1use anyhow::{Context as _, Result, anyhow};
2use collections::BTreeMap;
3use credentials_provider::CredentialsProvider;
4use editor::{Editor, EditorElement, EditorStyle};
5use futures::{FutureExt, Stream, StreamExt, future::BoxFuture};
6use google_ai::{
7 FunctionDeclaration, GenerateContentResponse, GoogleModelMode, Part, SystemInstruction,
8 ThinkingConfig, UsageMetadata,
9};
10use gpui::{
11 AnyView, App, AsyncApp, Context, Entity, FontStyle, Subscription, Task, TextStyle, WhiteSpace,
12};
13use http_client::HttpClient;
14use language_model::{
15 AuthenticateError, ConfigurationViewTargetAgent, LanguageModelCompletionError,
16 LanguageModelCompletionEvent, LanguageModelToolChoice, LanguageModelToolSchemaFormat,
17 LanguageModelToolUse, LanguageModelToolUseId, MessageContent, StopReason,
18};
19use language_model::{
20 LanguageModel, LanguageModelId, LanguageModelName, LanguageModelProvider,
21 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
22 LanguageModelRequest, RateLimiter, Role,
23};
24use schemars::JsonSchema;
25use serde::{Deserialize, Serialize};
26use settings::{Settings, SettingsStore};
27use std::pin::Pin;
28use std::sync::{
29 Arc,
30 atomic::{self, AtomicU64},
31};
32use strum::IntoEnumIterator;
33use theme::ThemeSettings;
34use ui::{Icon, IconName, List, Tooltip, prelude::*};
35use util::ResultExt;
36
37use crate::AllLanguageModelSettings;
38use crate::ui::InstructionListItem;
39
40use super::anthropic::ApiKey;
41
42const PROVIDER_ID: LanguageModelProviderId = language_model::GOOGLE_PROVIDER_ID;
43const PROVIDER_NAME: LanguageModelProviderName = language_model::GOOGLE_PROVIDER_NAME;
44
45#[derive(Default, Clone, Debug, PartialEq)]
46pub struct GoogleSettings {
47 pub api_url: String,
48 pub available_models: Vec<AvailableModel>,
49}
50
51#[derive(Clone, Copy, Debug, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
52#[serde(tag = "type", rename_all = "lowercase")]
53pub enum ModelMode {
54 #[default]
55 Default,
56 Thinking {
57 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
58 budget_tokens: Option<u32>,
59 },
60}
61
62impl From<ModelMode> for GoogleModelMode {
63 fn from(value: ModelMode) -> Self {
64 match value {
65 ModelMode::Default => GoogleModelMode::Default,
66 ModelMode::Thinking { budget_tokens } => GoogleModelMode::Thinking { budget_tokens },
67 }
68 }
69}
70
71impl From<GoogleModelMode> for ModelMode {
72 fn from(value: GoogleModelMode) -> Self {
73 match value {
74 GoogleModelMode::Default => ModelMode::Default,
75 GoogleModelMode::Thinking { budget_tokens } => ModelMode::Thinking { budget_tokens },
76 }
77 }
78}
79
80#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
81pub struct AvailableModel {
82 name: String,
83 display_name: Option<String>,
84 max_tokens: u64,
85 mode: Option<ModelMode>,
86}
87
88pub struct GoogleLanguageModelProvider {
89 http_client: Arc<dyn HttpClient>,
90 state: gpui::Entity<State>,
91}
92
93pub struct State {
94 api_key: Option<String>,
95 api_key_from_env: bool,
96 _subscription: Subscription,
97}
98
99const GEMINI_API_KEY_VAR: &str = "GEMINI_API_KEY";
100const GOOGLE_AI_API_KEY_VAR: &str = "GOOGLE_AI_API_KEY";
101
102impl State {
103 fn is_authenticated(&self) -> bool {
104 self.api_key.is_some()
105 }
106
107 fn reset_api_key(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
108 let credentials_provider = <dyn CredentialsProvider>::global(cx);
109 let api_url = AllLanguageModelSettings::get_global(cx)
110 .google
111 .api_url
112 .clone();
113 cx.spawn(async move |this, cx| {
114 credentials_provider
115 .delete_credentials(&api_url, cx)
116 .await
117 .log_err();
118 this.update(cx, |this, cx| {
119 this.api_key = None;
120 this.api_key_from_env = false;
121 cx.notify();
122 })
123 })
124 }
125
126 fn set_api_key(&mut self, api_key: String, cx: &mut Context<Self>) -> Task<Result<()>> {
127 let credentials_provider = <dyn CredentialsProvider>::global(cx);
128 let api_url = AllLanguageModelSettings::get_global(cx)
129 .google
130 .api_url
131 .clone();
132 cx.spawn(async move |this, cx| {
133 credentials_provider
134 .write_credentials(&api_url, "Bearer", api_key.as_bytes(), cx)
135 .await?;
136 this.update(cx, |this, cx| {
137 this.api_key = Some(api_key);
138 cx.notify();
139 })
140 })
141 }
142
143 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
144 if self.is_authenticated() {
145 return Task::ready(Ok(()));
146 }
147
148 let credentials_provider = <dyn CredentialsProvider>::global(cx);
149 let api_url = AllLanguageModelSettings::get_global(cx)
150 .google
151 .api_url
152 .clone();
153
154 cx.spawn(async move |this, cx| {
155 let (api_key, from_env) = if let Ok(api_key) = std::env::var(GOOGLE_AI_API_KEY_VAR) {
156 (api_key, true)
157 } else if let Ok(api_key) = std::env::var(GEMINI_API_KEY_VAR) {
158 (api_key, true)
159 } else {
160 let (_, api_key) = credentials_provider
161 .read_credentials(&api_url, cx)
162 .await?
163 .ok_or(AuthenticateError::CredentialsNotFound)?;
164 (
165 String::from_utf8(api_key).context("invalid {PROVIDER_NAME} API key")?,
166 false,
167 )
168 };
169
170 this.update(cx, |this, cx| {
171 this.api_key = Some(api_key);
172 this.api_key_from_env = from_env;
173 cx.notify();
174 })?;
175
176 Ok(())
177 })
178 }
179}
180
181impl GoogleLanguageModelProvider {
182 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
183 let state = cx.new(|cx| State {
184 api_key: None,
185 api_key_from_env: false,
186 _subscription: cx.observe_global::<SettingsStore>(|_, cx| {
187 cx.notify();
188 }),
189 });
190
191 Self { http_client, state }
192 }
193
194 fn create_language_model(&self, model: google_ai::Model) -> Arc<dyn LanguageModel> {
195 Arc::new(GoogleLanguageModel {
196 id: LanguageModelId::from(model.id().to_string()),
197 model,
198 state: self.state.clone(),
199 http_client: self.http_client.clone(),
200 request_limiter: RateLimiter::new(4),
201 })
202 }
203
204 pub fn api_key(cx: &mut App) -> Task<Result<ApiKey>> {
205 let credentials_provider = <dyn CredentialsProvider>::global(cx);
206 let api_url = AllLanguageModelSettings::get_global(cx)
207 .google
208 .api_url
209 .clone();
210
211 if let Ok(key) = std::env::var(GEMINI_API_KEY_VAR) {
212 Task::ready(Ok(ApiKey {
213 key,
214 from_env: true,
215 }))
216 } else {
217 cx.spawn(async move |cx| {
218 let (_, api_key) = credentials_provider
219 .read_credentials(&api_url, cx)
220 .await?
221 .ok_or(AuthenticateError::CredentialsNotFound)?;
222
223 Ok(ApiKey {
224 key: String::from_utf8(api_key).context("invalid {PROVIDER_NAME} API key")?,
225 from_env: false,
226 })
227 })
228 }
229 }
230}
231
232impl LanguageModelProviderState for GoogleLanguageModelProvider {
233 type ObservableEntity = State;
234
235 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
236 Some(self.state.clone())
237 }
238}
239
240impl LanguageModelProvider for GoogleLanguageModelProvider {
241 fn id(&self) -> LanguageModelProviderId {
242 PROVIDER_ID
243 }
244
245 fn name(&self) -> LanguageModelProviderName {
246 PROVIDER_NAME
247 }
248
249 fn icon(&self) -> IconName {
250 IconName::AiGoogle
251 }
252
253 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
254 Some(self.create_language_model(google_ai::Model::default()))
255 }
256
257 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
258 Some(self.create_language_model(google_ai::Model::default_fast()))
259 }
260
261 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
262 let mut models = BTreeMap::default();
263
264 // Add base models from google_ai::Model::iter()
265 for model in google_ai::Model::iter() {
266 if !matches!(model, google_ai::Model::Custom { .. }) {
267 models.insert(model.id().to_string(), model);
268 }
269 }
270
271 // Override with available models from settings
272 for model in &AllLanguageModelSettings::get_global(cx)
273 .google
274 .available_models
275 {
276 models.insert(
277 model.name.clone(),
278 google_ai::Model::Custom {
279 name: model.name.clone(),
280 display_name: model.display_name.clone(),
281 max_tokens: model.max_tokens,
282 mode: model.mode.unwrap_or_default().into(),
283 },
284 );
285 }
286
287 models
288 .into_values()
289 .map(|model| {
290 Arc::new(GoogleLanguageModel {
291 id: LanguageModelId::from(model.id().to_string()),
292 model,
293 state: self.state.clone(),
294 http_client: self.http_client.clone(),
295 request_limiter: RateLimiter::new(4),
296 }) as Arc<dyn LanguageModel>
297 })
298 .collect()
299 }
300
301 fn is_authenticated(&self, cx: &App) -> bool {
302 self.state.read(cx).is_authenticated()
303 }
304
305 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
306 self.state.update(cx, |state, cx| state.authenticate(cx))
307 }
308
309 fn configuration_view(
310 &self,
311 target_agent: language_model::ConfigurationViewTargetAgent,
312 window: &mut Window,
313 cx: &mut App,
314 ) -> AnyView {
315 cx.new(|cx| ConfigurationView::new(self.state.clone(), target_agent, window, cx))
316 .into()
317 }
318
319 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
320 self.state.update(cx, |state, cx| state.reset_api_key(cx))
321 }
322}
323
324pub struct GoogleLanguageModel {
325 id: LanguageModelId,
326 model: google_ai::Model,
327 state: gpui::Entity<State>,
328 http_client: Arc<dyn HttpClient>,
329 request_limiter: RateLimiter,
330}
331
332impl GoogleLanguageModel {
333 fn stream_completion(
334 &self,
335 request: google_ai::GenerateContentRequest,
336 cx: &AsyncApp,
337 ) -> BoxFuture<
338 'static,
339 Result<futures::stream::BoxStream<'static, Result<GenerateContentResponse>>>,
340 > {
341 let http_client = self.http_client.clone();
342
343 let Ok((api_key, api_url)) = cx.read_entity(&self.state, |state, cx| {
344 let settings = &AllLanguageModelSettings::get_global(cx).google;
345 (state.api_key.clone(), settings.api_url.clone())
346 }) else {
347 return futures::future::ready(Err(anyhow!("App state dropped"))).boxed();
348 };
349
350 async move {
351 let api_key = api_key.context("Missing Google API key")?;
352 let request = google_ai::stream_generate_content(
353 http_client.as_ref(),
354 &api_url,
355 &api_key,
356 request,
357 );
358 request.await.context("failed to stream completion")
359 }
360 .boxed()
361 }
362}
363
364impl LanguageModel for GoogleLanguageModel {
365 fn id(&self) -> LanguageModelId {
366 self.id.clone()
367 }
368
369 fn name(&self) -> LanguageModelName {
370 LanguageModelName::from(self.model.display_name().to_string())
371 }
372
373 fn provider_id(&self) -> LanguageModelProviderId {
374 PROVIDER_ID
375 }
376
377 fn provider_name(&self) -> LanguageModelProviderName {
378 PROVIDER_NAME
379 }
380
381 fn supports_tools(&self) -> bool {
382 self.model.supports_tools()
383 }
384
385 fn supports_images(&self) -> bool {
386 self.model.supports_images()
387 }
388
389 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
390 match choice {
391 LanguageModelToolChoice::Auto
392 | LanguageModelToolChoice::Any
393 | LanguageModelToolChoice::None => true,
394 }
395 }
396
397 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
398 LanguageModelToolSchemaFormat::JsonSchemaSubset
399 }
400
401 fn telemetry_id(&self) -> String {
402 format!("google/{}", self.model.request_id())
403 }
404
405 fn max_token_count(&self) -> u64 {
406 self.model.max_token_count()
407 }
408
409 fn max_output_tokens(&self) -> Option<u64> {
410 self.model.max_output_tokens()
411 }
412
413 fn count_tokens(
414 &self,
415 request: LanguageModelRequest,
416 cx: &App,
417 ) -> BoxFuture<'static, Result<u64>> {
418 let model_id = self.model.request_id().to_string();
419 let request = into_google(request, model_id, self.model.mode());
420 let http_client = self.http_client.clone();
421 let api_key = self.state.read(cx).api_key.clone();
422
423 let settings = &AllLanguageModelSettings::get_global(cx).google;
424 let api_url = settings.api_url.clone();
425
426 async move {
427 let api_key = api_key.context("Missing Google API key")?;
428 let response = google_ai::count_tokens(
429 http_client.as_ref(),
430 &api_url,
431 &api_key,
432 google_ai::CountTokensRequest {
433 generate_content_request: request,
434 },
435 )
436 .await?;
437 Ok(response.total_tokens)
438 }
439 .boxed()
440 }
441
442 fn stream_completion(
443 &self,
444 request: LanguageModelRequest,
445 cx: &AsyncApp,
446 ) -> BoxFuture<
447 'static,
448 Result<
449 futures::stream::BoxStream<
450 'static,
451 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
452 >,
453 LanguageModelCompletionError,
454 >,
455 > {
456 let request = into_google(
457 request,
458 self.model.request_id().to_string(),
459 self.model.mode(),
460 );
461 let request = self.stream_completion(request, cx);
462 let future = self.request_limiter.stream(async move {
463 let response = request.await.map_err(LanguageModelCompletionError::from)?;
464 Ok(GoogleEventMapper::new().map_stream(response))
465 });
466 async move { Ok(future.await?.boxed()) }.boxed()
467 }
468}
469
470pub fn into_google(
471 mut request: LanguageModelRequest,
472 model_id: String,
473 mode: GoogleModelMode,
474) -> google_ai::GenerateContentRequest {
475 fn map_content(content: Vec<MessageContent>) -> Vec<Part> {
476 content
477 .into_iter()
478 .flat_map(|content| match content {
479 language_model::MessageContent::Text(text) => {
480 if !text.is_empty() {
481 vec![Part::TextPart(google_ai::TextPart { text })]
482 } else {
483 vec![]
484 }
485 }
486 language_model::MessageContent::Thinking {
487 text: _,
488 signature: Some(signature),
489 } => {
490 if !signature.is_empty() {
491 vec![Part::ThoughtPart(google_ai::ThoughtPart {
492 thought: true,
493 thought_signature: signature,
494 })]
495 } else {
496 vec![]
497 }
498 }
499 language_model::MessageContent::Thinking { .. } => {
500 vec![]
501 }
502 language_model::MessageContent::RedactedThinking(_) => vec![],
503 language_model::MessageContent::Image(image) => {
504 vec![Part::InlineDataPart(google_ai::InlineDataPart {
505 inline_data: google_ai::GenerativeContentBlob {
506 mime_type: "image/png".to_string(),
507 data: image.source.to_string(),
508 },
509 })]
510 }
511 language_model::MessageContent::ToolUse(tool_use) => {
512 vec![Part::FunctionCallPart(google_ai::FunctionCallPart {
513 function_call: google_ai::FunctionCall {
514 name: tool_use.name.to_string(),
515 args: tool_use.input,
516 },
517 })]
518 }
519 language_model::MessageContent::ToolResult(tool_result) => {
520 match tool_result.content {
521 language_model::LanguageModelToolResultContent::Text(text) => {
522 vec![Part::FunctionResponsePart(
523 google_ai::FunctionResponsePart {
524 function_response: google_ai::FunctionResponse {
525 name: tool_result.tool_name.to_string(),
526 // The API expects a valid JSON object
527 response: serde_json::json!({
528 "output": text
529 }),
530 },
531 },
532 )]
533 }
534 language_model::LanguageModelToolResultContent::Image(image) => {
535 vec![
536 Part::FunctionResponsePart(google_ai::FunctionResponsePart {
537 function_response: google_ai::FunctionResponse {
538 name: tool_result.tool_name.to_string(),
539 // The API expects a valid JSON object
540 response: serde_json::json!({
541 "output": "Tool responded with an image"
542 }),
543 },
544 }),
545 Part::InlineDataPart(google_ai::InlineDataPart {
546 inline_data: google_ai::GenerativeContentBlob {
547 mime_type: "image/png".to_string(),
548 data: image.source.to_string(),
549 },
550 }),
551 ]
552 }
553 }
554 }
555 })
556 .collect()
557 }
558
559 let system_instructions = if request
560 .messages
561 .first()
562 .is_some_and(|msg| matches!(msg.role, Role::System))
563 {
564 let message = request.messages.remove(0);
565 Some(SystemInstruction {
566 parts: map_content(message.content),
567 })
568 } else {
569 None
570 };
571
572 google_ai::GenerateContentRequest {
573 model: google_ai::ModelName { model_id },
574 system_instruction: system_instructions,
575 contents: request
576 .messages
577 .into_iter()
578 .filter_map(|message| {
579 let parts = map_content(message.content);
580 if parts.is_empty() {
581 None
582 } else {
583 Some(google_ai::Content {
584 parts,
585 role: match message.role {
586 Role::User => google_ai::Role::User,
587 Role::Assistant => google_ai::Role::Model,
588 Role::System => google_ai::Role::User, // Google AI doesn't have a system role
589 },
590 })
591 }
592 })
593 .collect(),
594 generation_config: Some(google_ai::GenerationConfig {
595 candidate_count: Some(1),
596 stop_sequences: Some(request.stop),
597 max_output_tokens: None,
598 temperature: request.temperature.map(|t| t as f64).or(Some(1.0)),
599 thinking_config: match (request.thinking_allowed, mode) {
600 (true, GoogleModelMode::Thinking { budget_tokens }) => {
601 budget_tokens.map(|thinking_budget| ThinkingConfig { thinking_budget })
602 }
603 _ => None,
604 },
605 top_p: None,
606 top_k: None,
607 }),
608 safety_settings: None,
609 tools: (!request.tools.is_empty()).then(|| {
610 vec![google_ai::Tool {
611 function_declarations: request
612 .tools
613 .into_iter()
614 .map(|tool| FunctionDeclaration {
615 name: tool.name,
616 description: tool.description,
617 parameters: tool.input_schema,
618 })
619 .collect(),
620 }]
621 }),
622 tool_config: request.tool_choice.map(|choice| google_ai::ToolConfig {
623 function_calling_config: google_ai::FunctionCallingConfig {
624 mode: match choice {
625 LanguageModelToolChoice::Auto => google_ai::FunctionCallingMode::Auto,
626 LanguageModelToolChoice::Any => google_ai::FunctionCallingMode::Any,
627 LanguageModelToolChoice::None => google_ai::FunctionCallingMode::None,
628 },
629 allowed_function_names: None,
630 },
631 }),
632 }
633}
634
635pub struct GoogleEventMapper {
636 usage: UsageMetadata,
637 stop_reason: StopReason,
638}
639
640impl GoogleEventMapper {
641 pub fn new() -> Self {
642 Self {
643 usage: UsageMetadata::default(),
644 stop_reason: StopReason::EndTurn,
645 }
646 }
647
648 pub fn map_stream(
649 mut self,
650 events: Pin<Box<dyn Send + Stream<Item = Result<GenerateContentResponse>>>>,
651 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
652 {
653 events
654 .map(Some)
655 .chain(futures::stream::once(async { None }))
656 .flat_map(move |event| {
657 futures::stream::iter(match event {
658 Some(Ok(event)) => self.map_event(event),
659 Some(Err(error)) => {
660 vec![Err(LanguageModelCompletionError::from(error))]
661 }
662 None => vec![Ok(LanguageModelCompletionEvent::Stop(self.stop_reason))],
663 })
664 })
665 }
666
667 pub fn map_event(
668 &mut self,
669 event: GenerateContentResponse,
670 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
671 static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
672
673 let mut events: Vec<_> = Vec::new();
674 let mut wants_to_use_tool = false;
675 if let Some(usage_metadata) = event.usage_metadata {
676 update_usage(&mut self.usage, &usage_metadata);
677 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
678 convert_usage(&self.usage),
679 )))
680 }
681 if let Some(candidates) = event.candidates {
682 for candidate in candidates {
683 if let Some(finish_reason) = candidate.finish_reason.as_deref() {
684 self.stop_reason = match finish_reason {
685 "STOP" => StopReason::EndTurn,
686 "MAX_TOKENS" => StopReason::MaxTokens,
687 _ => {
688 log::error!("Unexpected google finish_reason: {finish_reason}");
689 StopReason::EndTurn
690 }
691 };
692 }
693 candidate
694 .content
695 .parts
696 .into_iter()
697 .for_each(|part| match part {
698 Part::TextPart(text_part) => {
699 events.push(Ok(LanguageModelCompletionEvent::Text(text_part.text)))
700 }
701 Part::InlineDataPart(_) => {}
702 Part::FunctionCallPart(function_call_part) => {
703 wants_to_use_tool = true;
704 let name: Arc<str> = function_call_part.function_call.name.into();
705 let next_tool_id =
706 TOOL_CALL_COUNTER.fetch_add(1, atomic::Ordering::SeqCst);
707 let id: LanguageModelToolUseId =
708 format!("{}-{}", name, next_tool_id).into();
709
710 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
711 LanguageModelToolUse {
712 id,
713 name,
714 is_input_complete: true,
715 raw_input: function_call_part.function_call.args.to_string(),
716 input: function_call_part.function_call.args,
717 },
718 )));
719 }
720 Part::FunctionResponsePart(_) => {}
721 Part::ThoughtPart(part) => {
722 events.push(Ok(LanguageModelCompletionEvent::Thinking {
723 text: "(Encrypted thought)".to_string(), // TODO: Can we populate this from thought summaries?
724 signature: Some(part.thought_signature),
725 }));
726 }
727 });
728 }
729 }
730
731 // Even when Gemini wants to use a Tool, the API
732 // responds with `finish_reason: STOP`
733 if wants_to_use_tool {
734 self.stop_reason = StopReason::ToolUse;
735 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
736 }
737 events
738 }
739}
740
741pub fn count_google_tokens(
742 request: LanguageModelRequest,
743 cx: &App,
744) -> BoxFuture<'static, Result<u64>> {
745 // We couldn't use the GoogleLanguageModelProvider to count tokens because the github copilot doesn't have the access to google_ai directly.
746 // So we have to use tokenizer from tiktoken_rs to count tokens.
747 cx.background_spawn(async move {
748 let messages = request
749 .messages
750 .into_iter()
751 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
752 role: match message.role {
753 Role::User => "user".into(),
754 Role::Assistant => "assistant".into(),
755 Role::System => "system".into(),
756 },
757 content: Some(message.string_contents()),
758 name: None,
759 function_call: None,
760 })
761 .collect::<Vec<_>>();
762
763 // Tiktoken doesn't yet support these models, so we manually use the
764 // same tokenizer as GPT-4.
765 tiktoken_rs::num_tokens_from_messages("gpt-4", &messages).map(|tokens| tokens as u64)
766 })
767 .boxed()
768}
769
770fn update_usage(usage: &mut UsageMetadata, new: &UsageMetadata) {
771 if let Some(prompt_token_count) = new.prompt_token_count {
772 usage.prompt_token_count = Some(prompt_token_count);
773 }
774 if let Some(cached_content_token_count) = new.cached_content_token_count {
775 usage.cached_content_token_count = Some(cached_content_token_count);
776 }
777 if let Some(candidates_token_count) = new.candidates_token_count {
778 usage.candidates_token_count = Some(candidates_token_count);
779 }
780 if let Some(tool_use_prompt_token_count) = new.tool_use_prompt_token_count {
781 usage.tool_use_prompt_token_count = Some(tool_use_prompt_token_count);
782 }
783 if let Some(thoughts_token_count) = new.thoughts_token_count {
784 usage.thoughts_token_count = Some(thoughts_token_count);
785 }
786 if let Some(total_token_count) = new.total_token_count {
787 usage.total_token_count = Some(total_token_count);
788 }
789}
790
791fn convert_usage(usage: &UsageMetadata) -> language_model::TokenUsage {
792 let prompt_tokens = usage.prompt_token_count.unwrap_or(0);
793 let cached_tokens = usage.cached_content_token_count.unwrap_or(0);
794 let input_tokens = prompt_tokens - cached_tokens;
795 let output_tokens = usage.candidates_token_count.unwrap_or(0);
796
797 language_model::TokenUsage {
798 input_tokens,
799 output_tokens,
800 cache_read_input_tokens: cached_tokens,
801 cache_creation_input_tokens: 0,
802 }
803}
804
805struct ConfigurationView {
806 api_key_editor: Entity<Editor>,
807 state: gpui::Entity<State>,
808 target_agent: language_model::ConfigurationViewTargetAgent,
809 load_credentials_task: Option<Task<()>>,
810}
811
812impl ConfigurationView {
813 fn new(
814 state: gpui::Entity<State>,
815 target_agent: language_model::ConfigurationViewTargetAgent,
816 window: &mut Window,
817 cx: &mut Context<Self>,
818 ) -> Self {
819 cx.observe(&state, |_, _, cx| {
820 cx.notify();
821 })
822 .detach();
823
824 let load_credentials_task = Some(cx.spawn_in(window, {
825 let state = state.clone();
826 async move |this, cx| {
827 if let Some(task) = state
828 .update(cx, |state, cx| state.authenticate(cx))
829 .log_err()
830 {
831 // We don't log an error, because "not signed in" is also an error.
832 let _ = task.await;
833 }
834 this.update(cx, |this, cx| {
835 this.load_credentials_task = None;
836 cx.notify();
837 })
838 .log_err();
839 }
840 }));
841
842 Self {
843 api_key_editor: cx.new(|cx| {
844 let mut editor = Editor::single_line(window, cx);
845 editor.set_placeholder_text("AIzaSy...", window, cx);
846 editor
847 }),
848 target_agent,
849 state,
850 load_credentials_task,
851 }
852 }
853
854 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
855 let api_key = self.api_key_editor.read(cx).text(cx);
856 if api_key.is_empty() {
857 return;
858 }
859
860 let state = self.state.clone();
861 cx.spawn_in(window, async move |_, cx| {
862 state
863 .update(cx, |state, cx| state.set_api_key(api_key, cx))?
864 .await
865 })
866 .detach_and_log_err(cx);
867
868 cx.notify();
869 }
870
871 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
872 self.api_key_editor
873 .update(cx, |editor, cx| editor.set_text("", window, cx));
874
875 let state = self.state.clone();
876 cx.spawn_in(window, async move |_, cx| {
877 state.update(cx, |state, cx| state.reset_api_key(cx))?.await
878 })
879 .detach_and_log_err(cx);
880
881 cx.notify();
882 }
883
884 fn render_api_key_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
885 let settings = ThemeSettings::get_global(cx);
886 let text_style = TextStyle {
887 color: cx.theme().colors().text,
888 font_family: settings.ui_font.family.clone(),
889 font_features: settings.ui_font.features.clone(),
890 font_fallbacks: settings.ui_font.fallbacks.clone(),
891 font_size: rems(0.875).into(),
892 font_weight: settings.ui_font.weight,
893 font_style: FontStyle::Normal,
894 line_height: relative(1.3),
895 white_space: WhiteSpace::Normal,
896 ..Default::default()
897 };
898 EditorElement::new(
899 &self.api_key_editor,
900 EditorStyle {
901 background: cx.theme().colors().editor_background,
902 local_player: cx.theme().players().local(),
903 text: text_style,
904 ..Default::default()
905 },
906 )
907 }
908
909 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
910 !self.state.read(cx).is_authenticated()
911 }
912}
913
914impl Render for ConfigurationView {
915 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
916 let env_var_set = self.state.read(cx).api_key_from_env;
917
918 if self.load_credentials_task.is_some() {
919 div().child(Label::new("Loading credentials...")).into_any()
920 } else if self.should_render_editor(cx) {
921 v_flex()
922 .size_full()
923 .on_action(cx.listener(Self::save_api_key))
924 .child(Label::new(format!("To use {}, you need to add an API key. Follow these steps:", match &self.target_agent {
925 ConfigurationViewTargetAgent::ZedAgent => "Zed's agent with Google AI".into(),
926 ConfigurationViewTargetAgent::Other(agent) => agent.clone(),
927 })))
928 .child(
929 List::new()
930 .child(InstructionListItem::new(
931 "Create one by visiting",
932 Some("Google AI's console"),
933 Some("https://aistudio.google.com/app/apikey"),
934 ))
935 .child(InstructionListItem::text_only(
936 "Paste your API key below and hit enter to start using the assistant",
937 )),
938 )
939 .child(
940 h_flex()
941 .w_full()
942 .my_2()
943 .px_2()
944 .py_1()
945 .bg(cx.theme().colors().editor_background)
946 .border_1()
947 .border_color(cx.theme().colors().border)
948 .rounded_sm()
949 .child(self.render_api_key_editor(cx)),
950 )
951 .child(
952 Label::new(
953 format!("You can also assign the {GEMINI_API_KEY_VAR} environment variable and restart Zed."),
954 )
955 .size(LabelSize::Small).color(Color::Muted),
956 )
957 .into_any()
958 } else {
959 h_flex()
960 .mt_1()
961 .p_1()
962 .justify_between()
963 .rounded_md()
964 .border_1()
965 .border_color(cx.theme().colors().border)
966 .bg(cx.theme().colors().background)
967 .child(
968 h_flex()
969 .gap_1()
970 .child(Icon::new(IconName::Check).color(Color::Success))
971 .child(Label::new(if env_var_set {
972 format!("API key set in {GEMINI_API_KEY_VAR} environment variable.")
973 } else {
974 "API key configured.".to_string()
975 })),
976 )
977 .child(
978 Button::new("reset-key", "Reset Key")
979 .label_size(LabelSize::Small)
980 .icon(Some(IconName::Trash))
981 .icon_size(IconSize::Small)
982 .icon_position(IconPosition::Start)
983 .disabled(env_var_set)
984 .when(env_var_set, |this| {
985 this.tooltip(Tooltip::text(format!("To reset your API key, make sure {GEMINI_API_KEY_VAR} and {GOOGLE_AI_API_KEY_VAR} environment variables are unset.")))
986 })
987 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx))),
988 )
989 .into_any()
990 }
991 }
992}