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