cx.notify())
78 })
79 }
80}
81
82impl CloudLanguageModelProvider {
83 pub fn new(user_store: Model<UserStore>, client: Arc<Client>, cx: &mut AppContext) -> Self {
84 let mut status_rx = client.status();
85 let status = *status_rx.borrow();
86
87 let state = cx.new_model(|cx| State {
88 client: client.clone(),
89 user_store,
90 status,
91 _subscription: cx.observe_global::<SettingsStore>(|_, cx| {
92 cx.notify();
93 }),
94 });
95
96 let state_ref = state.downgrade();
97 let maintain_client_status = cx.spawn(|mut cx| async move {
98 while let Some(status) = status_rx.next().await {
99 if let Some(this) = state_ref.upgrade() {
100 _ = this.update(&mut cx, |this, cx| {
101 if this.status != status {
102 this.status = status;
103 cx.notify();
104 }
105 });
106 } else {
107 break;
108 }
109 }
110 });
111
112 Self {
113 client,
114 state,
115 llm_api_token: LlmApiToken::default(),
116 _maintain_client_status: maintain_client_status,
117 }
118 }
119}
120
121impl LanguageModelProviderState for CloudLanguageModelProvider {
122 type ObservableEntity = State;
123
124 fn observable_entity(&self) -> Option<gpui::Model<Self::ObservableEntity>> {
125 Some(self.state.clone())
126 }
127}
128
129impl LanguageModelProvider for CloudLanguageModelProvider {
130 fn id(&self) -> LanguageModelProviderId {
131 LanguageModelProviderId(PROVIDER_ID.into())
132 }
133
134 fn name(&self) -> LanguageModelProviderName {
135 LanguageModelProviderName(PROVIDER_NAME.into())
136 }
137
138 fn icon(&self) -> IconName {
139 IconName::AiZed
140 }
141
142 fn provided_models(&self, cx: &AppContext) -> Vec<Arc<dyn LanguageModel>> {
143 let mut models = BTreeMap::default();
144
145 let is_user = !cx.has_flag::<LanguageModels>();
146 if is_user {
147 models.insert(
148 anthropic::Model::Claude3_5Sonnet.id().to_string(),
149 CloudModel::Anthropic(anthropic::Model::Claude3_5Sonnet),
150 );
151 } else {
152 for model in anthropic::Model::iter() {
153 if !matches!(model, anthropic::Model::Custom { .. }) {
154 models.insert(model.id().to_string(), CloudModel::Anthropic(model));
155 }
156 }
157 for model in open_ai::Model::iter() {
158 if !matches!(model, open_ai::Model::Custom { .. }) {
159 models.insert(model.id().to_string(), CloudModel::OpenAi(model));
160 }
161 }
162 for model in google_ai::Model::iter() {
163 if !matches!(model, google_ai::Model::Custom { .. }) {
164 models.insert(model.id().to_string(), CloudModel::Google(model));
165 }
166 }
167 for model in ZedModel::iter() {
168 models.insert(model.id().to_string(), CloudModel::Zed(model));
169 }
170
171 // Override with available models from settings
172 for model in &AllLanguageModelSettings::get_global(cx)
173 .zed_dot_dev
174 .available_models
175 {
176 let model = match model.provider {
177 AvailableProvider::Anthropic => {
178 CloudModel::Anthropic(anthropic::Model::Custom {
179 name: model.name.clone(),
180 max_tokens: model.max_tokens,
181 tool_override: model.tool_override.clone(),
182 })
183 }
184 AvailableProvider::OpenAi => CloudModel::OpenAi(open_ai::Model::Custom {
185 name: model.name.clone(),
186 max_tokens: model.max_tokens,
187 }),
188 AvailableProvider::Google => CloudModel::Google(google_ai::Model::Custom {
189 name: model.name.clone(),
190 max_tokens: model.max_tokens,
191 }),
192 };
193 models.insert(model.id().to_string(), model.clone());
194 }
195 }
196
197 models
198 .into_values()
199 .map(|model| {
200 Arc::new(CloudLanguageModel {
201 id: LanguageModelId::from(model.id().to_string()),
202 model,
203 llm_api_token: self.llm_api_token.clone(),
204 client: self.client.clone(),
205 request_limiter: RateLimiter::new(4),
206 }) as Arc<dyn LanguageModel>
207 })
208 .collect()
209 }
210
211 fn is_authenticated(&self, cx: &AppContext) -> bool {
212 !self.state.read(cx).is_signed_out()
213 }
214
215 fn authenticate(&self, _cx: &mut AppContext) -> Task<Result<()>> {
216 Task::ready(Ok(()))
217 }
218
219 fn configuration_view(&self, cx: &mut WindowContext) -> AnyView {
220 cx.new_view(|_cx| ConfigurationView {
221 state: self.state.clone(),
222 })
223 .into()
224 }
225
226 fn reset_credentials(&self, _cx: &mut AppContext) -> Task<Result<()>> {
227 Task::ready(Ok(()))
228 }
229}
230
231struct LlmServiceFeatureFlag;
232
233impl FeatureFlag for LlmServiceFeatureFlag {
234 const NAME: &'static str = "llm-service";
235
236 fn enabled_for_staff() -> bool {
237 false
238 }
239}
240
241pub struct CloudLanguageModel {
242 id: LanguageModelId,
243 model: CloudModel,
244 llm_api_token: LlmApiToken,
245 client: Arc<Client>,
246 request_limiter: RateLimiter,
247}
248
249#[derive(Clone, Default)]
250struct LlmApiToken(Arc<RwLock<Option<String>>>);
251
252impl CloudLanguageModel {
253 async fn perform_llm_completion(
254 client: Arc<Client>,
255 llm_api_token: LlmApiToken,
256 body: PerformCompletionParams,
257 ) -> Result<Response<AsyncBody>> {
258 let http_client = &client.http_client();
259
260 let mut token = llm_api_token.acquire(&client).await?;
261 let mut did_retry = false;
262
263 let response = loop {
264 let request = http_client::Request::builder()
265 .method(Method::POST)
266 .uri(http_client.build_zed_llm_url("/completion", &[])?.as_ref())
267 .header("Content-Type", "application/json")
268 .header("Authorization", format!("Bearer {token}"))
269 .body(serde_json::to_string(&body)?.into())?;
270 let response = http_client.send(request).await?;
271 if response.status().is_success() {
272 break response;
273 } else if !did_retry
274 && response
275 .headers()
276 .get(EXPIRED_LLM_TOKEN_HEADER_NAME)
277 .is_some()
278 {
279 did_retry = true;
280 token = llm_api_token.refresh(&client).await?;
281 } else {
282 break Err(anyhow!(
283 "cloud language model completion failed with status {}",
284 response.status()
285 ))?;
286 }
287 };
288
289 Ok(response)
290 }
291}
292
293impl LanguageModel for CloudLanguageModel {
294 fn id(&self) -> LanguageModelId {
295 self.id.clone()
296 }
297
298 fn name(&self) -> LanguageModelName {
299 LanguageModelName::from(self.model.display_name().to_string())
300 }
301
302 fn provider_id(&self) -> LanguageModelProviderId {
303 LanguageModelProviderId(PROVIDER_ID.into())
304 }
305
306 fn provider_name(&self) -> LanguageModelProviderName {
307 LanguageModelProviderName(PROVIDER_NAME.into())
308 }
309
310 fn telemetry_id(&self) -> String {
311 format!("zed.dev/{}", self.model.id())
312 }
313
314 fn availability(&self) -> LanguageModelAvailability {
315 self.model.availability()
316 }
317
318 fn max_token_count(&self) -> usize {
319 self.model.max_token_count()
320 }
321
322 fn count_tokens(
323 &self,
324 request: LanguageModelRequest,
325 cx: &AppContext,
326 ) -> BoxFuture<'static, Result<usize>> {
327 match self.model.clone() {
328 CloudModel::Anthropic(_) => count_anthropic_tokens(request, cx),
329 CloudModel::OpenAi(model) => count_open_ai_tokens(request, model, cx),
330 CloudModel::Google(model) => {
331 let client = self.client.clone();
332 let request = request.into_google(model.id().into());
333 let request = google_ai::CountTokensRequest {
334 contents: request.contents,
335 };
336 async move {
337 let request = serde_json::to_string(&request)?;
338 let response = client
339 .request(proto::CountLanguageModelTokens {
340 provider: proto::LanguageModelProvider::Google as i32,
341 request,
342 })
343 .await?;
344 Ok(response.token_count as usize)
345 }
346 .boxed()
347 }
348 CloudModel::Zed(_) => {
349 count_open_ai_tokens(request, open_ai::Model::ThreePointFiveTurbo, cx)
350 }
351 }
352 }
353
354 fn stream_completion(
355 &self,
356 request: LanguageModelRequest,
357 cx: &AsyncAppContext,
358 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
359 match &self.model {
360 CloudModel::Anthropic(model) => {
361 let request = request.into_anthropic(model.id().into());
362 let client = self.client.clone();
363
364 if cx
365 .update(|cx| cx.has_flag::<LlmServiceFeatureFlag>())
366 .unwrap_or(false)
367 {
368 let llm_api_token = self.llm_api_token.clone();
369 let future = self.request_limiter.stream(async move {
370 let response = Self::perform_llm_completion(
371 client.clone(),
372 llm_api_token,
373 PerformCompletionParams {
374 provider: client::LanguageModelProvider::Anthropic,
375 model: request.model.clone(),
376 provider_request: RawValue::from_string(serde_json::to_string(
377 &request,
378 )?)?,
379 },
380 )
381 .await?;
382 let body = BufReader::new(response.into_body());
383 let stream =
384 futures::stream::try_unfold(body, move |mut body| async move {
385 let mut buffer = String::new();
386 match body.read_line(&mut buffer).await {
387 Ok(0) => Ok(None),
388 Ok(_) => {
389 let event: anthropic::Event =
390 serde_json::from_str(&buffer)?;
391 Ok(Some((event, body)))
392 }
393 Err(e) => Err(e.into()),
394 }
395 });
396
397 Ok(anthropic::extract_text_from_events(stream))
398 });
399 async move { Ok(future.await?.boxed()) }.boxed()
400 } else {
401 let future = self.request_limiter.stream(async move {
402 let request = serde_json::to_string(&request)?;
403 let stream = client
404 .request_stream(proto::StreamCompleteWithLanguageModel {
405 provider: proto::LanguageModelProvider::Anthropic as i32,
406 request,
407 })
408 .await?
409 .map(|event| Ok(serde_json::from_str(&event?.event)?));
410 Ok(anthropic::extract_text_from_events(stream))
411 });
412 async move { Ok(future.await?.boxed()) }.boxed()
413 }
414 }
415 CloudModel::OpenAi(model) => {
416 let client = self.client.clone();
417 let request = request.into_open_ai(model.id().into());
418
419 if cx
420 .update(|cx| cx.has_flag::<LlmServiceFeatureFlag>())
421 .unwrap_or(false)
422 {
423 let llm_api_token = self.llm_api_token.clone();
424 let future = self.request_limiter.stream(async move {
425 let response = Self::perform_llm_completion(
426 client.clone(),
427 llm_api_token,
428 PerformCompletionParams {
429 provider: client::LanguageModelProvider::OpenAi,
430 model: request.model.clone(),
431 provider_request: RawValue::from_string(serde_json::to_string(
432 &request,
433 )?)?,
434 },
435 )
436 .await?;
437 let body = BufReader::new(response.into_body());
438 let stream =
439 futures::stream::try_unfold(body, move |mut body| async move {
440 let mut buffer = String::new();
441 match body.read_line(&mut buffer).await {
442 Ok(0) => Ok(None),
443 Ok(_) => {
444 let event: open_ai::ResponseStreamEvent =
445 serde_json::from_str(&buffer)?;
446 Ok(Some((event, body)))
447 }
448 Err(e) => Err(e.into()),
449 }
450 });
451
452 Ok(open_ai::extract_text_from_events(stream))
453 });
454 async move { Ok(future.await?.boxed()) }.boxed()
455 } else {
456 let future = self.request_limiter.stream(async move {
457 let request = serde_json::to_string(&request)?;
458 let stream = client
459 .request_stream(proto::StreamCompleteWithLanguageModel {
460 provider: proto::LanguageModelProvider::OpenAi as i32,
461 request,
462 })
463 .await?;
464 Ok(open_ai::extract_text_from_events(
465 stream.map(|item| Ok(serde_json::from_str(&item?.event)?)),
466 ))
467 });
468 async move { Ok(future.await?.boxed()) }.boxed()
469 }
470 }
471 CloudModel::Google(model) => {
472 let client = self.client.clone();
473 let request = request.into_google(model.id().into());
474
475 if cx
476 .update(|cx| cx.has_flag::<LlmServiceFeatureFlag>())
477 .unwrap_or(false)
478 {
479 let llm_api_token = self.llm_api_token.clone();
480 let future = self.request_limiter.stream(async move {
481 let response = Self::perform_llm_completion(
482 client.clone(),
483 llm_api_token,
484 PerformCompletionParams {
485 provider: client::LanguageModelProvider::Google,
486 model: request.model.clone(),
487 provider_request: RawValue::from_string(serde_json::to_string(
488 &request,
489 )?)?,
490 },
491 )
492 .await?;
493 let body = BufReader::new(response.into_body());
494 let stream =
495 futures::stream::try_unfold(body, move |mut body| async move {
496 let mut buffer = String::new();
497 match body.read_line(&mut buffer).await {
498 Ok(0) => Ok(None),
499 Ok(_) => {
500 let event: google_ai::GenerateContentResponse =
501 serde_json::from_str(&buffer)?;
502 Ok(Some((event, body)))
503 }
504 Err(e) => Err(e.into()),
505 }
506 });
507
508 Ok(google_ai::extract_text_from_events(stream))
509 });
510 async move { Ok(future.await?.boxed()) }.boxed()
511 } else {
512 let future = self.request_limiter.stream(async move {
513 let request = serde_json::to_string(&request)?;
514 let stream = client
515 .request_stream(proto::StreamCompleteWithLanguageModel {
516 provider: proto::LanguageModelProvider::Google as i32,
517 request,
518 })
519 .await?;
520 Ok(google_ai::extract_text_from_events(
521 stream.map(|item| Ok(serde_json::from_str(&item?.event)?)),
522 ))
523 });
524 async move { Ok(future.await?.boxed()) }.boxed()
525 }
526 }
527 CloudModel::Zed(model) => {
528 let client = self.client.clone();
529 let mut request = request.into_open_ai(model.id().into());
530 request.max_tokens = Some(4000);
531
532 if cx
533 .update(|cx| cx.has_flag::<LlmServiceFeatureFlag>())
534 .unwrap_or(false)
535 {
536 let llm_api_token = self.llm_api_token.clone();
537 let future = self.request_limiter.stream(async move {
538 let response = Self::perform_llm_completion(
539 client.clone(),
540 llm_api_token,
541 PerformCompletionParams {
542 provider: client::LanguageModelProvider::Zed,
543 model: request.model.clone(),
544 provider_request: RawValue::from_string(serde_json::to_string(
545 &request,
546 )?)?,
547 },
548 )
549 .await?;
550 let body = BufReader::new(response.into_body());
551 let stream =
552 futures::stream::try_unfold(body, move |mut body| async move {
553 let mut buffer = String::new();
554 match body.read_line(&mut buffer).await {
555 Ok(0) => Ok(None),
556 Ok(_) => {
557 let event: open_ai::ResponseStreamEvent =
558 serde_json::from_str(&buffer)?;
559 Ok(Some((event, body)))
560 }
561 Err(e) => Err(e.into()),
562 }
563 });
564
565 Ok(open_ai::extract_text_from_events(stream))
566 });
567 async move { Ok(future.await?.boxed()) }.boxed()
568 } else {
569 let future = self.request_limiter.stream(async move {
570 let request = serde_json::to_string(&request)?;
571 let stream = client
572 .request_stream(proto::StreamCompleteWithLanguageModel {
573 provider: proto::LanguageModelProvider::Zed as i32,
574 request,
575 })
576 .await?;
577 Ok(open_ai::extract_text_from_events(
578 stream.map(|item| Ok(serde_json::from_str(&item?.event)?)),
579 ))
580 });
581 async move { Ok(future.await?.boxed()) }.boxed()
582 }
583 }
584 }
585 }
586
587 fn use_any_tool(
588 &self,
589 request: LanguageModelRequest,
590 tool_name: String,
591 tool_description: String,
592 input_schema: serde_json::Value,
593 _cx: &AsyncAppContext,
594 ) -> BoxFuture<'static, Result<serde_json::Value>> {
595 match &self.model {
596 CloudModel::Anthropic(model) => {
597 let client = self.client.clone();
598 let mut request = request.into_anthropic(model.tool_model_id().into());
599 request.tool_choice = Some(anthropic::ToolChoice::Tool {
600 name: tool_name.clone(),
601 });
602 request.tools = vec![anthropic::Tool {
603 name: tool_name.clone(),
604 description: tool_description,
605 input_schema,
606 }];
607
608 self.request_limiter
609 .run(async move {
610 let request = serde_json::to_string(&request)?;
611 let response = client
612 .request(proto::CompleteWithLanguageModel {
613 provider: proto::LanguageModelProvider::Anthropic as i32,
614 request,
615 })
616 .await?;
617 let response: anthropic::Response =
618 serde_json::from_str(&response.completion)?;
619 response
620 .content
621 .into_iter()
622 .find_map(|content| {
623 if let anthropic::Content::ToolUse { name, input, .. } = content {
624 if name == tool_name {
625 Some(input)
626 } else {
627 None
628 }
629 } else {
630 None
631 }
632 })
633 .context("tool not used")
634 })
635 .boxed()
636 }
637 CloudModel::OpenAi(_) => {
638 future::ready(Err(anyhow!("tool use not implemented for OpenAI"))).boxed()
639 }
640 CloudModel::Google(_) => {
641 future::ready(Err(anyhow!("tool use not implemented for Google AI"))).boxed()
642 }
643 CloudModel::Zed(_) => {
644 future::ready(Err(anyhow!("tool use not implemented for Zed models"))).boxed()
645 }
646 }
647 }
648}
649
650impl LlmApiToken {
651 async fn acquire(&self, client: &Arc<Client>) -> Result<String> {
652 let lock = self.0.upgradable_read().await;
653 if let Some(token) = lock.as_ref() {
654 Ok(token.to_string())
655 } else {
656 Self::fetch(RwLockUpgradableReadGuard::upgrade(lock).await, &client).await
657 }
658 }
659
660 async fn refresh(&self, client: &Arc<Client>) -> Result<String> {
661 Self::fetch(self.0.write().await, &client).await
662 }
663
664 async fn fetch<'a>(
665 mut lock: RwLockWriteGuard<'a, Option<String>>,
666 client: &Arc<Client>,
667 ) -> Result<String> {
668 let response = client.request(proto::GetLlmToken {}).await?;
669 *lock = Some(response.token.clone());
670 Ok(response.token.clone())
671 }
672}
673
674struct ConfigurationView {
675 state: gpui::Model<State>,
676}
677
678impl ConfigurationView {
679 fn authenticate(&mut self, cx: &mut ViewContext<Self>) {
680 self.state.update(cx, |state, cx| {
681 state.authenticate(cx).detach_and_log_err(cx);
682 });
683 cx.notify();
684 }
685}
686
687impl Render for ConfigurationView {
688 fn render(&mut self, cx: &mut ViewContext<Self>) -> impl IntoElement {
689 const ZED_AI_URL: &str = "https://zed.dev/ai";
690 const ACCOUNT_SETTINGS_URL: &str = "https://zed.dev/account";
691
692 let is_connected = !self.state.read(cx).is_signed_out();
693 let plan = self.state.read(cx).user_store.read(cx).current_plan();
694
695 let is_pro = plan == Some(proto::Plan::ZedPro);
696
697 if is_connected {
698 v_flex()
699 .gap_3()
700 .max_w_4_5()
701 .child(Label::new(
702 if is_pro {
703 "You have full access to Zed's hosted models from Anthropic, OpenAI, Google with faster speeds and higher limits through Zed Pro."
704 } else {
705 "You have basic access to models from Anthropic, OpenAI, Google and more through the Zed AI Free plan."
706 }))
707 .child(
708 if is_pro {
709 h_flex().child(
710 Button::new("manage_settings", "Manage Subscription")
711 .style(ButtonStyle::Filled)
712 .on_click(cx.listener(|_, _, cx| {
713 cx.open_url(ACCOUNT_SETTINGS_URL)
714 })))
715 } else {
716 h_flex()
717 .gap_2()
718 .child(
719 Button::new("learn_more", "Learn more")
720 .style(ButtonStyle::Subtle)
721 .on_click(cx.listener(|_, _, cx| {
722 cx.open_url(ZED_AI_URL)
723 })))
724 .child(
725 Button::new("upgrade", "Upgrade")
726 .style(ButtonStyle::Subtle)
727 .color(Color::Accent)
728 .on_click(cx.listener(|_, _, cx| {
729 cx.open_url(ACCOUNT_SETTINGS_URL)
730 })))
731 },
732 )
733 } else {
734 v_flex()
735 .gap_6()
736 .child(Label::new("Use the zed.dev to access language models."))
737 .child(
738 v_flex()
739 .gap_2()
740 .child(
741 Button::new("sign_in", "Sign in")
742 .icon_color(Color::Muted)
743 .icon(IconName::Github)
744 .icon_position(IconPosition::Start)
745 .style(ButtonStyle::Filled)
746 .full_width()
747 .on_click(cx.listener(move |this, _, cx| this.authenticate(cx))),
748 )
749 .child(
750 div().flex().w_full().items_center().child(
751 Label::new("Sign in to enable collaboration.")
752 .color(Color::Muted)
753 .size(LabelSize::Small),
754 ),
755 ),
756 )
757 }
758 }
759}