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