anthropic.rs

  1use crate::{
  2    settings::AllLanguageModelSettings, LanguageModel, LanguageModelCacheConfiguration,
  3    LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
  4    LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest, RateLimiter, Role,
  5};
  6use crate::{LanguageModelCompletionEvent, LanguageModelToolUse, StopReason};
  7use anthropic::{AnthropicError, ContentDelta, Event, ResponseContent};
  8use anyhow::{anyhow, Context as _, Result};
  9use collections::{BTreeMap, HashMap};
 10use editor::{Editor, EditorElement, EditorStyle};
 11use futures::Stream;
 12use futures::{future::BoxFuture, stream::BoxStream, FutureExt, StreamExt, TryStreamExt as _};
 13use gpui::{
 14    AnyView, AppContext, AsyncAppContext, FontStyle, ModelContext, Subscription, Task, TextStyle,
 15    View, WhiteSpace,
 16};
 17use http_client::HttpClient;
 18use schemars::JsonSchema;
 19use serde::{Deserialize, Serialize};
 20use settings::{Settings, SettingsStore};
 21use std::pin::Pin;
 22use std::str::FromStr;
 23use std::{sync::Arc, time::Duration};
 24use strum::IntoEnumIterator;
 25use theme::ThemeSettings;
 26use ui::{prelude::*, Icon, IconName, Tooltip};
 27use util::{maybe, ResultExt};
 28
 29const PROVIDER_ID: &str = "anthropic";
 30const PROVIDER_NAME: &str = "Anthropic";
 31
 32#[derive(Default, Clone, Debug, PartialEq)]
 33pub struct AnthropicSettings {
 34    pub api_url: String,
 35    pub low_speed_timeout: Option<Duration>,
 36    /// Extend Zed's list of Anthropic models.
 37    pub available_models: Vec<AvailableModel>,
 38    pub needs_setting_migration: bool,
 39}
 40
 41#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
 42pub struct AvailableModel {
 43    /// The model's name in the Anthropic API. e.g. claude-3-5-sonnet-20240620
 44    pub name: String,
 45    /// The model's name in Zed's UI, such as in the model selector dropdown menu in the assistant panel.
 46    pub display_name: Option<String>,
 47    /// The model's context window size.
 48    pub max_tokens: usize,
 49    /// A model `name` to substitute when calling tools, in case the primary model doesn't support tool calling.
 50    pub tool_override: Option<String>,
 51    /// Configuration of Anthropic's caching API.
 52    pub cache_configuration: Option<LanguageModelCacheConfiguration>,
 53    pub max_output_tokens: Option<u32>,
 54}
 55
 56pub struct AnthropicLanguageModelProvider {
 57    http_client: Arc<dyn HttpClient>,
 58    state: gpui::Model<State>,
 59}
 60
 61const ANTHROPIC_API_KEY_VAR: &str = "ANTHROPIC_API_KEY";
 62
 63pub struct State {
 64    api_key: Option<String>,
 65    api_key_from_env: bool,
 66    _subscription: Subscription,
 67}
 68
 69impl State {
 70    fn reset_api_key(&self, cx: &mut ModelContext<Self>) -> Task<Result<()>> {
 71        let delete_credentials =
 72            cx.delete_credentials(&AllLanguageModelSettings::get_global(cx).anthropic.api_url);
 73        cx.spawn(|this, mut cx| async move {
 74            delete_credentials.await.ok();
 75            this.update(&mut cx, |this, cx| {
 76                this.api_key = None;
 77                this.api_key_from_env = false;
 78                cx.notify();
 79            })
 80        })
 81    }
 82
 83    fn set_api_key(&mut self, api_key: String, cx: &mut ModelContext<Self>) -> Task<Result<()>> {
 84        let write_credentials = cx.write_credentials(
 85            AllLanguageModelSettings::get_global(cx)
 86                .anthropic
 87                .api_url
 88                .as_str(),
 89            "Bearer",
 90            api_key.as_bytes(),
 91        );
 92        cx.spawn(|this, mut cx| async move {
 93            write_credentials.await?;
 94
 95            this.update(&mut cx, |this, cx| {
 96                this.api_key = Some(api_key);
 97                cx.notify();
 98            })
 99        })
100    }
101
102    fn is_authenticated(&self) -> bool {
103        self.api_key.is_some()
104    }
105
106    fn authenticate(&self, cx: &mut ModelContext<Self>) -> Task<Result<()>> {
107        if self.is_authenticated() {
108            Task::ready(Ok(()))
109        } else {
110            let api_url = AllLanguageModelSettings::get_global(cx)
111                .anthropic
112                .api_url
113                .clone();
114
115            cx.spawn(|this, mut cx| async move {
116                let (api_key, from_env) = if let Ok(api_key) = std::env::var(ANTHROPIC_API_KEY_VAR)
117                {
118                    (api_key, true)
119                } else {
120                    let (_, api_key) = cx
121                        .update(|cx| cx.read_credentials(&api_url))?
122                        .await?
123                        .ok_or_else(|| anyhow!("credentials not found"))?;
124                    (String::from_utf8(api_key)?, false)
125                };
126
127                this.update(&mut cx, |this, cx| {
128                    this.api_key = Some(api_key);
129                    this.api_key_from_env = from_env;
130                    cx.notify();
131                })
132            })
133        }
134    }
135}
136
137impl AnthropicLanguageModelProvider {
138    pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut AppContext) -> Self {
139        let state = cx.new_model(|cx| State {
140            api_key: None,
141            api_key_from_env: false,
142            _subscription: cx.observe_global::<SettingsStore>(|_, cx| {
143                cx.notify();
144            }),
145        });
146
147        Self { http_client, state }
148    }
149}
150
151impl LanguageModelProviderState for AnthropicLanguageModelProvider {
152    type ObservableEntity = State;
153
154    fn observable_entity(&self) -> Option<gpui::Model<Self::ObservableEntity>> {
155        Some(self.state.clone())
156    }
157}
158
159impl LanguageModelProvider for AnthropicLanguageModelProvider {
160    fn id(&self) -> LanguageModelProviderId {
161        LanguageModelProviderId(PROVIDER_ID.into())
162    }
163
164    fn name(&self) -> LanguageModelProviderName {
165        LanguageModelProviderName(PROVIDER_NAME.into())
166    }
167
168    fn icon(&self) -> IconName {
169        IconName::AiAnthropic
170    }
171
172    fn provided_models(&self, cx: &AppContext) -> Vec<Arc<dyn LanguageModel>> {
173        let mut models = BTreeMap::default();
174
175        // Add base models from anthropic::Model::iter()
176        for model in anthropic::Model::iter() {
177            if !matches!(model, anthropic::Model::Custom { .. }) {
178                models.insert(model.id().to_string(), model);
179            }
180        }
181
182        // Override with available models from settings
183        for model in AllLanguageModelSettings::get_global(cx)
184            .anthropic
185            .available_models
186            .iter()
187        {
188            models.insert(
189                model.name.clone(),
190                anthropic::Model::Custom {
191                    name: model.name.clone(),
192                    display_name: model.display_name.clone(),
193                    max_tokens: model.max_tokens,
194                    tool_override: model.tool_override.clone(),
195                    cache_configuration: model.cache_configuration.as_ref().map(|config| {
196                        anthropic::AnthropicModelCacheConfiguration {
197                            max_cache_anchors: config.max_cache_anchors,
198                            should_speculate: config.should_speculate,
199                            min_total_token: config.min_total_token,
200                        }
201                    }),
202                    max_output_tokens: model.max_output_tokens,
203                },
204            );
205        }
206
207        models
208            .into_values()
209            .map(|model| {
210                Arc::new(AnthropicModel {
211                    id: LanguageModelId::from(model.id().to_string()),
212                    model,
213                    state: self.state.clone(),
214                    http_client: self.http_client.clone(),
215                    request_limiter: RateLimiter::new(4),
216                }) as Arc<dyn LanguageModel>
217            })
218            .collect()
219    }
220
221    fn is_authenticated(&self, cx: &AppContext) -> bool {
222        self.state.read(cx).is_authenticated()
223    }
224
225    fn authenticate(&self, cx: &mut AppContext) -> Task<Result<()>> {
226        self.state.update(cx, |state, cx| state.authenticate(cx))
227    }
228
229    fn configuration_view(&self, cx: &mut WindowContext) -> AnyView {
230        cx.new_view(|cx| ConfigurationView::new(self.state.clone(), cx))
231            .into()
232    }
233
234    fn reset_credentials(&self, cx: &mut AppContext) -> Task<Result<()>> {
235        self.state.update(cx, |state, cx| state.reset_api_key(cx))
236    }
237}
238
239pub struct AnthropicModel {
240    id: LanguageModelId,
241    model: anthropic::Model,
242    state: gpui::Model<State>,
243    http_client: Arc<dyn HttpClient>,
244    request_limiter: RateLimiter,
245}
246
247pub fn count_anthropic_tokens(
248    request: LanguageModelRequest,
249    cx: &AppContext,
250) -> BoxFuture<'static, Result<usize>> {
251    cx.background_executor()
252        .spawn(async move {
253            let messages = request.messages;
254            let mut tokens_from_images = 0;
255            let mut string_messages = Vec::with_capacity(messages.len());
256
257            for message in messages {
258                use crate::MessageContent;
259
260                let mut string_contents = String::new();
261
262                for content in message.content {
263                    match content {
264                        MessageContent::Text(text) => {
265                            string_contents.push_str(&text);
266                        }
267                        MessageContent::Image(image) => {
268                            tokens_from_images += image.estimate_tokens();
269                        }
270                        MessageContent::ToolUse(_tool_use) => {
271                            // TODO: Estimate token usage from tool uses.
272                        }
273                        MessageContent::ToolResult(tool_result) => {
274                            string_contents.push_str(&tool_result.content);
275                        }
276                    }
277                }
278
279                if !string_contents.is_empty() {
280                    string_messages.push(tiktoken_rs::ChatCompletionRequestMessage {
281                        role: match message.role {
282                            Role::User => "user".into(),
283                            Role::Assistant => "assistant".into(),
284                            Role::System => "system".into(),
285                        },
286                        content: Some(string_contents),
287                        name: None,
288                        function_call: None,
289                    });
290                }
291            }
292
293            // Tiktoken doesn't yet support these models, so we manually use the
294            // same tokenizer as GPT-4.
295            tiktoken_rs::num_tokens_from_messages("gpt-4", &string_messages)
296                .map(|tokens| tokens + tokens_from_images)
297        })
298        .boxed()
299}
300
301impl AnthropicModel {
302    fn stream_completion(
303        &self,
304        request: anthropic::Request,
305        cx: &AsyncAppContext,
306    ) -> BoxFuture<'static, Result<BoxStream<'static, Result<anthropic::Event, AnthropicError>>>>
307    {
308        let http_client = self.http_client.clone();
309
310        let Ok((api_key, api_url, low_speed_timeout)) = cx.read_model(&self.state, |state, cx| {
311            let settings = &AllLanguageModelSettings::get_global(cx).anthropic;
312            (
313                state.api_key.clone(),
314                settings.api_url.clone(),
315                settings.low_speed_timeout,
316            )
317        }) else {
318            return futures::future::ready(Err(anyhow!("App state dropped"))).boxed();
319        };
320
321        async move {
322            let api_key = api_key.ok_or_else(|| anyhow!("Missing Anthropic API Key"))?;
323            let request = anthropic::stream_completion(
324                http_client.as_ref(),
325                &api_url,
326                &api_key,
327                request,
328                low_speed_timeout,
329            );
330            request.await.context("failed to stream completion")
331        }
332        .boxed()
333    }
334}
335
336impl LanguageModel for AnthropicModel {
337    fn id(&self) -> LanguageModelId {
338        self.id.clone()
339    }
340
341    fn name(&self) -> LanguageModelName {
342        LanguageModelName::from(self.model.display_name().to_string())
343    }
344
345    fn provider_id(&self) -> LanguageModelProviderId {
346        LanguageModelProviderId(PROVIDER_ID.into())
347    }
348
349    fn provider_name(&self) -> LanguageModelProviderName {
350        LanguageModelProviderName(PROVIDER_NAME.into())
351    }
352
353    fn telemetry_id(&self) -> String {
354        format!("anthropic/{}", self.model.id())
355    }
356
357    fn max_token_count(&self) -> usize {
358        self.model.max_token_count()
359    }
360
361    fn max_output_tokens(&self) -> Option<u32> {
362        Some(self.model.max_output_tokens())
363    }
364
365    fn count_tokens(
366        &self,
367        request: LanguageModelRequest,
368        cx: &AppContext,
369    ) -> BoxFuture<'static, Result<usize>> {
370        count_anthropic_tokens(request, cx)
371    }
372
373    fn stream_completion(
374        &self,
375        request: LanguageModelRequest,
376        cx: &AsyncAppContext,
377    ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>> {
378        let request =
379            request.into_anthropic(self.model.id().into(), self.model.max_output_tokens());
380        let request = self.stream_completion(request, cx);
381        let future = self.request_limiter.stream(async move {
382            let response = request.await.map_err(|err| anyhow!(err))?;
383            Ok(map_to_language_model_completion_events(response))
384        });
385        async move { Ok(future.await?.boxed()) }.boxed()
386    }
387
388    fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
389        self.model
390            .cache_configuration()
391            .map(|config| LanguageModelCacheConfiguration {
392                max_cache_anchors: config.max_cache_anchors,
393                should_speculate: config.should_speculate,
394                min_total_token: config.min_total_token,
395            })
396    }
397
398    fn use_any_tool(
399        &self,
400        request: LanguageModelRequest,
401        tool_name: String,
402        tool_description: String,
403        input_schema: serde_json::Value,
404        cx: &AsyncAppContext,
405    ) -> BoxFuture<'static, Result<BoxStream<'static, Result<String>>>> {
406        let mut request = request.into_anthropic(
407            self.model.tool_model_id().into(),
408            self.model.max_output_tokens(),
409        );
410        request.tool_choice = Some(anthropic::ToolChoice::Tool {
411            name: tool_name.clone(),
412        });
413        request.tools = vec![anthropic::Tool {
414            name: tool_name.clone(),
415            description: tool_description,
416            input_schema,
417        }];
418
419        let response = self.stream_completion(request, cx);
420        self.request_limiter
421            .run(async move {
422                let response = response.await?;
423                Ok(anthropic::extract_tool_args_from_events(
424                    tool_name,
425                    Box::pin(response.map_err(|e| anyhow!(e))),
426                )
427                .await?
428                .boxed())
429            })
430            .boxed()
431    }
432}
433
434pub fn map_to_language_model_completion_events(
435    events: Pin<Box<dyn Send + Stream<Item = Result<Event, AnthropicError>>>>,
436) -> impl Stream<Item = Result<LanguageModelCompletionEvent>> {
437    struct RawToolUse {
438        id: String,
439        name: String,
440        input_json: String,
441    }
442
443    struct State {
444        events: Pin<Box<dyn Send + Stream<Item = Result<Event, AnthropicError>>>>,
445        tool_uses_by_index: HashMap<usize, RawToolUse>,
446    }
447
448    futures::stream::unfold(
449        State {
450            events,
451            tool_uses_by_index: HashMap::default(),
452        },
453        |mut state| async move {
454            while let Some(event) = state.events.next().await {
455                match event {
456                    Ok(event) => match event {
457                        Event::ContentBlockStart {
458                            index,
459                            content_block,
460                        } => match content_block {
461                            ResponseContent::Text { text } => {
462                                return Some((
463                                    Some(Ok(LanguageModelCompletionEvent::Text(text))),
464                                    state,
465                                ));
466                            }
467                            ResponseContent::ToolUse { id, name, .. } => {
468                                state.tool_uses_by_index.insert(
469                                    index,
470                                    RawToolUse {
471                                        id,
472                                        name,
473                                        input_json: String::new(),
474                                    },
475                                );
476
477                                return Some((None, state));
478                            }
479                        },
480                        Event::ContentBlockDelta { index, delta } => match delta {
481                            ContentDelta::TextDelta { text } => {
482                                return Some((
483                                    Some(Ok(LanguageModelCompletionEvent::Text(text))),
484                                    state,
485                                ));
486                            }
487                            ContentDelta::InputJsonDelta { partial_json } => {
488                                if let Some(tool_use) = state.tool_uses_by_index.get_mut(&index) {
489                                    tool_use.input_json.push_str(&partial_json);
490                                    return Some((None, state));
491                                }
492                            }
493                        },
494                        Event::ContentBlockStop { index } => {
495                            if let Some(tool_use) = state.tool_uses_by_index.remove(&index) {
496                                return Some((
497                                    Some(maybe!({
498                                        Ok(LanguageModelCompletionEvent::ToolUse(
499                                            LanguageModelToolUse {
500                                                id: tool_use.id,
501                                                name: tool_use.name,
502                                                input: serde_json::Value::from_str(
503                                                    &tool_use.input_json,
504                                                )
505                                                .map_err(|err| anyhow!(err))?,
506                                            },
507                                        ))
508                                    })),
509                                    state,
510                                ));
511                            }
512                        }
513                        Event::MessageDelta { delta, .. } => {
514                            if let Some(stop_reason) = delta.stop_reason.as_deref() {
515                                let stop_reason = match stop_reason {
516                                    "end_turn" => StopReason::EndTurn,
517                                    "max_tokens" => StopReason::MaxTokens,
518                                    "tool_use" => StopReason::ToolUse,
519                                    _ => StopReason::EndTurn,
520                                };
521
522                                return Some((
523                                    Some(Ok(LanguageModelCompletionEvent::Stop(stop_reason))),
524                                    state,
525                                ));
526                            }
527                        }
528                        Event::Error { error } => {
529                            return Some((
530                                Some(Err(anyhow!(AnthropicError::ApiError(error)))),
531                                state,
532                            ));
533                        }
534                        _ => {}
535                    },
536                    Err(err) => {
537                        return Some((Some(Err(anyhow!(err))), state));
538                    }
539                }
540            }
541
542            None
543        },
544    )
545    .filter_map(|event| async move { event })
546}
547
548struct ConfigurationView {
549    api_key_editor: View<Editor>,
550    state: gpui::Model<State>,
551    load_credentials_task: Option<Task<()>>,
552}
553
554impl ConfigurationView {
555    const PLACEHOLDER_TEXT: &'static str = "sk-ant-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
556
557    fn new(state: gpui::Model<State>, cx: &mut ViewContext<Self>) -> Self {
558        cx.observe(&state, |_, _, cx| {
559            cx.notify();
560        })
561        .detach();
562
563        let load_credentials_task = Some(cx.spawn({
564            let state = state.clone();
565            |this, mut cx| async move {
566                if let Some(task) = state
567                    .update(&mut cx, |state, cx| state.authenticate(cx))
568                    .log_err()
569                {
570                    // We don't log an error, because "not signed in" is also an error.
571                    let _ = task.await;
572                }
573                this.update(&mut cx, |this, cx| {
574                    this.load_credentials_task = None;
575                    cx.notify();
576                })
577                .log_err();
578            }
579        }));
580
581        Self {
582            api_key_editor: cx.new_view(|cx| {
583                let mut editor = Editor::single_line(cx);
584                editor.set_placeholder_text(Self::PLACEHOLDER_TEXT, cx);
585                editor
586            }),
587            state,
588            load_credentials_task,
589        }
590    }
591
592    fn save_api_key(&mut self, _: &menu::Confirm, cx: &mut ViewContext<Self>) {
593        let api_key = self.api_key_editor.read(cx).text(cx);
594        if api_key.is_empty() {
595            return;
596        }
597
598        let state = self.state.clone();
599        cx.spawn(|_, mut cx| async move {
600            state
601                .update(&mut cx, |state, cx| state.set_api_key(api_key, cx))?
602                .await
603        })
604        .detach_and_log_err(cx);
605
606        cx.notify();
607    }
608
609    fn reset_api_key(&mut self, cx: &mut ViewContext<Self>) {
610        self.api_key_editor
611            .update(cx, |editor, cx| editor.set_text("", cx));
612
613        let state = self.state.clone();
614        cx.spawn(|_, mut cx| async move {
615            state
616                .update(&mut cx, |state, cx| state.reset_api_key(cx))?
617                .await
618        })
619        .detach_and_log_err(cx);
620
621        cx.notify();
622    }
623
624    fn render_api_key_editor(&self, cx: &mut ViewContext<Self>) -> impl IntoElement {
625        let settings = ThemeSettings::get_global(cx);
626        let text_style = TextStyle {
627            color: cx.theme().colors().text,
628            font_family: settings.ui_font.family.clone(),
629            font_features: settings.ui_font.features.clone(),
630            font_fallbacks: settings.ui_font.fallbacks.clone(),
631            font_size: rems(0.875).into(),
632            font_weight: settings.ui_font.weight,
633            font_style: FontStyle::Normal,
634            line_height: relative(1.3),
635            background_color: None,
636            underline: None,
637            strikethrough: None,
638            white_space: WhiteSpace::Normal,
639            truncate: None,
640        };
641        EditorElement::new(
642            &self.api_key_editor,
643            EditorStyle {
644                background: cx.theme().colors().editor_background,
645                local_player: cx.theme().players().local(),
646                text: text_style,
647                ..Default::default()
648            },
649        )
650    }
651
652    fn should_render_editor(&self, cx: &mut ViewContext<Self>) -> bool {
653        !self.state.read(cx).is_authenticated()
654    }
655}
656
657impl Render for ConfigurationView {
658    fn render(&mut self, cx: &mut ViewContext<Self>) -> impl IntoElement {
659        const ANTHROPIC_CONSOLE_URL: &str = "https://console.anthropic.com/settings/keys";
660        const INSTRUCTIONS: [&str; 3] = [
661            "To use Zed's assistant with Anthropic, you need to add an API key. Follow these steps:",
662            "- Create one at:",
663            "- Paste your API key below and hit enter to use the assistant:",
664        ];
665        let env_var_set = self.state.read(cx).api_key_from_env;
666
667        if self.load_credentials_task.is_some() {
668            div().child(Label::new("Loading credentials...")).into_any()
669        } else if self.should_render_editor(cx) {
670            v_flex()
671                .size_full()
672                .on_action(cx.listener(Self::save_api_key))
673                .child(Label::new(INSTRUCTIONS[0]))
674                .child(h_flex().child(Label::new(INSTRUCTIONS[1])).child(
675                    Button::new("anthropic_console", ANTHROPIC_CONSOLE_URL)
676                        .style(ButtonStyle::Subtle)
677                        .icon(IconName::ExternalLink)
678                        .icon_size(IconSize::XSmall)
679                        .icon_color(Color::Muted)
680                        .on_click(move |_, cx| cx.open_url(ANTHROPIC_CONSOLE_URL))
681                    )
682                )
683                .child(Label::new(INSTRUCTIONS[2]))
684                .child(
685                    h_flex()
686                        .w_full()
687                        .my_2()
688                        .px_2()
689                        .py_1()
690                        .bg(cx.theme().colors().editor_background)
691                        .rounded_md()
692                        .child(self.render_api_key_editor(cx)),
693                )
694                .child(
695                    Label::new(
696                        format!("You can also assign the {ANTHROPIC_API_KEY_VAR} environment variable and restart Zed."),
697                    )
698                    .size(LabelSize::Small),
699                )
700                .into_any()
701        } else {
702            h_flex()
703                .size_full()
704                .justify_between()
705                .child(
706                    h_flex()
707                        .gap_1()
708                        .child(Icon::new(IconName::Check).color(Color::Success))
709                        .child(Label::new(if env_var_set {
710                            format!("API key set in {ANTHROPIC_API_KEY_VAR} environment variable.")
711                        } else {
712                            "API key configured.".to_string()
713                        })),
714                )
715                .child(
716                    Button::new("reset-key", "Reset key")
717                        .icon(Some(IconName::Trash))
718                        .icon_size(IconSize::Small)
719                        .icon_position(IconPosition::Start)
720                        .disabled(env_var_set)
721                        .when(env_var_set, |this| {
722                            this.tooltip(|cx| Tooltip::text(format!("To reset your API key, unset the {ANTHROPIC_API_KEY_VAR} environment variable."), cx))
723                        })
724                        .on_click(cx.listener(|this, _, cx| this.reset_api_key(cx))),
725                )
726                .into_any()
727        }
728    }
729}