1use std::pin::Pin;
2use std::str::FromStr;
3use std::sync::Arc;
4
5use crate::ui::InstructionListItem;
6use anyhow::{Context as _, Result, anyhow};
7use aws_config::stalled_stream_protection::StalledStreamProtectionConfig;
8use aws_config::{BehaviorVersion, Region};
9use aws_credential_types::Credentials;
10use aws_http_client::AwsHttpClient;
11use bedrock::bedrock_client::Client as BedrockClient;
12use bedrock::bedrock_client::config::timeout::TimeoutConfig;
13use bedrock::bedrock_client::types::{
14 ContentBlockDelta, ContentBlockStart, ConverseStreamOutput, ReasoningContentBlockDelta,
15 StopReason,
16};
17use bedrock::{
18 BedrockAnyToolChoice, BedrockAutoToolChoice, BedrockBlob, BedrockError, BedrockInnerContent,
19 BedrockMessage, BedrockModelMode, BedrockStreamingResponse, BedrockThinkingBlock,
20 BedrockThinkingTextBlock, BedrockTool, BedrockToolChoice, BedrockToolConfig,
21 BedrockToolInputSchema, BedrockToolResultBlock, BedrockToolResultContentBlock,
22 BedrockToolResultStatus, BedrockToolSpec, BedrockToolUseBlock, Model, value_to_aws_document,
23};
24use collections::{BTreeMap, HashMap};
25use credentials_provider::CredentialsProvider;
26use editor::{Editor, EditorElement, EditorStyle};
27use futures::{FutureExt, Stream, StreamExt, future::BoxFuture, stream::BoxStream};
28use gpui::{
29 AnyView, App, AsyncApp, Context, Entity, FontStyle, FontWeight, Subscription, Task, TextStyle,
30 WhiteSpace,
31};
32use gpui_tokio::Tokio;
33use http_client::HttpClient;
34use language_model::{
35 AuthenticateError, LanguageModel, LanguageModelCacheConfiguration,
36 LanguageModelCompletionError, LanguageModelCompletionEvent, LanguageModelId, LanguageModelName,
37 LanguageModelProvider, LanguageModelProviderId, LanguageModelProviderName,
38 LanguageModelProviderState, LanguageModelRequest, LanguageModelToolChoice,
39 LanguageModelToolResultContent, LanguageModelToolUse, MessageContent, RateLimiter, Role,
40 TokenUsage,
41};
42use schemars::JsonSchema;
43use serde::{Deserialize, Serialize};
44use serde_json::Value;
45use settings::{Settings, SettingsStore};
46use smol::lock::OnceCell;
47use strum::{EnumIter, IntoEnumIterator, IntoStaticStr};
48use theme::ThemeSettings;
49use tokio::runtime::Handle;
50use ui::{Icon, IconName, List, Tooltip, prelude::*};
51use util::{ResultExt, default};
52
53use crate::AllLanguageModelSettings;
54
55const PROVIDER_ID: &str = "amazon-bedrock";
56const PROVIDER_NAME: &str = "Amazon Bedrock";
57
58#[derive(Default, Clone, Deserialize, Serialize, PartialEq, Debug)]
59pub struct BedrockCredentials {
60 pub access_key_id: String,
61 pub secret_access_key: String,
62 pub session_token: Option<String>,
63 pub region: String,
64}
65
66#[derive(Default, Clone, Debug, PartialEq)]
67pub struct AmazonBedrockSettings {
68 pub available_models: Vec<AvailableModel>,
69 pub region: Option<String>,
70 pub endpoint: Option<String>,
71 pub profile_name: Option<String>,
72 pub role_arn: Option<String>,
73 pub authentication_method: Option<BedrockAuthMethod>,
74}
75
76#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, EnumIter, IntoStaticStr, JsonSchema)]
77pub enum BedrockAuthMethod {
78 #[serde(rename = "named_profile")]
79 NamedProfile,
80 #[serde(rename = "sso")]
81 SingleSignOn,
82 /// IMDSv2, PodIdentity, env vars, etc.
83 #[serde(rename = "default")]
84 Automatic,
85}
86
87#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
88pub struct AvailableModel {
89 pub name: String,
90 pub display_name: Option<String>,
91 pub max_tokens: u64,
92 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
93 pub max_output_tokens: Option<u64>,
94 pub default_temperature: Option<f32>,
95 pub mode: Option<ModelMode>,
96}
97
98#[derive(Clone, Debug, Default, PartialEq, Serialize, Deserialize, JsonSchema)]
99#[serde(tag = "type", rename_all = "lowercase")]
100pub enum ModelMode {
101 #[default]
102 Default,
103 Thinking {
104 /// The maximum number of tokens to use for reasoning. Must be lower than the model's `max_output_tokens`.
105 budget_tokens: Option<u64>,
106 },
107}
108
109impl From<ModelMode> for BedrockModelMode {
110 fn from(value: ModelMode) -> Self {
111 match value {
112 ModelMode::Default => BedrockModelMode::Default,
113 ModelMode::Thinking { budget_tokens } => BedrockModelMode::Thinking { budget_tokens },
114 }
115 }
116}
117
118impl From<BedrockModelMode> for ModelMode {
119 fn from(value: BedrockModelMode) -> Self {
120 match value {
121 BedrockModelMode::Default => ModelMode::Default,
122 BedrockModelMode::Thinking { budget_tokens } => ModelMode::Thinking { budget_tokens },
123 }
124 }
125}
126
127/// The URL of the base AWS service.
128///
129/// Right now we're just using this as the key to store the AWS credentials
130/// under in the keychain.
131const AMAZON_AWS_URL: &str = "https://amazonaws.com";
132
133// These environment variables all use a `ZED_` prefix because we don't want to overwrite the user's AWS credentials.
134const ZED_BEDROCK_ACCESS_KEY_ID_VAR: &str = "ZED_ACCESS_KEY_ID";
135const ZED_BEDROCK_SECRET_ACCESS_KEY_VAR: &str = "ZED_SECRET_ACCESS_KEY";
136const ZED_BEDROCK_SESSION_TOKEN_VAR: &str = "ZED_SESSION_TOKEN";
137const ZED_AWS_PROFILE_VAR: &str = "ZED_AWS_PROFILE";
138const ZED_BEDROCK_REGION_VAR: &str = "ZED_AWS_REGION";
139const ZED_AWS_CREDENTIALS_VAR: &str = "ZED_AWS_CREDENTIALS";
140const ZED_AWS_ENDPOINT_VAR: &str = "ZED_AWS_ENDPOINT";
141
142pub struct State {
143 credentials: Option<BedrockCredentials>,
144 settings: Option<AmazonBedrockSettings>,
145 credentials_from_env: bool,
146 _subscription: Subscription,
147}
148
149impl State {
150 fn reset_credentials(&self, cx: &mut Context<Self>) -> Task<Result<()>> {
151 let credentials_provider = <dyn CredentialsProvider>::global(cx);
152 cx.spawn(async move |this, cx| {
153 credentials_provider
154 .delete_credentials(AMAZON_AWS_URL, &cx)
155 .await
156 .log_err();
157 this.update(cx, |this, cx| {
158 this.credentials = None;
159 this.credentials_from_env = false;
160 this.settings = None;
161 cx.notify();
162 })
163 })
164 }
165
166 fn set_credentials(
167 &mut self,
168 credentials: BedrockCredentials,
169 cx: &mut Context<Self>,
170 ) -> Task<Result<()>> {
171 let credentials_provider = <dyn CredentialsProvider>::global(cx);
172 cx.spawn(async move |this, cx| {
173 credentials_provider
174 .write_credentials(
175 AMAZON_AWS_URL,
176 "Bearer",
177 &serde_json::to_vec(&credentials)?,
178 &cx,
179 )
180 .await?;
181 this.update(cx, |this, cx| {
182 this.credentials = Some(credentials);
183 cx.notify();
184 })
185 })
186 }
187
188 fn is_authenticated(&self) -> bool {
189 let derived = self
190 .settings
191 .as_ref()
192 .and_then(|s| s.authentication_method.as_ref());
193 let creds = self.credentials.as_ref();
194
195 derived.is_some() || creds.is_some()
196 }
197
198 fn authenticate(&self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
199 if self.is_authenticated() {
200 return Task::ready(Ok(()));
201 }
202
203 let credentials_provider = <dyn CredentialsProvider>::global(cx);
204 cx.spawn(async move |this, cx| {
205 let (credentials, from_env) =
206 if let Ok(credentials) = std::env::var(ZED_AWS_CREDENTIALS_VAR) {
207 (credentials, true)
208 } else {
209 let (_, credentials) = credentials_provider
210 .read_credentials(AMAZON_AWS_URL, &cx)
211 .await?
212 .ok_or_else(|| AuthenticateError::CredentialsNotFound)?;
213 (
214 String::from_utf8(credentials)
215 .context("invalid {PROVIDER_NAME} credentials")?,
216 false,
217 )
218 };
219
220 let credentials: BedrockCredentials =
221 serde_json::from_str(&credentials).context("failed to parse credentials")?;
222
223 this.update(cx, |this, cx| {
224 this.credentials = Some(credentials);
225 this.credentials_from_env = from_env;
226 cx.notify();
227 })?;
228
229 Ok(())
230 })
231 }
232}
233
234pub struct BedrockLanguageModelProvider {
235 http_client: AwsHttpClient,
236 handler: tokio::runtime::Handle,
237 state: gpui::Entity<State>,
238}
239
240impl BedrockLanguageModelProvider {
241 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
242 let state = cx.new(|cx| State {
243 credentials: None,
244 settings: Some(AllLanguageModelSettings::get_global(cx).bedrock.clone()),
245 credentials_from_env: false,
246 _subscription: cx.observe_global::<SettingsStore>(|_, cx| {
247 cx.notify();
248 }),
249 });
250
251 let tokio_handle = Tokio::handle(cx);
252
253 let coerced_client = AwsHttpClient::new(http_client.clone(), tokio_handle.clone());
254
255 Self {
256 http_client: coerced_client,
257 handler: tokio_handle.clone(),
258 state,
259 }
260 }
261
262 fn create_language_model(&self, model: bedrock::Model) -> Arc<dyn LanguageModel> {
263 Arc::new(BedrockModel {
264 id: LanguageModelId::from(model.id().to_string()),
265 model,
266 http_client: self.http_client.clone(),
267 handler: self.handler.clone(),
268 state: self.state.clone(),
269 client: OnceCell::new(),
270 request_limiter: RateLimiter::new(4),
271 })
272 }
273}
274
275impl LanguageModelProvider for BedrockLanguageModelProvider {
276 fn id(&self) -> LanguageModelProviderId {
277 LanguageModelProviderId(PROVIDER_ID.into())
278 }
279
280 fn name(&self) -> LanguageModelProviderName {
281 LanguageModelProviderName(PROVIDER_NAME.into())
282 }
283
284 fn icon(&self) -> IconName {
285 IconName::AiBedrock
286 }
287
288 fn default_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
289 Some(self.create_language_model(bedrock::Model::default()))
290 }
291
292 fn default_fast_model(&self, _cx: &App) -> Option<Arc<dyn LanguageModel>> {
293 Some(self.create_language_model(bedrock::Model::default_fast()))
294 }
295
296 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
297 let mut models = BTreeMap::default();
298
299 for model in bedrock::Model::iter() {
300 if !matches!(model, bedrock::Model::Custom { .. }) {
301 // TODO: Sonnet 3.7 vs. 3.7 Thinking bug is here.
302 models.insert(model.id().to_string(), model);
303 }
304 }
305
306 // Override with available models from settings
307 for model in AllLanguageModelSettings::get_global(cx)
308 .bedrock
309 .available_models
310 .iter()
311 {
312 models.insert(
313 model.name.clone(),
314 bedrock::Model::Custom {
315 name: model.name.clone(),
316 display_name: model.display_name.clone(),
317 max_tokens: model.max_tokens,
318 max_output_tokens: model.max_output_tokens,
319 default_temperature: model.default_temperature,
320 },
321 );
322 }
323
324 models
325 .into_values()
326 .map(|model| self.create_language_model(model))
327 .collect()
328 }
329
330 fn is_authenticated(&self, cx: &App) -> bool {
331 self.state.read(cx).is_authenticated()
332 }
333
334 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
335 self.state.update(cx, |state, cx| state.authenticate(cx))
336 }
337
338 fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView {
339 cx.new(|cx| ConfigurationView::new(self.state.clone(), window, cx))
340 .into()
341 }
342
343 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
344 self.state
345 .update(cx, |state, cx| state.reset_credentials(cx))
346 }
347}
348
349impl LanguageModelProviderState for BedrockLanguageModelProvider {
350 type ObservableEntity = State;
351
352 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
353 Some(self.state.clone())
354 }
355}
356
357struct BedrockModel {
358 id: LanguageModelId,
359 model: Model,
360 http_client: AwsHttpClient,
361 handler: tokio::runtime::Handle,
362 client: OnceCell<BedrockClient>,
363 state: gpui::Entity<State>,
364 request_limiter: RateLimiter,
365}
366
367impl BedrockModel {
368 fn get_or_init_client(&self, cx: &AsyncApp) -> anyhow::Result<&BedrockClient> {
369 self.client
370 .get_or_try_init_blocking(|| {
371 let (auth_method, credentials, endpoint, region, settings) =
372 cx.read_entity(&self.state, |state, _cx| {
373 let auth_method = state
374 .settings
375 .as_ref()
376 .and_then(|s| s.authentication_method.clone());
377
378 let endpoint = state.settings.as_ref().and_then(|s| s.endpoint.clone());
379
380 let region = state
381 .settings
382 .as_ref()
383 .and_then(|s| s.region.clone())
384 .unwrap_or(String::from("us-east-1"));
385
386 (
387 auth_method,
388 state.credentials.clone(),
389 endpoint,
390 region,
391 state.settings.clone(),
392 )
393 })?;
394
395 let mut config_builder = aws_config::defaults(BehaviorVersion::latest())
396 .stalled_stream_protection(StalledStreamProtectionConfig::disabled())
397 .http_client(self.http_client.clone())
398 .region(Region::new(region))
399 .timeout_config(TimeoutConfig::disabled());
400
401 if let Some(endpoint_url) = endpoint {
402 if !endpoint_url.is_empty() {
403 config_builder = config_builder.endpoint_url(endpoint_url);
404 }
405 }
406
407 match auth_method {
408 None => {
409 if let Some(creds) = credentials {
410 let aws_creds = Credentials::new(
411 creds.access_key_id,
412 creds.secret_access_key,
413 creds.session_token,
414 None,
415 "zed-bedrock-provider",
416 );
417 config_builder = config_builder.credentials_provider(aws_creds);
418 }
419 }
420 Some(BedrockAuthMethod::NamedProfile)
421 | Some(BedrockAuthMethod::SingleSignOn) => {
422 // Currently NamedProfile and SSO behave the same way but only the instructions change
423 // Until we support BearerAuth through SSO, this will not change.
424 let profile_name = settings
425 .and_then(|s| s.profile_name)
426 .unwrap_or_else(|| "default".to_string());
427
428 if !profile_name.is_empty() {
429 config_builder = config_builder.profile_name(profile_name);
430 }
431 }
432 Some(BedrockAuthMethod::Automatic) => {
433 // Use default credential provider chain
434 }
435 }
436
437 let config = self.handler.block_on(config_builder.load());
438 anyhow::Ok(BedrockClient::new(&config))
439 })
440 .context("initializing Bedrock client")?;
441
442 self.client.get().context("Bedrock client not initialized")
443 }
444
445 fn stream_completion(
446 &self,
447 request: bedrock::Request,
448 cx: &AsyncApp,
449 ) -> Result<
450 BoxFuture<'static, BoxStream<'static, Result<BedrockStreamingResponse, BedrockError>>>,
451 > {
452 let runtime_client = self
453 .get_or_init_client(cx)
454 .cloned()
455 .context("Bedrock client not initialized")?;
456 let owned_handle = self.handler.clone();
457
458 Ok(async move {
459 let request = bedrock::stream_completion(runtime_client, request, owned_handle);
460 request.await.unwrap_or_else(|e| {
461 futures::stream::once(async move { Err(BedrockError::ClientError(e)) }).boxed()
462 })
463 }
464 .boxed())
465 }
466}
467
468impl LanguageModel for BedrockModel {
469 fn id(&self) -> LanguageModelId {
470 self.id.clone()
471 }
472
473 fn name(&self) -> LanguageModelName {
474 LanguageModelName::from(self.model.display_name().to_string())
475 }
476
477 fn provider_id(&self) -> LanguageModelProviderId {
478 LanguageModelProviderId(PROVIDER_ID.into())
479 }
480
481 fn provider_name(&self) -> LanguageModelProviderName {
482 LanguageModelProviderName(PROVIDER_NAME.into())
483 }
484
485 fn supports_tools(&self) -> bool {
486 self.model.supports_tool_use()
487 }
488
489 fn supports_images(&self) -> bool {
490 false
491 }
492
493 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
494 match choice {
495 LanguageModelToolChoice::Auto | LanguageModelToolChoice::Any => {
496 self.model.supports_tool_use()
497 }
498 LanguageModelToolChoice::None => false,
499 }
500 }
501
502 fn telemetry_id(&self) -> String {
503 format!("bedrock/{}", self.model.id())
504 }
505
506 fn max_token_count(&self) -> u64 {
507 self.model.max_token_count()
508 }
509
510 fn max_output_tokens(&self) -> Option<u64> {
511 Some(self.model.max_output_tokens())
512 }
513
514 fn count_tokens(
515 &self,
516 request: LanguageModelRequest,
517 cx: &App,
518 ) -> BoxFuture<'static, Result<u64>> {
519 get_bedrock_tokens(request, cx)
520 }
521
522 fn stream_completion(
523 &self,
524 request: LanguageModelRequest,
525 cx: &AsyncApp,
526 ) -> BoxFuture<
527 'static,
528 Result<
529 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
530 LanguageModelCompletionError,
531 >,
532 > {
533 let Ok(region) = cx.read_entity(&self.state, |state, _cx| {
534 // Get region - from credentials or directly from settings
535 let credentials_region = state.credentials.as_ref().map(|s| s.region.clone());
536 let settings_region = state.settings.as_ref().and_then(|s| s.region.clone());
537
538 // Use credentials region if available, otherwise use settings region, finally fall back to default
539 credentials_region
540 .or(settings_region)
541 .unwrap_or(String::from("us-east-1"))
542 }) else {
543 return async move { Err(anyhow::anyhow!("App State Dropped").into()) }.boxed();
544 };
545
546 let model_id = match self.model.cross_region_inference_id(®ion) {
547 Ok(s) => s,
548 Err(e) => {
549 return async move { Err(e.into()) }.boxed();
550 }
551 };
552
553 let request = match into_bedrock(
554 request,
555 model_id,
556 self.model.default_temperature(),
557 self.model.max_output_tokens(),
558 self.model.mode(),
559 ) {
560 Ok(request) => request,
561 Err(err) => return futures::future::ready(Err(err.into())).boxed(),
562 };
563
564 let owned_handle = self.handler.clone();
565
566 let request = self.stream_completion(request, cx);
567 let future = self.request_limiter.stream(async move {
568 let response = request.map_err(|err| anyhow!(err))?.await;
569 Ok(map_to_language_model_completion_events(
570 response,
571 owned_handle,
572 ))
573 });
574 async move { Ok(future.await?.boxed()) }.boxed()
575 }
576
577 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
578 None
579 }
580}
581
582pub fn into_bedrock(
583 request: LanguageModelRequest,
584 model: String,
585 default_temperature: f32,
586 max_output_tokens: u64,
587 mode: BedrockModelMode,
588) -> Result<bedrock::Request> {
589 let mut new_messages: Vec<BedrockMessage> = Vec::new();
590 let mut system_message = String::new();
591
592 for message in request.messages {
593 if message.contents_empty() {
594 continue;
595 }
596
597 match message.role {
598 Role::User | Role::Assistant => {
599 let bedrock_message_content: Vec<BedrockInnerContent> = message
600 .content
601 .into_iter()
602 .filter_map(|content| match content {
603 MessageContent::Text(text) => {
604 if !text.is_empty() {
605 Some(BedrockInnerContent::Text(text))
606 } else {
607 None
608 }
609 }
610 MessageContent::Thinking { text, signature } => {
611 let thinking = BedrockThinkingTextBlock::builder()
612 .text(text)
613 .set_signature(signature)
614 .build()
615 .context("failed to build reasoning block")
616 .log_err()?;
617
618 Some(BedrockInnerContent::ReasoningContent(
619 BedrockThinkingBlock::ReasoningText(thinking),
620 ))
621 }
622 MessageContent::RedactedThinking(blob) => {
623 let redacted =
624 BedrockThinkingBlock::RedactedContent(BedrockBlob::new(blob));
625
626 Some(BedrockInnerContent::ReasoningContent(redacted))
627 }
628 MessageContent::ToolUse(tool_use) => BedrockToolUseBlock::builder()
629 .name(tool_use.name.to_string())
630 .tool_use_id(tool_use.id.to_string())
631 .input(value_to_aws_document(&tool_use.input))
632 .build()
633 .context("failed to build Bedrock tool use block")
634 .log_err()
635 .map(BedrockInnerContent::ToolUse),
636 MessageContent::ToolResult(tool_result) => {
637 BedrockToolResultBlock::builder()
638 .tool_use_id(tool_result.tool_use_id.to_string())
639 .content(match tool_result.content {
640 LanguageModelToolResultContent::Text(text) => {
641 BedrockToolResultContentBlock::Text(text.to_string())
642 }
643 LanguageModelToolResultContent::Image(_) => {
644 BedrockToolResultContentBlock::Text(
645 // TODO: Bedrock image support
646 "[Tool responded with an image, but Zed doesn't support these in Bedrock models yet]".to_string()
647 )
648 }
649 })
650 .status({
651 if tool_result.is_error {
652 BedrockToolResultStatus::Error
653 } else {
654 BedrockToolResultStatus::Success
655 }
656 })
657 .build()
658 .context("failed to build Bedrock tool result block")
659 .log_err()
660 .map(BedrockInnerContent::ToolResult)
661 }
662 _ => None,
663 })
664 .collect();
665 let bedrock_role = match message.role {
666 Role::User => bedrock::BedrockRole::User,
667 Role::Assistant => bedrock::BedrockRole::Assistant,
668 Role::System => unreachable!("System role should never occur here"),
669 };
670 if let Some(last_message) = new_messages.last_mut() {
671 if last_message.role == bedrock_role {
672 last_message.content.extend(bedrock_message_content);
673 continue;
674 }
675 }
676 new_messages.push(
677 BedrockMessage::builder()
678 .role(bedrock_role)
679 .set_content(Some(bedrock_message_content))
680 .build()
681 .context("failed to build Bedrock message")?,
682 );
683 }
684 Role::System => {
685 if !system_message.is_empty() {
686 system_message.push_str("\n\n");
687 }
688 system_message.push_str(&message.string_contents());
689 }
690 }
691 }
692
693 let tool_spec: Vec<BedrockTool> = request
694 .tools
695 .iter()
696 .filter_map(|tool| {
697 Some(BedrockTool::ToolSpec(
698 BedrockToolSpec::builder()
699 .name(tool.name.clone())
700 .description(tool.description.clone())
701 .input_schema(BedrockToolInputSchema::Json(value_to_aws_document(
702 &tool.input_schema,
703 )))
704 .build()
705 .log_err()?,
706 ))
707 })
708 .collect();
709
710 let tool_choice = match request.tool_choice {
711 Some(LanguageModelToolChoice::Auto) | None => {
712 BedrockToolChoice::Auto(BedrockAutoToolChoice::builder().build())
713 }
714 Some(LanguageModelToolChoice::Any) => {
715 BedrockToolChoice::Any(BedrockAnyToolChoice::builder().build())
716 }
717 Some(LanguageModelToolChoice::None) => {
718 anyhow::bail!("LanguageModelToolChoice::None is not supported");
719 }
720 };
721 let tool_config: BedrockToolConfig = BedrockToolConfig::builder()
722 .set_tools(Some(tool_spec))
723 .tool_choice(tool_choice)
724 .build()?;
725
726 Ok(bedrock::Request {
727 model,
728 messages: new_messages,
729 max_tokens: max_output_tokens,
730 system: Some(system_message),
731 tools: Some(tool_config),
732 thinking: if let BedrockModelMode::Thinking { budget_tokens } = mode {
733 Some(bedrock::Thinking::Enabled { budget_tokens })
734 } else {
735 None
736 },
737 metadata: None,
738 stop_sequences: Vec::new(),
739 temperature: request.temperature.or(Some(default_temperature)),
740 top_k: None,
741 top_p: None,
742 })
743}
744
745// TODO: just call the ConverseOutput.usage() method:
746// https://docs.rs/aws-sdk-bedrockruntime/latest/aws_sdk_bedrockruntime/operation/converse/struct.ConverseOutput.html#method.output
747pub fn get_bedrock_tokens(
748 request: LanguageModelRequest,
749 cx: &App,
750) -> BoxFuture<'static, Result<u64>> {
751 cx.background_executor()
752 .spawn(async move {
753 let messages = request.messages;
754 let mut tokens_from_images = 0;
755 let mut string_messages = Vec::with_capacity(messages.len());
756
757 for message in messages {
758 use language_model::MessageContent;
759
760 let mut string_contents = String::new();
761
762 for content in message.content {
763 match content {
764 MessageContent::Text(text) | MessageContent::Thinking { text, .. } => {
765 string_contents.push_str(&text);
766 }
767 MessageContent::RedactedThinking(_) => {}
768 MessageContent::Image(image) => {
769 tokens_from_images += image.estimate_tokens();
770 }
771 MessageContent::ToolUse(_tool_use) => {
772 // TODO: Estimate token usage from tool uses.
773 }
774 MessageContent::ToolResult(tool_result) => match tool_result.content {
775 LanguageModelToolResultContent::Text(text) => {
776 string_contents.push_str(&text);
777 }
778 LanguageModelToolResultContent::Image(image) => {
779 tokens_from_images += image.estimate_tokens();
780 }
781 },
782 }
783 }
784
785 if !string_contents.is_empty() {
786 string_messages.push(tiktoken_rs::ChatCompletionRequestMessage {
787 role: match message.role {
788 Role::User => "user".into(),
789 Role::Assistant => "assistant".into(),
790 Role::System => "system".into(),
791 },
792 content: Some(string_contents),
793 name: None,
794 function_call: None,
795 });
796 }
797 }
798
799 // Tiktoken doesn't yet support these models, so we manually use the
800 // same tokenizer as GPT-4.
801 tiktoken_rs::num_tokens_from_messages("gpt-4", &string_messages)
802 .map(|tokens| (tokens + tokens_from_images) as u64)
803 })
804 .boxed()
805}
806
807pub fn map_to_language_model_completion_events(
808 events: Pin<Box<dyn Send + Stream<Item = Result<BedrockStreamingResponse, BedrockError>>>>,
809 handle: Handle,
810) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
811 struct RawToolUse {
812 id: String,
813 name: String,
814 input_json: String,
815 }
816
817 struct State {
818 events: Pin<Box<dyn Send + Stream<Item = Result<BedrockStreamingResponse, BedrockError>>>>,
819 tool_uses_by_index: HashMap<i32, RawToolUse>,
820 }
821
822 futures::stream::unfold(
823 State {
824 events,
825 tool_uses_by_index: HashMap::default(),
826 },
827 move |mut state: State| {
828 let inner_handle = handle.clone();
829 async move {
830 inner_handle
831 .spawn(async {
832 while let Some(event) = state.events.next().await {
833 match event {
834 Ok(event) => match event {
835 ConverseStreamOutput::ContentBlockDelta(cb_delta) => {
836 match cb_delta.delta {
837 Some(ContentBlockDelta::Text(text_out)) => {
838 let completion_event =
839 LanguageModelCompletionEvent::Text(text_out);
840 return Some((Some(Ok(completion_event)), state));
841 }
842
843 Some(ContentBlockDelta::ToolUse(text_out)) => {
844 if let Some(tool_use) = state
845 .tool_uses_by_index
846 .get_mut(&cb_delta.content_block_index)
847 {
848 tool_use.input_json.push_str(text_out.input());
849 }
850 }
851
852 Some(ContentBlockDelta::ReasoningContent(thinking)) => {
853 match thinking {
854 ReasoningContentBlockDelta::RedactedContent(
855 redacted,
856 ) => {
857 let thinking_event =
858 LanguageModelCompletionEvent::Thinking {
859 text: String::from_utf8(
860 redacted.into_inner(),
861 )
862 .unwrap_or("REDACTED".to_string()),
863 signature: None,
864 };
865
866 return Some((
867 Some(Ok(thinking_event)),
868 state,
869 ));
870 }
871 ReasoningContentBlockDelta::Signature(
872 signature,
873 ) => {
874 return Some((
875 Some(Ok(LanguageModelCompletionEvent::Thinking {
876 text: "".to_string(),
877 signature: Some(signature)
878 })),
879 state,
880 ));
881 }
882 ReasoningContentBlockDelta::Text(thoughts) => {
883 let thinking_event =
884 LanguageModelCompletionEvent::Thinking {
885 text: thoughts.to_string(),
886 signature: None
887 };
888
889 return Some((
890 Some(Ok(thinking_event)),
891 state,
892 ));
893 }
894 _ => {}
895 }
896 }
897 _ => {}
898 }
899 }
900 ConverseStreamOutput::ContentBlockStart(cb_start) => {
901 if let Some(ContentBlockStart::ToolUse(text_out)) =
902 cb_start.start
903 {
904 let tool_use = RawToolUse {
905 id: text_out.tool_use_id,
906 name: text_out.name,
907 input_json: String::new(),
908 };
909
910 state
911 .tool_uses_by_index
912 .insert(cb_start.content_block_index, tool_use);
913 }
914 }
915 ConverseStreamOutput::ContentBlockStop(cb_stop) => {
916 if let Some(tool_use) = state
917 .tool_uses_by_index
918 .remove(&cb_stop.content_block_index)
919 {
920 let tool_use_event = LanguageModelToolUse {
921 id: tool_use.id.into(),
922 name: tool_use.name.into(),
923 is_input_complete: true,
924 raw_input: tool_use.input_json.clone(),
925 input: if tool_use.input_json.is_empty() {
926 Value::Null
927 } else {
928 serde_json::Value::from_str(
929 &tool_use.input_json,
930 )
931 .map_err(|err| anyhow!(err))
932 .unwrap()
933 },
934 };
935
936 return Some((
937 Some(Ok(LanguageModelCompletionEvent::ToolUse(
938 tool_use_event,
939 ))),
940 state,
941 ));
942 }
943 }
944
945 ConverseStreamOutput::Metadata(cb_meta) => {
946 if let Some(metadata) = cb_meta.usage {
947 let completion_event =
948 LanguageModelCompletionEvent::UsageUpdate(
949 TokenUsage {
950 input_tokens: metadata.input_tokens as u64,
951 output_tokens: metadata.output_tokens
952 as u64,
953 cache_creation_input_tokens: default(),
954 cache_read_input_tokens: default(),
955 },
956 );
957 return Some((Some(Ok(completion_event)), state));
958 }
959 }
960 ConverseStreamOutput::MessageStop(message_stop) => {
961 let reason = match message_stop.stop_reason {
962 StopReason::ContentFiltered => {
963 LanguageModelCompletionEvent::Stop(
964 language_model::StopReason::EndTurn,
965 )
966 }
967 StopReason::EndTurn => {
968 LanguageModelCompletionEvent::Stop(
969 language_model::StopReason::EndTurn,
970 )
971 }
972 StopReason::GuardrailIntervened => {
973 LanguageModelCompletionEvent::Stop(
974 language_model::StopReason::EndTurn,
975 )
976 }
977 StopReason::MaxTokens => {
978 LanguageModelCompletionEvent::Stop(
979 language_model::StopReason::EndTurn,
980 )
981 }
982 StopReason::StopSequence => {
983 LanguageModelCompletionEvent::Stop(
984 language_model::StopReason::EndTurn,
985 )
986 }
987 StopReason::ToolUse => {
988 LanguageModelCompletionEvent::Stop(
989 language_model::StopReason::ToolUse,
990 )
991 }
992 _ => LanguageModelCompletionEvent::Stop(
993 language_model::StopReason::EndTurn,
994 ),
995 };
996 return Some((Some(Ok(reason)), state));
997 }
998 _ => {}
999 },
1000
1001 Err(err) => return Some((Some(Err(anyhow!(err).into())), state)),
1002 }
1003 }
1004 None
1005 })
1006 .await
1007 .log_err()
1008 .flatten()
1009 }
1010 },
1011 )
1012 .filter_map(|event| async move { event })
1013}
1014
1015struct ConfigurationView {
1016 access_key_id_editor: Entity<Editor>,
1017 secret_access_key_editor: Entity<Editor>,
1018 session_token_editor: Entity<Editor>,
1019 region_editor: Entity<Editor>,
1020 state: gpui::Entity<State>,
1021 load_credentials_task: Option<Task<()>>,
1022}
1023
1024impl ConfigurationView {
1025 const PLACEHOLDER_ACCESS_KEY_ID_TEXT: &'static str = "XXXXXXXXXXXXXXXX";
1026 const PLACEHOLDER_SECRET_ACCESS_KEY_TEXT: &'static str =
1027 "XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX";
1028 const PLACEHOLDER_SESSION_TOKEN_TEXT: &'static str = "XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX";
1029 const PLACEHOLDER_REGION: &'static str = "us-east-1";
1030
1031 fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
1032 cx.observe(&state, |_, _, cx| {
1033 cx.notify();
1034 })
1035 .detach();
1036
1037 let load_credentials_task = Some(cx.spawn({
1038 let state = state.clone();
1039 async move |this, cx| {
1040 if let Some(task) = state
1041 .update(cx, |state, cx| state.authenticate(cx))
1042 .log_err()
1043 {
1044 // We don't log an error, because "not signed in" is also an error.
1045 let _ = task.await;
1046 }
1047 this.update(cx, |this, cx| {
1048 this.load_credentials_task = None;
1049 cx.notify();
1050 })
1051 .log_err();
1052 }
1053 }));
1054
1055 Self {
1056 access_key_id_editor: cx.new(|cx| {
1057 let mut editor = Editor::single_line(window, cx);
1058 editor.set_placeholder_text(Self::PLACEHOLDER_ACCESS_KEY_ID_TEXT, cx);
1059 editor
1060 }),
1061 secret_access_key_editor: cx.new(|cx| {
1062 let mut editor = Editor::single_line(window, cx);
1063 editor.set_placeholder_text(Self::PLACEHOLDER_SECRET_ACCESS_KEY_TEXT, cx);
1064 editor
1065 }),
1066 session_token_editor: cx.new(|cx| {
1067 let mut editor = Editor::single_line(window, cx);
1068 editor.set_placeholder_text(Self::PLACEHOLDER_SESSION_TOKEN_TEXT, cx);
1069 editor
1070 }),
1071 region_editor: cx.new(|cx| {
1072 let mut editor = Editor::single_line(window, cx);
1073 editor.set_placeholder_text(Self::PLACEHOLDER_REGION, cx);
1074 editor
1075 }),
1076 state,
1077 load_credentials_task,
1078 }
1079 }
1080
1081 fn save_credentials(
1082 &mut self,
1083 _: &menu::Confirm,
1084 _window: &mut Window,
1085 cx: &mut Context<Self>,
1086 ) {
1087 let access_key_id = self
1088 .access_key_id_editor
1089 .read(cx)
1090 .text(cx)
1091 .to_string()
1092 .trim()
1093 .to_string();
1094 let secret_access_key = self
1095 .secret_access_key_editor
1096 .read(cx)
1097 .text(cx)
1098 .to_string()
1099 .trim()
1100 .to_string();
1101 let session_token = self
1102 .session_token_editor
1103 .read(cx)
1104 .text(cx)
1105 .to_string()
1106 .trim()
1107 .to_string();
1108 let session_token = if session_token.is_empty() {
1109 None
1110 } else {
1111 Some(session_token)
1112 };
1113 let region = self
1114 .region_editor
1115 .read(cx)
1116 .text(cx)
1117 .to_string()
1118 .trim()
1119 .to_string();
1120 let region = if region.is_empty() {
1121 "us-east-1".to_string()
1122 } else {
1123 region
1124 };
1125
1126 let state = self.state.clone();
1127 cx.spawn(async move |_, cx| {
1128 state
1129 .update(cx, |state, cx| {
1130 let credentials: BedrockCredentials = BedrockCredentials {
1131 region: region.clone(),
1132 access_key_id: access_key_id.clone(),
1133 secret_access_key: secret_access_key.clone(),
1134 session_token: session_token.clone(),
1135 };
1136
1137 state.set_credentials(credentials, cx)
1138 })?
1139 .await
1140 })
1141 .detach_and_log_err(cx);
1142 }
1143
1144 fn reset_credentials(&mut self, window: &mut Window, cx: &mut Context<Self>) {
1145 self.access_key_id_editor
1146 .update(cx, |editor, cx| editor.set_text("", window, cx));
1147 self.secret_access_key_editor
1148 .update(cx, |editor, cx| editor.set_text("", window, cx));
1149 self.session_token_editor
1150 .update(cx, |editor, cx| editor.set_text("", window, cx));
1151 self.region_editor
1152 .update(cx, |editor, cx| editor.set_text("", window, cx));
1153
1154 let state = self.state.clone();
1155 cx.spawn(async move |_, cx| {
1156 state
1157 .update(cx, |state, cx| state.reset_credentials(cx))?
1158 .await
1159 })
1160 .detach_and_log_err(cx);
1161 }
1162
1163 fn make_text_style(&self, cx: &Context<Self>) -> TextStyle {
1164 let settings = ThemeSettings::get_global(cx);
1165 TextStyle {
1166 color: cx.theme().colors().text,
1167 font_family: settings.ui_font.family.clone(),
1168 font_features: settings.ui_font.features.clone(),
1169 font_fallbacks: settings.ui_font.fallbacks.clone(),
1170 font_size: rems(0.875).into(),
1171 font_weight: settings.ui_font.weight,
1172 font_style: FontStyle::Normal,
1173 line_height: relative(1.3),
1174 background_color: None,
1175 underline: None,
1176 strikethrough: None,
1177 white_space: WhiteSpace::Normal,
1178 text_overflow: None,
1179 text_align: Default::default(),
1180 line_clamp: None,
1181 }
1182 }
1183
1184 fn make_input_styles(&self, cx: &Context<Self>) -> Div {
1185 let bg_color = cx.theme().colors().editor_background;
1186 let border_color = cx.theme().colors().border;
1187
1188 h_flex()
1189 .w_full()
1190 .px_2()
1191 .py_1()
1192 .bg(bg_color)
1193 .border_1()
1194 .border_color(border_color)
1195 .rounded_sm()
1196 }
1197
1198 fn should_render_editor(&self, cx: &Context<Self>) -> bool {
1199 self.state.read(cx).is_authenticated()
1200 }
1201}
1202
1203impl Render for ConfigurationView {
1204 fn render(&mut self, _window: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1205 let env_var_set = self.state.read(cx).credentials_from_env;
1206 let bedrock_settings = self.state.read(cx).settings.as_ref();
1207 let bedrock_method = bedrock_settings
1208 .as_ref()
1209 .and_then(|s| s.authentication_method.clone());
1210
1211 if self.load_credentials_task.is_some() {
1212 return div().child(Label::new("Loading credentials...")).into_any();
1213 }
1214
1215 if self.should_render_editor(cx) {
1216 return h_flex()
1217 .mt_1()
1218 .p_1()
1219 .justify_between()
1220 .rounded_md()
1221 .border_1()
1222 .border_color(cx.theme().colors().border)
1223 .bg(cx.theme().colors().background)
1224 .child(
1225 h_flex()
1226 .gap_1()
1227 .child(Icon::new(IconName::Check).color(Color::Success))
1228 .child(Label::new(if env_var_set {
1229 format!("Access Key ID is set in {ZED_BEDROCK_ACCESS_KEY_ID_VAR}, Secret Key is set in {ZED_BEDROCK_SECRET_ACCESS_KEY_VAR}, Region is set in {ZED_BEDROCK_REGION_VAR} environment variables.")
1230 } else {
1231 match bedrock_method {
1232 Some(BedrockAuthMethod::Automatic) => "You are using automatic credentials".into(),
1233 Some(BedrockAuthMethod::NamedProfile) => {
1234 "You are using named profile".into()
1235 },
1236 Some(BedrockAuthMethod::SingleSignOn) => "You are using a single sign on profile".into(),
1237 None => "You are using static credentials".into(),
1238 }
1239 })),
1240 )
1241 .child(
1242 Button::new("reset-key", "Reset Key")
1243 .icon(Some(IconName::Trash))
1244 .icon_size(IconSize::Small)
1245 .icon_position(IconPosition::Start)
1246 .disabled(env_var_set || bedrock_method.is_some())
1247 .when(env_var_set, |this| {
1248 this.tooltip(Tooltip::text(format!("To reset your credentials, unset the {ZED_BEDROCK_ACCESS_KEY_ID_VAR}, {ZED_BEDROCK_SECRET_ACCESS_KEY_VAR}, and {ZED_BEDROCK_REGION_VAR} environment variables.")))
1249 })
1250 .when(bedrock_method.is_some(), |this| {
1251 this.tooltip(Tooltip::text("You cannot reset credentials as they're being derived, check Zed settings to understand how"))
1252 })
1253 .on_click(cx.listener(|this, _, window, cx| this.reset_credentials(window, cx))),
1254 )
1255 .into_any();
1256 }
1257
1258 v_flex()
1259 .size_full()
1260 .on_action(cx.listener(ConfigurationView::save_credentials))
1261 .child(Label::new("To use Zed's assistant with Bedrock, you can set a custom authentication strategy through the settings.json, or use static credentials."))
1262 .child(Label::new("But, to access models on AWS, you need to:").mt_1())
1263 .child(
1264 List::new()
1265 .child(
1266 InstructionListItem::new(
1267 "Grant permissions to the strategy you'll use according to the:",
1268 Some("Prerequisites"),
1269 Some("https://docs.aws.amazon.com/bedrock/latest/userguide/inference-prereq.html"),
1270 )
1271 )
1272 .child(
1273 InstructionListItem::new(
1274 "Select the models you would like access to:",
1275 Some("Bedrock Model Catalog"),
1276 Some("https://us-east-1.console.aws.amazon.com/bedrock/home?region=us-east-1#/modelaccess"),
1277 )
1278 )
1279 )
1280 .child(self.render_static_credentials_ui(cx))
1281 .child(self.render_common_fields(cx))
1282 .child(
1283 Label::new(
1284 format!("You can also assign the {ZED_BEDROCK_ACCESS_KEY_ID_VAR}, {ZED_BEDROCK_SECRET_ACCESS_KEY_VAR} AND {ZED_BEDROCK_REGION_VAR} environment variables and restart Zed."),
1285 )
1286 .size(LabelSize::Small)
1287 .color(Color::Muted)
1288 .my_1(),
1289 )
1290 .child(
1291 Label::new(
1292 format!("Optionally, if your environment uses AWS CLI profiles, you can set {ZED_AWS_PROFILE_VAR}; if it requires a custom endpoint, you can set {ZED_AWS_ENDPOINT_VAR}; and if it requires a Session Token, you can set {ZED_BEDROCK_SESSION_TOKEN_VAR}."),
1293 )
1294 .size(LabelSize::Small)
1295 .color(Color::Muted),
1296 )
1297 .into_any()
1298 }
1299}
1300
1301impl ConfigurationView {
1302 fn render_access_key_id_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1303 let text_style = self.make_text_style(cx);
1304
1305 EditorElement::new(
1306 &self.access_key_id_editor,
1307 EditorStyle {
1308 background: cx.theme().colors().editor_background,
1309 local_player: cx.theme().players().local(),
1310 text: text_style,
1311 ..Default::default()
1312 },
1313 )
1314 }
1315
1316 fn render_secret_key_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1317 let text_style = self.make_text_style(cx);
1318
1319 EditorElement::new(
1320 &self.secret_access_key_editor,
1321 EditorStyle {
1322 background: cx.theme().colors().editor_background,
1323 local_player: cx.theme().players().local(),
1324 text: text_style,
1325 ..Default::default()
1326 },
1327 )
1328 }
1329
1330 fn render_session_token_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1331 let text_style = self.make_text_style(cx);
1332
1333 EditorElement::new(
1334 &self.session_token_editor,
1335 EditorStyle {
1336 background: cx.theme().colors().editor_background,
1337 local_player: cx.theme().players().local(),
1338 text: text_style,
1339 ..Default::default()
1340 },
1341 )
1342 }
1343
1344 fn render_region_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1345 let text_style = self.make_text_style(cx);
1346
1347 EditorElement::new(
1348 &self.region_editor,
1349 EditorStyle {
1350 background: cx.theme().colors().editor_background,
1351 local_player: cx.theme().players().local(),
1352 text: text_style,
1353 ..Default::default()
1354 },
1355 )
1356 }
1357
1358 fn render_static_credentials_ui(&self, cx: &mut Context<Self>) -> AnyElement {
1359 v_flex()
1360 .my_2()
1361 .gap_1p5()
1362 .child(
1363 Label::new("Static Keys")
1364 .size(LabelSize::Default)
1365 .weight(FontWeight::BOLD),
1366 )
1367 .child(
1368 Label::new(
1369 "This method uses your AWS access key ID and secret access key directly.",
1370 )
1371 )
1372 .child(
1373 List::new()
1374 .child(InstructionListItem::new(
1375 "Create an IAM user in the AWS console with programmatic access",
1376 Some("IAM Console"),
1377 Some("https://us-east-1.console.aws.amazon.com/iam/home?region=us-east-1#/users"),
1378 ))
1379 .child(InstructionListItem::new(
1380 "Attach the necessary Bedrock permissions to this ",
1381 Some("user"),
1382 Some("https://docs.aws.amazon.com/bedrock/latest/userguide/inference-prereq.html"),
1383 ))
1384 .child(InstructionListItem::text_only(
1385 "Copy the access key ID and secret access key when provided",
1386 ))
1387 .child(InstructionListItem::text_only(
1388 "Enter these credentials below",
1389 )),
1390 )
1391 .child(
1392 v_flex()
1393 .gap_0p5()
1394 .child(Label::new("Access Key ID").size(LabelSize::Small))
1395 .child(
1396 self.make_input_styles(cx)
1397 .child(self.render_access_key_id_editor(cx)),
1398 ),
1399 )
1400 .child(
1401 v_flex()
1402 .gap_0p5()
1403 .child(Label::new("Secret Access Key").size(LabelSize::Small))
1404 .child(self.make_input_styles(cx).child(self.render_secret_key_editor(cx))),
1405 )
1406 .child(
1407 v_flex()
1408 .gap_0p5()
1409 .child(Label::new("Session Token (Optional)").size(LabelSize::Small))
1410 .child(
1411 self.make_input_styles(cx)
1412 .child(self.render_session_token_editor(cx)),
1413 ),
1414 )
1415 .into_any_element()
1416 }
1417
1418 fn render_common_fields(&self, cx: &mut Context<Self>) -> AnyElement {
1419 v_flex()
1420 .gap_0p5()
1421 .child(Label::new("Region").size(LabelSize::Small))
1422 .child(
1423 self.make_input_styles(cx)
1424 .child(self.render_region_editor(cx)),
1425 )
1426 .into_any_element()
1427 }
1428}