1use anyhow::{Context as _, Result};
2use collections::BTreeMap;
3use credentials_provider::CredentialsProvider;
4use futures::{FutureExt, Stream, StreamExt, future::BoxFuture};
5use google_ai::{
6 FunctionDeclaration, GenerateContentResponse, GoogleModelMode, Part, SystemInstruction,
7 ThinkingConfig, UsageMetadata,
8};
9use gpui::{AnyView, App, AsyncApp, Context, Entity, SharedString, Task, Window};
10use http_client::HttpClient;
11use language_model::{
12 AuthenticateError, ConfigurationViewTargetAgent, EnvVar, LanguageModelCompletionError,
13 LanguageModelCompletionEvent, LanguageModelToolChoice, LanguageModelToolSchemaFormat,
14 LanguageModelToolUse, LanguageModelToolUseId, MessageContent, StopReason,
15};
16use language_model::{
17 GOOGLE_PROVIDER_ID, GOOGLE_PROVIDER_NAME, IconOrSvg, LanguageModel, LanguageModelId,
18 LanguageModelName, LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
19 LanguageModelProviderState, LanguageModelRequest, RateLimiter, Role,
20};
21use schemars::JsonSchema;
22use serde::{Deserialize, Serialize};
23pub use settings::GoogleAvailableModel as AvailableModel;
24use settings::{Settings, SettingsStore};
25use std::pin::Pin;
26use std::sync::{
27 Arc, LazyLock,
28 atomic::{self, AtomicU64},
29};
30use strum::IntoEnumIterator;
31use ui::{ButtonLink, ConfiguredApiCard, List, ListBulletItem, prelude::*};
32use ui_input::InputField;
33use util::ResultExt;
34
35use language_model::ApiKeyState;
36
37const PROVIDER_ID: LanguageModelProviderId = GOOGLE_PROVIDER_ID;
38const PROVIDER_NAME: LanguageModelProviderName = GOOGLE_PROVIDER_NAME;
39
40#[derive(Default, Clone, Debug, PartialEq)]
41pub struct GoogleSettings {
42 pub api_url: String,
43 pub available_models: Vec<AvailableModel>,
44}
45
46#[derive(Clone, Copy, Debug, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
47#[serde(tag = "type", rename_all = "lowercase")]
48pub enum ModelMode {
49 #[default]
50 Default,
51 Thinking {
52 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
53 budget_tokens: Option<u32>,
54 },
55}
56
57pub struct GoogleLanguageModelProvider {
58 http_client: Arc<dyn HttpClient>,
59 state: Entity<State>,
60}
61
62pub struct State {
63 api_key_state: ApiKeyState,
64 credentials_provider: Arc<dyn CredentialsProvider>,
65}
66
67const GEMINI_API_KEY_VAR_NAME: &str = "GEMINI_API_KEY";
68const GOOGLE_AI_API_KEY_VAR_NAME: &str = "GOOGLE_AI_API_KEY";
69
70static API_KEY_ENV_VAR: LazyLock<EnvVar> = LazyLock::new(|| {
71 // Try GEMINI_API_KEY first as primary, fallback to GOOGLE_AI_API_KEY
72 EnvVar::new(GEMINI_API_KEY_VAR_NAME.into()).or(EnvVar::new(GOOGLE_AI_API_KEY_VAR_NAME.into()))
73});
74
75impl State {
76 fn is_authenticated(&self) -> bool {
77 self.api_key_state.has_key()
78 }
79
80 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
81 let credentials_provider = self.credentials_provider.clone();
82 let api_url = GoogleLanguageModelProvider::api_url(cx);
83 self.api_key_state.store(
84 api_url,
85 api_key,
86 |this| &mut this.api_key_state,
87 credentials_provider,
88 cx,
89 )
90 }
91
92 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
93 let credentials_provider = self.credentials_provider.clone();
94 let api_url = GoogleLanguageModelProvider::api_url(cx);
95 self.api_key_state.load_if_needed(
96 api_url,
97 |this| &mut this.api_key_state,
98 credentials_provider,
99 cx,
100 )
101 }
102}
103
104impl GoogleLanguageModelProvider {
105 pub fn new(
106 http_client: Arc<dyn HttpClient>,
107 credentials_provider: Arc<dyn CredentialsProvider>,
108 cx: &mut App,
109 ) -> Self {
110 let state = cx.new(|cx| {
111 cx.observe_global::<SettingsStore>(|this: &mut State, cx| {
112 let credentials_provider = this.credentials_provider.clone();
113 let api_url = Self::api_url(cx);
114 this.api_key_state.handle_url_change(
115 api_url,
116 |this| &mut this.api_key_state,
117 credentials_provider,
118 cx,
119 );
120 cx.notify();
121 })
122 .detach();
123 State {
124 api_key_state: ApiKeyState::new(Self::api_url(cx), (*API_KEY_ENV_VAR).clone()),
125 credentials_provider,
126 }
127 });
128
129 Self { http_client, state }
130 }
131
132 fn create_language_model(&self, model: google_ai::Model) -> Arc<dyn LanguageModel> {
133 Arc::new(GoogleLanguageModel {
134 id: LanguageModelId::from(model.id().to_string()),
135 model,
136 state: self.state.clone(),
137 http_client: self.http_client.clone(),
138 request_limiter: RateLimiter::new(4),
139 })
140 }
141
142 fn settings(cx: &App) -> &GoogleSettings {
143 &crate::AllLanguageModelSettings::get_global(cx).google
144 }
145
146 fn api_url(cx: &App) -> SharedString {
147 let api_url = &Self::settings(cx).api_url;
148 if api_url.is_empty() {
149 google_ai::API_URL.into()
150 } else {
151 SharedString::new(api_url.as_str())
152 }
153 }
154}
155
156impl LanguageModelProviderState for GoogleLanguageModelProvider {
157 type ObservableEntity = State;
158
159 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
160 Some(self.state.clone())
161 }
162}
163
164impl LanguageModelProvider for GoogleLanguageModelProvider {
165 fn id(&self) -> LanguageModelProviderId {
166 PROVIDER_ID
167 }
168
169 fn name(&self) -> LanguageModelProviderName {
170 PROVIDER_NAME
171 }
172
173 fn icon(&self) -> IconOrSvg {
174 IconOrSvg::Icon(IconName::AiGoogle)
175 }
176
177 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
178 Some(self.create_language_model(google_ai::Model::default()))
179 }
180
181 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
182 Some(self.create_language_model(google_ai::Model::default_fast()))
183 }
184
185 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
186 let mut models = BTreeMap::default();
187
188 // Add base models from google_ai::Model::iter()
189 for model in google_ai::Model::iter() {
190 if !matches!(model, google_ai::Model::Custom { .. }) {
191 models.insert(model.id().to_string(), model);
192 }
193 }
194
195 // Override with available models from settings
196 for model in &GoogleLanguageModelProvider::settings(cx).available_models {
197 models.insert(
198 model.name.clone(),
199 google_ai::Model::Custom {
200 name: model.name.clone(),
201 display_name: model.display_name.clone(),
202 max_tokens: model.max_tokens,
203 mode: model.mode.unwrap_or_default(),
204 },
205 );
206 }
207
208 models
209 .into_values()
210 .map(|model| {
211 Arc::new(GoogleLanguageModel {
212 id: LanguageModelId::from(model.id().to_string()),
213 model,
214 state: self.state.clone(),
215 http_client: self.http_client.clone(),
216 request_limiter: RateLimiter::new(4),
217 }) as Arc<dyn LanguageModel>
218 })
219 .collect()
220 }
221
222 fn is_authenticated(&self, cx: &App) -> bool {
223 self.state.read(cx).is_authenticated()
224 }
225
226 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
227 self.state.update(cx, |state, cx| state.authenticate(cx))
228 }
229
230 fn configuration_view(
231 &self,
232 target_agent: language_model::ConfigurationViewTargetAgent,
233 window: &mut Window,
234 cx: &mut App,
235 ) -> AnyView {
236 cx.new(|cx| ConfigurationView::new(self.state.clone(), target_agent, window, cx))
237 .into()
238 }
239
240 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
241 self.state
242 .update(cx, |state, cx| state.set_api_key(None, cx))
243 }
244}
245
246pub struct GoogleLanguageModel {
247 id: LanguageModelId,
248 model: google_ai::Model,
249 state: Entity<State>,
250 http_client: Arc<dyn HttpClient>,
251 request_limiter: RateLimiter,
252}
253
254impl GoogleLanguageModel {
255 fn stream_completion(
256 &self,
257 request: google_ai::GenerateContentRequest,
258 cx: &AsyncApp,
259 ) -> BoxFuture<
260 'static,
261 Result<futures::stream::BoxStream<'static, Result<GenerateContentResponse>>>,
262 > {
263 let http_client = self.http_client.clone();
264
265 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
266 let api_url = GoogleLanguageModelProvider::api_url(cx);
267 (state.api_key_state.key(&api_url), api_url)
268 });
269
270 async move {
271 let api_key = api_key.context("Missing Google API key")?;
272 let request = google_ai::stream_generate_content(
273 http_client.as_ref(),
274 &api_url,
275 &api_key,
276 request,
277 );
278 request.await.context("failed to stream completion")
279 }
280 .boxed()
281 }
282}
283
284impl LanguageModel for GoogleLanguageModel {
285 fn id(&self) -> LanguageModelId {
286 self.id.clone()
287 }
288
289 fn name(&self) -> LanguageModelName {
290 LanguageModelName::from(self.model.display_name().to_string())
291 }
292
293 fn provider_id(&self) -> LanguageModelProviderId {
294 PROVIDER_ID
295 }
296
297 fn provider_name(&self) -> LanguageModelProviderName {
298 PROVIDER_NAME
299 }
300
301 fn supports_tools(&self) -> bool {
302 self.model.supports_tools()
303 }
304
305 fn supports_images(&self) -> bool {
306 self.model.supports_images()
307 }
308
309 fn supports_thinking(&self) -> bool {
310 matches!(self.model.mode(), GoogleModelMode::Thinking { .. })
311 }
312
313 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
314 match choice {
315 LanguageModelToolChoice::Auto
316 | LanguageModelToolChoice::Any
317 | LanguageModelToolChoice::None => true,
318 }
319 }
320
321 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
322 LanguageModelToolSchemaFormat::JsonSchemaSubset
323 }
324
325 fn telemetry_id(&self) -> String {
326 format!("google/{}", self.model.request_id())
327 }
328
329 fn max_token_count(&self) -> u64 {
330 self.model.max_token_count()
331 }
332
333 fn max_output_tokens(&self) -> Option<u64> {
334 self.model.max_output_tokens()
335 }
336
337 fn count_tokens(
338 &self,
339 request: LanguageModelRequest,
340 cx: &App,
341 ) -> BoxFuture<'static, Result<u64>> {
342 let model_id = self.model.request_id().to_string();
343 let request = into_google(request, model_id, self.model.mode());
344 let http_client = self.http_client.clone();
345 let api_url = GoogleLanguageModelProvider::api_url(cx);
346 let api_key = self.state.read(cx).api_key_state.key(&api_url);
347
348 async move {
349 let Some(api_key) = api_key else {
350 return Err(LanguageModelCompletionError::NoApiKey {
351 provider: PROVIDER_NAME,
352 }
353 .into());
354 };
355 let response = google_ai::count_tokens(
356 http_client.as_ref(),
357 &api_url,
358 &api_key,
359 google_ai::CountTokensRequest {
360 generate_content_request: request,
361 },
362 )
363 .await?;
364 Ok(response.total_tokens)
365 }
366 .boxed()
367 }
368
369 fn stream_completion(
370 &self,
371 request: LanguageModelRequest,
372 cx: &AsyncApp,
373 ) -> BoxFuture<
374 'static,
375 Result<
376 futures::stream::BoxStream<
377 'static,
378 Result<LanguageModelCompletionEvent, LanguageModelCompletionError>,
379 >,
380 LanguageModelCompletionError,
381 >,
382 > {
383 let request = into_google(
384 request,
385 self.model.request_id().to_string(),
386 self.model.mode(),
387 );
388 let request = self.stream_completion(request, cx);
389 let future = self.request_limiter.stream(async move {
390 let response = request.await.map_err(LanguageModelCompletionError::from)?;
391 Ok(GoogleEventMapper::new().map_stream(response))
392 });
393 async move { Ok(future.await?.boxed()) }.boxed()
394 }
395}
396
397pub fn into_google(
398 mut request: LanguageModelRequest,
399 model_id: String,
400 mode: GoogleModelMode,
401) -> google_ai::GenerateContentRequest {
402 fn map_content(content: Vec<MessageContent>) -> Vec<Part> {
403 content
404 .into_iter()
405 .flat_map(|content| match content {
406 language_model::MessageContent::Text(text) => {
407 if !text.is_empty() {
408 vec![Part::TextPart(google_ai::TextPart { text })]
409 } else {
410 vec![]
411 }
412 }
413 language_model::MessageContent::Thinking {
414 text: _,
415 signature: Some(signature),
416 } => {
417 if !signature.is_empty() {
418 vec![Part::ThoughtPart(google_ai::ThoughtPart {
419 thought: true,
420 thought_signature: signature,
421 })]
422 } else {
423 vec![]
424 }
425 }
426 language_model::MessageContent::Thinking { .. } => {
427 vec![]
428 }
429 language_model::MessageContent::RedactedThinking(_) => vec![],
430 language_model::MessageContent::Image(image) => {
431 vec![Part::InlineDataPart(google_ai::InlineDataPart {
432 inline_data: google_ai::GenerativeContentBlob {
433 mime_type: "image/png".to_string(),
434 data: image.source.to_string(),
435 },
436 })]
437 }
438 language_model::MessageContent::ToolUse(tool_use) => {
439 // Normalize empty string signatures to None
440 let thought_signature = tool_use.thought_signature.filter(|s| !s.is_empty());
441
442 vec![Part::FunctionCallPart(google_ai::FunctionCallPart {
443 function_call: google_ai::FunctionCall {
444 name: tool_use.name.to_string(),
445 args: tool_use.input,
446 },
447 thought_signature,
448 })]
449 }
450 language_model::MessageContent::ToolResult(tool_result) => {
451 match tool_result.content {
452 language_model::LanguageModelToolResultContent::Text(text) => {
453 vec![Part::FunctionResponsePart(
454 google_ai::FunctionResponsePart {
455 function_response: google_ai::FunctionResponse {
456 name: tool_result.tool_name.to_string(),
457 // The API expects a valid JSON object
458 response: serde_json::json!({
459 "output": text
460 }),
461 },
462 },
463 )]
464 }
465 language_model::LanguageModelToolResultContent::Image(image) => {
466 vec![
467 Part::FunctionResponsePart(google_ai::FunctionResponsePart {
468 function_response: google_ai::FunctionResponse {
469 name: tool_result.tool_name.to_string(),
470 // The API expects a valid JSON object
471 response: serde_json::json!({
472 "output": "Tool responded with an image"
473 }),
474 },
475 }),
476 Part::InlineDataPart(google_ai::InlineDataPart {
477 inline_data: google_ai::GenerativeContentBlob {
478 mime_type: "image/png".to_string(),
479 data: image.source.to_string(),
480 },
481 }),
482 ]
483 }
484 }
485 }
486 })
487 .collect()
488 }
489
490 let system_instructions = if request
491 .messages
492 .first()
493 .is_some_and(|msg| matches!(msg.role, Role::System))
494 {
495 let message = request.messages.remove(0);
496 Some(SystemInstruction {
497 parts: map_content(message.content),
498 })
499 } else {
500 None
501 };
502
503 google_ai::GenerateContentRequest {
504 model: google_ai::ModelName { model_id },
505 system_instruction: system_instructions,
506 contents: request
507 .messages
508 .into_iter()
509 .filter_map(|message| {
510 let parts = map_content(message.content);
511 if parts.is_empty() {
512 None
513 } else {
514 Some(google_ai::Content {
515 parts,
516 role: match message.role {
517 Role::User => google_ai::Role::User,
518 Role::Assistant => google_ai::Role::Model,
519 Role::System => google_ai::Role::User, // Google AI doesn't have a system role
520 },
521 })
522 }
523 })
524 .collect(),
525 generation_config: Some(google_ai::GenerationConfig {
526 candidate_count: Some(1),
527 stop_sequences: Some(request.stop),
528 max_output_tokens: None,
529 temperature: request.temperature.map(|t| t as f64).or(Some(1.0)),
530 thinking_config: match (request.thinking_allowed, mode) {
531 (true, GoogleModelMode::Thinking { budget_tokens }) => {
532 budget_tokens.map(|thinking_budget| ThinkingConfig { thinking_budget })
533 }
534 _ => None,
535 },
536 top_p: None,
537 top_k: None,
538 }),
539 safety_settings: None,
540 tools: (!request.tools.is_empty()).then(|| {
541 vec![google_ai::Tool {
542 function_declarations: request
543 .tools
544 .into_iter()
545 .map(|tool| FunctionDeclaration {
546 name: tool.name,
547 description: tool.description,
548 parameters: tool.input_schema,
549 })
550 .collect(),
551 }]
552 }),
553 tool_config: request.tool_choice.map(|choice| google_ai::ToolConfig {
554 function_calling_config: google_ai::FunctionCallingConfig {
555 mode: match choice {
556 LanguageModelToolChoice::Auto => google_ai::FunctionCallingMode::Auto,
557 LanguageModelToolChoice::Any => google_ai::FunctionCallingMode::Any,
558 LanguageModelToolChoice::None => google_ai::FunctionCallingMode::None,
559 },
560 allowed_function_names: None,
561 },
562 }),
563 }
564}
565
566pub struct GoogleEventMapper {
567 usage: UsageMetadata,
568 stop_reason: StopReason,
569}
570
571impl GoogleEventMapper {
572 pub fn new() -> Self {
573 Self {
574 usage: UsageMetadata::default(),
575 stop_reason: StopReason::EndTurn,
576 }
577 }
578
579 pub fn map_stream(
580 mut self,
581 events: Pin<Box<dyn Send + Stream<Item = Result<GenerateContentResponse>>>>,
582 ) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>
583 {
584 events
585 .map(Some)
586 .chain(futures::stream::once(async { None }))
587 .flat_map(move |event| {
588 futures::stream::iter(match event {
589 Some(Ok(event)) => self.map_event(event),
590 Some(Err(error)) => {
591 vec![Err(LanguageModelCompletionError::from(error))]
592 }
593 None => vec![Ok(LanguageModelCompletionEvent::Stop(self.stop_reason))],
594 })
595 })
596 }
597
598 pub fn map_event(
599 &mut self,
600 event: GenerateContentResponse,
601 ) -> Vec<Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
602 static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
603
604 let mut events: Vec<_> = Vec::new();
605 let mut wants_to_use_tool = false;
606 if let Some(usage_metadata) = event.usage_metadata {
607 update_usage(&mut self.usage, &usage_metadata);
608 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(
609 convert_usage(&self.usage),
610 )))
611 }
612
613 if let Some(prompt_feedback) = event.prompt_feedback
614 && let Some(block_reason) = prompt_feedback.block_reason.as_deref()
615 {
616 self.stop_reason = match block_reason {
617 "SAFETY" | "OTHER" | "BLOCKLIST" | "PROHIBITED_CONTENT" | "IMAGE_SAFETY" => {
618 StopReason::Refusal
619 }
620 _ => {
621 log::error!("Unexpected Google block_reason: {block_reason}");
622 StopReason::Refusal
623 }
624 };
625 events.push(Ok(LanguageModelCompletionEvent::Stop(self.stop_reason)));
626
627 return events;
628 }
629
630 if let Some(candidates) = event.candidates {
631 for candidate in candidates {
632 if let Some(finish_reason) = candidate.finish_reason.as_deref() {
633 self.stop_reason = match finish_reason {
634 "STOP" => StopReason::EndTurn,
635 "MAX_TOKENS" => StopReason::MaxTokens,
636 _ => {
637 log::error!("Unexpected google finish_reason: {finish_reason}");
638 StopReason::EndTurn
639 }
640 };
641 }
642 candidate
643 .content
644 .parts
645 .into_iter()
646 .for_each(|part| match part {
647 Part::TextPart(text_part) => {
648 events.push(Ok(LanguageModelCompletionEvent::Text(text_part.text)))
649 }
650 Part::InlineDataPart(_) => {}
651 Part::FunctionCallPart(function_call_part) => {
652 wants_to_use_tool = true;
653 let name: Arc<str> = function_call_part.function_call.name.into();
654 let next_tool_id =
655 TOOL_CALL_COUNTER.fetch_add(1, atomic::Ordering::SeqCst);
656 let id: LanguageModelToolUseId =
657 format!("{}-{}", name, next_tool_id).into();
658
659 // Normalize empty string signatures to None
660 let thought_signature = function_call_part
661 .thought_signature
662 .filter(|s| !s.is_empty());
663
664 events.push(Ok(LanguageModelCompletionEvent::ToolUse(
665 LanguageModelToolUse {
666 id,
667 name,
668 is_input_complete: true,
669 raw_input: function_call_part.function_call.args.to_string(),
670 input: function_call_part.function_call.args,
671 thought_signature,
672 },
673 )));
674 }
675 Part::FunctionResponsePart(_) => {}
676 Part::ThoughtPart(part) => {
677 events.push(Ok(LanguageModelCompletionEvent::Thinking {
678 text: "(Encrypted thought)".to_string(), // TODO: Can we populate this from thought summaries?
679 signature: Some(part.thought_signature),
680 }));
681 }
682 });
683 }
684 }
685
686 // Even when Gemini wants to use a Tool, the API
687 // responds with `finish_reason: STOP`
688 if wants_to_use_tool {
689 self.stop_reason = StopReason::ToolUse;
690 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
691 }
692 events
693 }
694}
695
696pub fn count_google_tokens(
697 request: LanguageModelRequest,
698 cx: &App,
699) -> BoxFuture<'static, Result<u64>> {
700 // We couldn't use the GoogleLanguageModelProvider to count tokens because the github copilot doesn't have the access to google_ai directly.
701 // So we have to use tokenizer from tiktoken_rs to count tokens.
702 cx.background_spawn(async move {
703 let messages = request
704 .messages
705 .into_iter()
706 .map(|message| tiktoken_rs::ChatCompletionRequestMessage {
707 role: match message.role {
708 Role::User => "user".into(),
709 Role::Assistant => "assistant".into(),
710 Role::System => "system".into(),
711 },
712 content: Some(message.string_contents()),
713 name: None,
714 function_call: None,
715 })
716 .collect::<Vec<_>>();
717
718 // Tiktoken doesn't yet support these models, so we manually use the
719 // same tokenizer as GPT-4.
720 tiktoken_rs::num_tokens_from_messages("gpt-4", &messages).map(|tokens| tokens as u64)
721 })
722 .boxed()
723}
724
725fn update_usage(usage: &mut UsageMetadata, new: &UsageMetadata) {
726 if let Some(prompt_token_count) = new.prompt_token_count {
727 usage.prompt_token_count = Some(prompt_token_count);
728 }
729 if let Some(cached_content_token_count) = new.cached_content_token_count {
730 usage.cached_content_token_count = Some(cached_content_token_count);
731 }
732 if let Some(candidates_token_count) = new.candidates_token_count {
733 usage.candidates_token_count = Some(candidates_token_count);
734 }
735 if let Some(tool_use_prompt_token_count) = new.tool_use_prompt_token_count {
736 usage.tool_use_prompt_token_count = Some(tool_use_prompt_token_count);
737 }
738 if let Some(thoughts_token_count) = new.thoughts_token_count {
739 usage.thoughts_token_count = Some(thoughts_token_count);
740 }
741 if let Some(total_token_count) = new.total_token_count {
742 usage.total_token_count = Some(total_token_count);
743 }
744}
745
746fn convert_usage(usage: &UsageMetadata) -> language_model::TokenUsage {
747 let prompt_tokens = usage.prompt_token_count.unwrap_or(0);
748 let cached_tokens = usage.cached_content_token_count.unwrap_or(0);
749 let input_tokens = prompt_tokens - cached_tokens;
750 let output_tokens = usage.candidates_token_count.unwrap_or(0);
751
752 language_model::TokenUsage {
753 input_tokens,
754 output_tokens,
755 cache_read_input_tokens: cached_tokens,
756 cache_creation_input_tokens: 0,
757 }
758}
759
760struct ConfigurationView {
761 api_key_editor: Entity<InputField>,
762 state: Entity<State>,
763 target_agent: language_model::ConfigurationViewTargetAgent,
764 load_credentials_task: Option<Task<()>>,
765}
766
767impl ConfigurationView {
768 fn new(
769 state: Entity<State>,
770 target_agent: language_model::ConfigurationViewTargetAgent,
771 window: &mut Window,
772 cx: &mut Context<Self>,
773 ) -> Self {
774 cx.observe(&state, |_, _, cx| {
775 cx.notify();
776 })
777 .detach();
778
779 let load_credentials_task = Some(cx.spawn_in(window, {
780 let state = state.clone();
781 async move |this, cx| {
782 if let Some(task) = Some(state.update(cx, |state, cx| state.authenticate(cx))) {
783 // We don't log an error, because "not signed in" is also an error.
784 let _ = task.await;
785 }
786 this.update(cx, |this, cx| {
787 this.load_credentials_task = None;
788 cx.notify();
789 })
790 .log_err();
791 }
792 }));
793
794 Self {
795 api_key_editor: cx.new(|cx| InputField::new(window, cx, "AIzaSy...")),
796 target_agent,
797 state,
798 load_credentials_task,
799 }
800 }
801
802 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
803 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
804 if api_key.is_empty() {
805 return;
806 }
807
808 // url changes can cause the editor to be displayed again
809 self.api_key_editor
810 .update(cx, |editor, cx| editor.set_text("", window, cx));
811
812 let state = self.state.clone();
813 cx.spawn_in(window, async move |_, cx| {
814 state
815 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
816 .await
817 })
818 .detach_and_log_err(cx);
819 }
820
821 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
822 self.api_key_editor
823 .update(cx, |editor, cx| editor.set_text("", window, cx));
824
825 let state = self.state.clone();
826 cx.spawn_in(window, async move |_, cx| {
827 state
828 .update(cx, |state, cx| state.set_api_key(None, cx))
829 .await
830 })
831 .detach_and_log_err(cx);
832 }
833
834 fn should_render_editor(&self, cx: &mut Context<Self>) -> bool {
835 !self.state.read(cx).is_authenticated()
836 }
837}
838
839impl Render for ConfigurationView {
840 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
841 let env_var_set = self.state.read(cx).api_key_state.is_from_env_var();
842 let configured_card_label = if env_var_set {
843 format!(
844 "API key set in {} environment variable",
845 API_KEY_ENV_VAR.name
846 )
847 } else {
848 let api_url = GoogleLanguageModelProvider::api_url(cx);
849 if api_url == google_ai::API_URL {
850 "API key configured".to_string()
851 } else {
852 format!("API key configured for {}", api_url)
853 }
854 };
855
856 if self.load_credentials_task.is_some() {
857 div()
858 .child(Label::new("Loading credentials..."))
859 .into_any_element()
860 } else if self.should_render_editor(cx) {
861 v_flex()
862 .size_full()
863 .on_action(cx.listener(Self::save_api_key))
864 .child(Label::new(format!("To use {}, you need to add an API key. Follow these steps:", match &self.target_agent {
865 ConfigurationViewTargetAgent::ZedAgent => "Zed's agent with Google AI".into(),
866 ConfigurationViewTargetAgent::Other(agent) => agent.clone(),
867 })))
868 .child(
869 List::new()
870 .child(
871 ListBulletItem::new("")
872 .child(Label::new("Create one by visiting"))
873 .child(ButtonLink::new("Google AI's console", "https://aistudio.google.com/app/apikey"))
874 )
875 .child(
876 ListBulletItem::new("Paste your API key below and hit enter to start using the agent")
877 )
878 )
879 .child(self.api_key_editor.clone())
880 .child(
881 Label::new(
882 format!("You can also set the {GEMINI_API_KEY_VAR_NAME} environment variable and restart Zed."),
883 )
884 .size(LabelSize::Small).color(Color::Muted),
885 )
886 .into_any_element()
887 } else {
888 ConfiguredApiCard::new(configured_card_label)
889 .disabled(env_var_set)
890 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
891 .when(env_var_set, |this| {
892 this.tooltip_label(format!("To reset your API key, make sure {GEMINI_API_KEY_VAR_NAME} and {GOOGLE_AI_API_KEY_VAR_NAME} environment variables are unset."))
893 })
894 .into_any_element()
895 }
896 }
897}
898
899#[cfg(test)]
900mod tests {
901 use super::*;
902 use google_ai::{
903 Content, FunctionCall, FunctionCallPart, GenerateContentCandidate, GenerateContentResponse,
904 Part, Role as GoogleRole, TextPart,
905 };
906 use language_model::{LanguageModelToolUseId, MessageContent, Role};
907 use serde_json::json;
908
909 #[test]
910 fn test_function_call_with_signature_creates_tool_use_with_signature() {
911 let mut mapper = GoogleEventMapper::new();
912
913 let response = GenerateContentResponse {
914 candidates: Some(vec![GenerateContentCandidate {
915 index: Some(0),
916 content: Content {
917 parts: vec![Part::FunctionCallPart(FunctionCallPart {
918 function_call: FunctionCall {
919 name: "test_function".to_string(),
920 args: json!({"arg": "value"}),
921 },
922 thought_signature: Some("test_signature_123".to_string()),
923 })],
924 role: GoogleRole::Model,
925 },
926 finish_reason: None,
927 finish_message: None,
928 safety_ratings: None,
929 citation_metadata: None,
930 }]),
931 prompt_feedback: None,
932 usage_metadata: None,
933 };
934
935 let events = mapper.map_event(response);
936
937 assert_eq!(events.len(), 2); // ToolUse event + Stop event
938
939 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
940 assert_eq!(tool_use.name.as_ref(), "test_function");
941 assert_eq!(
942 tool_use.thought_signature.as_deref(),
943 Some("test_signature_123")
944 );
945 } else {
946 panic!("Expected ToolUse event");
947 }
948 }
949
950 #[test]
951 fn test_function_call_without_signature_has_none() {
952 let mut mapper = GoogleEventMapper::new();
953
954 let response = GenerateContentResponse {
955 candidates: Some(vec![GenerateContentCandidate {
956 index: Some(0),
957 content: Content {
958 parts: vec![Part::FunctionCallPart(FunctionCallPart {
959 function_call: FunctionCall {
960 name: "test_function".to_string(),
961 args: json!({"arg": "value"}),
962 },
963 thought_signature: None,
964 })],
965 role: GoogleRole::Model,
966 },
967 finish_reason: None,
968 finish_message: None,
969 safety_ratings: None,
970 citation_metadata: None,
971 }]),
972 prompt_feedback: None,
973 usage_metadata: None,
974 };
975
976 let events = mapper.map_event(response);
977
978 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
979 assert_eq!(tool_use.thought_signature, None);
980 } else {
981 panic!("Expected ToolUse event");
982 }
983 }
984
985 #[test]
986 fn test_empty_string_signature_normalized_to_none() {
987 let mut mapper = GoogleEventMapper::new();
988
989 let response = GenerateContentResponse {
990 candidates: Some(vec![GenerateContentCandidate {
991 index: Some(0),
992 content: Content {
993 parts: vec![Part::FunctionCallPart(FunctionCallPart {
994 function_call: FunctionCall {
995 name: "test_function".to_string(),
996 args: json!({"arg": "value"}),
997 },
998 thought_signature: Some("".to_string()),
999 })],
1000 role: GoogleRole::Model,
1001 },
1002 finish_reason: None,
1003 finish_message: None,
1004 safety_ratings: None,
1005 citation_metadata: None,
1006 }]),
1007 prompt_feedback: None,
1008 usage_metadata: None,
1009 };
1010
1011 let events = mapper.map_event(response);
1012
1013 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1014 assert_eq!(tool_use.thought_signature, None);
1015 } else {
1016 panic!("Expected ToolUse event");
1017 }
1018 }
1019
1020 #[test]
1021 fn test_parallel_function_calls_preserve_signatures() {
1022 let mut mapper = GoogleEventMapper::new();
1023
1024 let response = GenerateContentResponse {
1025 candidates: Some(vec![GenerateContentCandidate {
1026 index: Some(0),
1027 content: Content {
1028 parts: vec![
1029 Part::FunctionCallPart(FunctionCallPart {
1030 function_call: FunctionCall {
1031 name: "function_1".to_string(),
1032 args: json!({"arg": "value1"}),
1033 },
1034 thought_signature: Some("signature_1".to_string()),
1035 }),
1036 Part::FunctionCallPart(FunctionCallPart {
1037 function_call: FunctionCall {
1038 name: "function_2".to_string(),
1039 args: json!({"arg": "value2"}),
1040 },
1041 thought_signature: None,
1042 }),
1043 ],
1044 role: GoogleRole::Model,
1045 },
1046 finish_reason: None,
1047 finish_message: None,
1048 safety_ratings: None,
1049 citation_metadata: None,
1050 }]),
1051 prompt_feedback: None,
1052 usage_metadata: None,
1053 };
1054
1055 let events = mapper.map_event(response);
1056
1057 assert_eq!(events.len(), 3); // 2 ToolUse events + Stop event
1058
1059 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1060 assert_eq!(tool_use.name.as_ref(), "function_1");
1061 assert_eq!(tool_use.thought_signature.as_deref(), Some("signature_1"));
1062 } else {
1063 panic!("Expected ToolUse event for function_1");
1064 }
1065
1066 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[1] {
1067 assert_eq!(tool_use.name.as_ref(), "function_2");
1068 assert_eq!(tool_use.thought_signature, None);
1069 } else {
1070 panic!("Expected ToolUse event for function_2");
1071 }
1072 }
1073
1074 #[test]
1075 fn test_tool_use_with_signature_converts_to_function_call_part() {
1076 let tool_use = language_model::LanguageModelToolUse {
1077 id: LanguageModelToolUseId::from("test_id"),
1078 name: "test_function".into(),
1079 raw_input: json!({"arg": "value"}).to_string(),
1080 input: json!({"arg": "value"}),
1081 is_input_complete: true,
1082 thought_signature: Some("test_signature_456".to_string()),
1083 };
1084
1085 let request = super::into_google(
1086 LanguageModelRequest {
1087 messages: vec![language_model::LanguageModelRequestMessage {
1088 role: Role::Assistant,
1089 content: vec![MessageContent::ToolUse(tool_use)],
1090 cache: false,
1091 reasoning_details: None,
1092 }],
1093 ..Default::default()
1094 },
1095 "gemini-2.5-flash".to_string(),
1096 GoogleModelMode::Default,
1097 );
1098
1099 assert_eq!(request.contents[0].parts.len(), 1);
1100 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1101 assert_eq!(fc_part.function_call.name, "test_function");
1102 assert_eq!(
1103 fc_part.thought_signature.as_deref(),
1104 Some("test_signature_456")
1105 );
1106 } else {
1107 panic!("Expected FunctionCallPart");
1108 }
1109 }
1110
1111 #[test]
1112 fn test_tool_use_without_signature_omits_field() {
1113 let tool_use = language_model::LanguageModelToolUse {
1114 id: LanguageModelToolUseId::from("test_id"),
1115 name: "test_function".into(),
1116 raw_input: json!({"arg": "value"}).to_string(),
1117 input: json!({"arg": "value"}),
1118 is_input_complete: true,
1119 thought_signature: None,
1120 };
1121
1122 let request = super::into_google(
1123 LanguageModelRequest {
1124 messages: vec![language_model::LanguageModelRequestMessage {
1125 role: Role::Assistant,
1126 content: vec![MessageContent::ToolUse(tool_use)],
1127 cache: false,
1128 reasoning_details: None,
1129 }],
1130 ..Default::default()
1131 },
1132 "gemini-2.5-flash".to_string(),
1133 GoogleModelMode::Default,
1134 );
1135
1136 assert_eq!(request.contents[0].parts.len(), 1);
1137 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1138 assert_eq!(fc_part.thought_signature, None);
1139 } else {
1140 panic!("Expected FunctionCallPart");
1141 }
1142 }
1143
1144 #[test]
1145 fn test_empty_signature_in_tool_use_normalized_to_none() {
1146 let tool_use = language_model::LanguageModelToolUse {
1147 id: LanguageModelToolUseId::from("test_id"),
1148 name: "test_function".into(),
1149 raw_input: json!({"arg": "value"}).to_string(),
1150 input: json!({"arg": "value"}),
1151 is_input_complete: true,
1152 thought_signature: Some("".to_string()),
1153 };
1154
1155 let request = super::into_google(
1156 LanguageModelRequest {
1157 messages: vec![language_model::LanguageModelRequestMessage {
1158 role: Role::Assistant,
1159 content: vec![MessageContent::ToolUse(tool_use)],
1160 cache: false,
1161 reasoning_details: None,
1162 }],
1163 ..Default::default()
1164 },
1165 "gemini-2.5-flash".to_string(),
1166 GoogleModelMode::Default,
1167 );
1168
1169 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1170 assert_eq!(fc_part.thought_signature, None);
1171 } else {
1172 panic!("Expected FunctionCallPart");
1173 }
1174 }
1175
1176 #[test]
1177 fn test_round_trip_preserves_signature() {
1178 let mut mapper = GoogleEventMapper::new();
1179
1180 // Simulate receiving a response from Google with a signature
1181 let response = GenerateContentResponse {
1182 candidates: Some(vec![GenerateContentCandidate {
1183 index: Some(0),
1184 content: Content {
1185 parts: vec![Part::FunctionCallPart(FunctionCallPart {
1186 function_call: FunctionCall {
1187 name: "test_function".to_string(),
1188 args: json!({"arg": "value"}),
1189 },
1190 thought_signature: Some("round_trip_sig".to_string()),
1191 })],
1192 role: GoogleRole::Model,
1193 },
1194 finish_reason: None,
1195 finish_message: None,
1196 safety_ratings: None,
1197 citation_metadata: None,
1198 }]),
1199 prompt_feedback: None,
1200 usage_metadata: None,
1201 };
1202
1203 let events = mapper.map_event(response);
1204
1205 let tool_use = if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1206 tool_use.clone()
1207 } else {
1208 panic!("Expected ToolUse event");
1209 };
1210
1211 // Convert back to Google format
1212 let request = super::into_google(
1213 LanguageModelRequest {
1214 messages: vec![language_model::LanguageModelRequestMessage {
1215 role: Role::Assistant,
1216 content: vec![MessageContent::ToolUse(tool_use)],
1217 cache: false,
1218 reasoning_details: None,
1219 }],
1220 ..Default::default()
1221 },
1222 "gemini-2.5-flash".to_string(),
1223 GoogleModelMode::Default,
1224 );
1225
1226 // Verify signature is preserved
1227 if let Part::FunctionCallPart(fc_part) = &request.contents[0].parts[0] {
1228 assert_eq!(fc_part.thought_signature.as_deref(), Some("round_trip_sig"));
1229 } else {
1230 panic!("Expected FunctionCallPart");
1231 }
1232 }
1233
1234 #[test]
1235 fn test_mixed_text_and_function_call_with_signature() {
1236 let mut mapper = GoogleEventMapper::new();
1237
1238 let response = GenerateContentResponse {
1239 candidates: Some(vec![GenerateContentCandidate {
1240 index: Some(0),
1241 content: Content {
1242 parts: vec![
1243 Part::TextPart(TextPart {
1244 text: "I'll help with that.".to_string(),
1245 }),
1246 Part::FunctionCallPart(FunctionCallPart {
1247 function_call: FunctionCall {
1248 name: "helper_function".to_string(),
1249 args: json!({"query": "help"}),
1250 },
1251 thought_signature: Some("mixed_sig".to_string()),
1252 }),
1253 ],
1254 role: GoogleRole::Model,
1255 },
1256 finish_reason: None,
1257 finish_message: None,
1258 safety_ratings: None,
1259 citation_metadata: None,
1260 }]),
1261 prompt_feedback: None,
1262 usage_metadata: None,
1263 };
1264
1265 let events = mapper.map_event(response);
1266
1267 assert_eq!(events.len(), 3); // Text event + ToolUse event + Stop event
1268
1269 if let Ok(LanguageModelCompletionEvent::Text(text)) = &events[0] {
1270 assert_eq!(text, "I'll help with that.");
1271 } else {
1272 panic!("Expected Text event");
1273 }
1274
1275 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[1] {
1276 assert_eq!(tool_use.name.as_ref(), "helper_function");
1277 assert_eq!(tool_use.thought_signature.as_deref(), Some("mixed_sig"));
1278 } else {
1279 panic!("Expected ToolUse event");
1280 }
1281 }
1282
1283 #[test]
1284 fn test_special_characters_in_signature_preserved() {
1285 let mut mapper = GoogleEventMapper::new();
1286
1287 let signature_with_special_chars = "sig<>\"'&%$#@!{}[]".to_string();
1288
1289 let response = GenerateContentResponse {
1290 candidates: Some(vec![GenerateContentCandidate {
1291 index: Some(0),
1292 content: Content {
1293 parts: vec![Part::FunctionCallPart(FunctionCallPart {
1294 function_call: FunctionCall {
1295 name: "test_function".to_string(),
1296 args: json!({"arg": "value"}),
1297 },
1298 thought_signature: Some(signature_with_special_chars.clone()),
1299 })],
1300 role: GoogleRole::Model,
1301 },
1302 finish_reason: None,
1303 finish_message: None,
1304 safety_ratings: None,
1305 citation_metadata: None,
1306 }]),
1307 prompt_feedback: None,
1308 usage_metadata: None,
1309 };
1310
1311 let events = mapper.map_event(response);
1312
1313 if let Ok(LanguageModelCompletionEvent::ToolUse(tool_use)) = &events[0] {
1314 assert_eq!(
1315 tool_use.thought_signature.as_deref(),
1316 Some(signature_with_special_chars.as_str())
1317 );
1318 } else {
1319 panic!("Expected ToolUse event");
1320 }
1321 }
1322}