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, WrappedTextContent,
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: usize,
92 pub cache_configuration: Option<LanguageModelCacheConfiguration>,
93 pub max_output_tokens: Option<u32>,
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) -> usize {
507 self.model.max_token_count()
508 }
509
510 fn max_output_tokens(&self) -> Option<u32> {
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<usize>> {
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 >,
531 > {
532 let Ok(region) = cx.read_entity(&self.state, |state, _cx| {
533 // Get region - from credentials or directly from settings
534 let region = state
535 .credentials
536 .as_ref()
537 .map(|s| s.region.clone())
538 .unwrap_or(String::from("us-east-1"));
539
540 region
541 }) else {
542 return async move {
543 anyhow::bail!("App State Dropped");
544 }
545 .boxed();
546 };
547
548 let model_id = match self.model.cross_region_inference_id(®ion) {
549 Ok(s) => s,
550 Err(e) => {
551 return async move { Err(e) }.boxed();
552 }
553 };
554
555 let request = match into_bedrock(
556 request,
557 model_id,
558 self.model.default_temperature(),
559 self.model.max_output_tokens(),
560 self.model.mode(),
561 ) {
562 Ok(request) => request,
563 Err(err) => return futures::future::ready(Err(err)).boxed(),
564 };
565
566 let owned_handle = self.handler.clone();
567
568 let request = self.stream_completion(request, cx);
569 let future = self.request_limiter.stream(async move {
570 let response = request.map_err(|err| anyhow!(err))?.await;
571 Ok(map_to_language_model_completion_events(
572 response,
573 owned_handle,
574 ))
575 });
576 async move { Ok(future.await?.boxed()) }.boxed()
577 }
578
579 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
580 None
581 }
582}
583
584pub fn into_bedrock(
585 request: LanguageModelRequest,
586 model: String,
587 default_temperature: f32,
588 max_output_tokens: u32,
589 mode: BedrockModelMode,
590) -> Result<bedrock::Request> {
591 let mut new_messages: Vec<BedrockMessage> = Vec::new();
592 let mut system_message = String::new();
593
594 for message in request.messages {
595 if message.contents_empty() {
596 continue;
597 }
598
599 match message.role {
600 Role::User | Role::Assistant => {
601 let bedrock_message_content: Vec<BedrockInnerContent> = message
602 .content
603 .into_iter()
604 .filter_map(|content| match content {
605 MessageContent::Text(text) => {
606 if !text.is_empty() {
607 Some(BedrockInnerContent::Text(text))
608 } else {
609 None
610 }
611 }
612 MessageContent::Thinking { text, signature } => {
613 let thinking = BedrockThinkingTextBlock::builder()
614 .text(text)
615 .set_signature(signature)
616 .build()
617 .context("failed to build reasoning block")
618 .log_err()?;
619
620 Some(BedrockInnerContent::ReasoningContent(
621 BedrockThinkingBlock::ReasoningText(thinking),
622 ))
623 }
624 MessageContent::RedactedThinking(blob) => {
625 let redacted =
626 BedrockThinkingBlock::RedactedContent(BedrockBlob::new(blob));
627
628 Some(BedrockInnerContent::ReasoningContent(redacted))
629 }
630 MessageContent::ToolUse(tool_use) => BedrockToolUseBlock::builder()
631 .name(tool_use.name.to_string())
632 .tool_use_id(tool_use.id.to_string())
633 .input(value_to_aws_document(&tool_use.input))
634 .build()
635 .context("failed to build Bedrock tool use block")
636 .log_err()
637 .map(BedrockInnerContent::ToolUse),
638 MessageContent::ToolResult(tool_result) => {
639 BedrockToolResultBlock::builder()
640 .tool_use_id(tool_result.tool_use_id.to_string())
641 .content(match tool_result.content {
642 LanguageModelToolResultContent::Text(text)
643 | LanguageModelToolResultContent::WrappedText(WrappedTextContent { text, .. }) => {
644 BedrockToolResultContentBlock::Text(text.to_string())
645 }
646 LanguageModelToolResultContent::Image(_) => {
647 BedrockToolResultContentBlock::Text(
648 // TODO: Bedrock image support
649 "[Tool responded with an image, but Zed doesn't support these in Bedrock models yet]".to_string()
650 )
651 }
652 })
653 .status({
654 if tool_result.is_error {
655 BedrockToolResultStatus::Error
656 } else {
657 BedrockToolResultStatus::Success
658 }
659 })
660 .build()
661 .context("failed to build Bedrock tool result block")
662 .log_err()
663 .map(BedrockInnerContent::ToolResult)
664 }
665 _ => None,
666 })
667 .collect();
668 let bedrock_role = match message.role {
669 Role::User => bedrock::BedrockRole::User,
670 Role::Assistant => bedrock::BedrockRole::Assistant,
671 Role::System => unreachable!("System role should never occur here"),
672 };
673 if let Some(last_message) = new_messages.last_mut() {
674 if last_message.role == bedrock_role {
675 last_message.content.extend(bedrock_message_content);
676 continue;
677 }
678 }
679 new_messages.push(
680 BedrockMessage::builder()
681 .role(bedrock_role)
682 .set_content(Some(bedrock_message_content))
683 .build()
684 .context("failed to build Bedrock message")?,
685 );
686 }
687 Role::System => {
688 if !system_message.is_empty() {
689 system_message.push_str("\n\n");
690 }
691 system_message.push_str(&message.string_contents());
692 }
693 }
694 }
695
696 let tool_spec: Vec<BedrockTool> = request
697 .tools
698 .iter()
699 .filter_map(|tool| {
700 Some(BedrockTool::ToolSpec(
701 BedrockToolSpec::builder()
702 .name(tool.name.clone())
703 .description(tool.description.clone())
704 .input_schema(BedrockToolInputSchema::Json(value_to_aws_document(
705 &tool.input_schema,
706 )))
707 .build()
708 .log_err()?,
709 ))
710 })
711 .collect();
712
713 let tool_choice = match request.tool_choice {
714 Some(LanguageModelToolChoice::Auto) | None => {
715 BedrockToolChoice::Auto(BedrockAutoToolChoice::builder().build())
716 }
717 Some(LanguageModelToolChoice::Any) => {
718 BedrockToolChoice::Any(BedrockAnyToolChoice::builder().build())
719 }
720 Some(LanguageModelToolChoice::None) => {
721 anyhow::bail!("LanguageModelToolChoice::None is not supported");
722 }
723 };
724 let tool_config: BedrockToolConfig = BedrockToolConfig::builder()
725 .set_tools(Some(tool_spec))
726 .tool_choice(tool_choice)
727 .build()?;
728
729 Ok(bedrock::Request {
730 model,
731 messages: new_messages,
732 max_tokens: max_output_tokens,
733 system: Some(system_message),
734 tools: Some(tool_config),
735 thinking: if let BedrockModelMode::Thinking { budget_tokens } = mode {
736 Some(bedrock::Thinking::Enabled { budget_tokens })
737 } else {
738 None
739 },
740 metadata: None,
741 stop_sequences: Vec::new(),
742 temperature: request.temperature.or(Some(default_temperature)),
743 top_k: None,
744 top_p: None,
745 })
746}
747
748// TODO: just call the ConverseOutput.usage() method:
749// https://docs.rs/aws-sdk-bedrockruntime/latest/aws_sdk_bedrockruntime/operation/converse/struct.ConverseOutput.html#method.output
750pub fn get_bedrock_tokens(
751 request: LanguageModelRequest,
752 cx: &App,
753) -> BoxFuture<'static, Result<usize>> {
754 cx.background_executor()
755 .spawn(async move {
756 let messages = request.messages;
757 let mut tokens_from_images = 0;
758 let mut string_messages = Vec::with_capacity(messages.len());
759
760 for message in messages {
761 use language_model::MessageContent;
762
763 let mut string_contents = String::new();
764
765 for content in message.content {
766 match content {
767 MessageContent::Text(text) | MessageContent::Thinking { text, .. } => {
768 string_contents.push_str(&text);
769 }
770 MessageContent::RedactedThinking(_) => {}
771 MessageContent::Image(image) => {
772 tokens_from_images += image.estimate_tokens();
773 }
774 MessageContent::ToolUse(_tool_use) => {
775 // TODO: Estimate token usage from tool uses.
776 }
777 MessageContent::ToolResult(tool_result) => match tool_result.content {
778 LanguageModelToolResultContent::Text(text)
779 | LanguageModelToolResultContent::WrappedText(WrappedTextContent {
780 text,
781 ..
782 }) => {
783 string_contents.push_str(&text);
784 }
785 LanguageModelToolResultContent::Image(image) => {
786 tokens_from_images += image.estimate_tokens();
787 }
788 },
789 }
790 }
791
792 if !string_contents.is_empty() {
793 string_messages.push(tiktoken_rs::ChatCompletionRequestMessage {
794 role: match message.role {
795 Role::User => "user".into(),
796 Role::Assistant => "assistant".into(),
797 Role::System => "system".into(),
798 },
799 content: Some(string_contents),
800 name: None,
801 function_call: None,
802 });
803 }
804 }
805
806 // Tiktoken doesn't yet support these models, so we manually use the
807 // same tokenizer as GPT-4.
808 tiktoken_rs::num_tokens_from_messages("gpt-4", &string_messages)
809 .map(|tokens| tokens + tokens_from_images)
810 })
811 .boxed()
812}
813
814pub fn map_to_language_model_completion_events(
815 events: Pin<Box<dyn Send + Stream<Item = Result<BedrockStreamingResponse, BedrockError>>>>,
816 handle: Handle,
817) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
818 struct RawToolUse {
819 id: String,
820 name: String,
821 input_json: String,
822 }
823
824 struct State {
825 events: Pin<Box<dyn Send + Stream<Item = Result<BedrockStreamingResponse, BedrockError>>>>,
826 tool_uses_by_index: HashMap<i32, RawToolUse>,
827 }
828
829 futures::stream::unfold(
830 State {
831 events,
832 tool_uses_by_index: HashMap::default(),
833 },
834 move |mut state: State| {
835 let inner_handle = handle.clone();
836 async move {
837 inner_handle
838 .spawn(async {
839 while let Some(event) = state.events.next().await {
840 match event {
841 Ok(event) => match event {
842 ConverseStreamOutput::ContentBlockDelta(cb_delta) => {
843 match cb_delta.delta {
844 Some(ContentBlockDelta::Text(text_out)) => {
845 let completion_event =
846 LanguageModelCompletionEvent::Text(text_out);
847 return Some((Some(Ok(completion_event)), state));
848 }
849
850 Some(ContentBlockDelta::ToolUse(text_out)) => {
851 if let Some(tool_use) = state
852 .tool_uses_by_index
853 .get_mut(&cb_delta.content_block_index)
854 {
855 tool_use.input_json.push_str(text_out.input());
856 }
857 }
858
859 Some(ContentBlockDelta::ReasoningContent(thinking)) => {
860 match thinking {
861 ReasoningContentBlockDelta::RedactedContent(
862 redacted,
863 ) => {
864 let thinking_event =
865 LanguageModelCompletionEvent::Thinking {
866 text: String::from_utf8(
867 redacted.into_inner(),
868 )
869 .unwrap_or("REDACTED".to_string()),
870 signature: None,
871 };
872
873 return Some((
874 Some(Ok(thinking_event)),
875 state,
876 ));
877 }
878 ReasoningContentBlockDelta::Signature(
879 signature,
880 ) => {
881 return Some((
882 Some(Ok(LanguageModelCompletionEvent::Thinking {
883 text: "".to_string(),
884 signature: Some(signature)
885 })),
886 state,
887 ));
888 }
889 ReasoningContentBlockDelta::Text(thoughts) => {
890 let thinking_event =
891 LanguageModelCompletionEvent::Thinking {
892 text: thoughts.to_string(),
893 signature: None
894 };
895
896 return Some((
897 Some(Ok(thinking_event)),
898 state,
899 ));
900 }
901 _ => {}
902 }
903 }
904 _ => {}
905 }
906 }
907 ConverseStreamOutput::ContentBlockStart(cb_start) => {
908 if let Some(ContentBlockStart::ToolUse(text_out)) =
909 cb_start.start
910 {
911 let tool_use = RawToolUse {
912 id: text_out.tool_use_id,
913 name: text_out.name,
914 input_json: String::new(),
915 };
916
917 state
918 .tool_uses_by_index
919 .insert(cb_start.content_block_index, tool_use);
920 }
921 }
922 ConverseStreamOutput::ContentBlockStop(cb_stop) => {
923 if let Some(tool_use) = state
924 .tool_uses_by_index
925 .remove(&cb_stop.content_block_index)
926 {
927 let tool_use_event = LanguageModelToolUse {
928 id: tool_use.id.into(),
929 name: tool_use.name.into(),
930 is_input_complete: true,
931 raw_input: tool_use.input_json.clone(),
932 input: if tool_use.input_json.is_empty() {
933 Value::Null
934 } else {
935 serde_json::Value::from_str(
936 &tool_use.input_json,
937 )
938 .map_err(|err| anyhow!(err))
939 .unwrap()
940 },
941 };
942
943 return Some((
944 Some(Ok(LanguageModelCompletionEvent::ToolUse(
945 tool_use_event,
946 ))),
947 state,
948 ));
949 }
950 }
951
952 ConverseStreamOutput::Metadata(cb_meta) => {
953 if let Some(metadata) = cb_meta.usage {
954 let completion_event =
955 LanguageModelCompletionEvent::UsageUpdate(
956 TokenUsage {
957 input_tokens: metadata.input_tokens as u32,
958 output_tokens: metadata.output_tokens
959 as u32,
960 cache_creation_input_tokens: default(),
961 cache_read_input_tokens: default(),
962 },
963 );
964 return Some((Some(Ok(completion_event)), state));
965 }
966 }
967 ConverseStreamOutput::MessageStop(message_stop) => {
968 let reason = match message_stop.stop_reason {
969 StopReason::ContentFiltered => {
970 LanguageModelCompletionEvent::Stop(
971 language_model::StopReason::EndTurn,
972 )
973 }
974 StopReason::EndTurn => {
975 LanguageModelCompletionEvent::Stop(
976 language_model::StopReason::EndTurn,
977 )
978 }
979 StopReason::GuardrailIntervened => {
980 LanguageModelCompletionEvent::Stop(
981 language_model::StopReason::EndTurn,
982 )
983 }
984 StopReason::MaxTokens => {
985 LanguageModelCompletionEvent::Stop(
986 language_model::StopReason::EndTurn,
987 )
988 }
989 StopReason::StopSequence => {
990 LanguageModelCompletionEvent::Stop(
991 language_model::StopReason::EndTurn,
992 )
993 }
994 StopReason::ToolUse => {
995 LanguageModelCompletionEvent::Stop(
996 language_model::StopReason::ToolUse,
997 )
998 }
999 _ => LanguageModelCompletionEvent::Stop(
1000 language_model::StopReason::EndTurn,
1001 ),
1002 };
1003 return Some((Some(Ok(reason)), state));
1004 }
1005 _ => {}
1006 },
1007
1008 Err(err) => return Some((Some(Err(anyhow!(err).into())), state)),
1009 }
1010 }
1011 None
1012 })
1013 .await
1014 .log_err()
1015 .flatten()
1016 }
1017 },
1018 )
1019 .filter_map(|event| async move { event })
1020}
1021
1022struct ConfigurationView {
1023 access_key_id_editor: Entity<Editor>,
1024 secret_access_key_editor: Entity<Editor>,
1025 session_token_editor: Entity<Editor>,
1026 region_editor: Entity<Editor>,
1027 state: gpui::Entity<State>,
1028 load_credentials_task: Option<Task<()>>,
1029}
1030
1031impl ConfigurationView {
1032 const PLACEHOLDER_ACCESS_KEY_ID_TEXT: &'static str = "XXXXXXXXXXXXXXXX";
1033 const PLACEHOLDER_SECRET_ACCESS_KEY_TEXT: &'static str =
1034 "XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX";
1035 const PLACEHOLDER_SESSION_TOKEN_TEXT: &'static str = "XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX";
1036 const PLACEHOLDER_REGION: &'static str = "us-east-1";
1037
1038 fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
1039 cx.observe(&state, |_, _, cx| {
1040 cx.notify();
1041 })
1042 .detach();
1043
1044 let load_credentials_task = Some(cx.spawn({
1045 let state = state.clone();
1046 async move |this, cx| {
1047 if let Some(task) = state
1048 .update(cx, |state, cx| state.authenticate(cx))
1049 .log_err()
1050 {
1051 // We don't log an error, because "not signed in" is also an error.
1052 let _ = task.await;
1053 }
1054 this.update(cx, |this, cx| {
1055 this.load_credentials_task = None;
1056 cx.notify();
1057 })
1058 .log_err();
1059 }
1060 }));
1061
1062 Self {
1063 access_key_id_editor: cx.new(|cx| {
1064 let mut editor = Editor::single_line(window, cx);
1065 editor.set_placeholder_text(Self::PLACEHOLDER_ACCESS_KEY_ID_TEXT, cx);
1066 editor
1067 }),
1068 secret_access_key_editor: cx.new(|cx| {
1069 let mut editor = Editor::single_line(window, cx);
1070 editor.set_placeholder_text(Self::PLACEHOLDER_SECRET_ACCESS_KEY_TEXT, cx);
1071 editor
1072 }),
1073 session_token_editor: cx.new(|cx| {
1074 let mut editor = Editor::single_line(window, cx);
1075 editor.set_placeholder_text(Self::PLACEHOLDER_SESSION_TOKEN_TEXT, cx);
1076 editor
1077 }),
1078 region_editor: cx.new(|cx| {
1079 let mut editor = Editor::single_line(window, cx);
1080 editor.set_placeholder_text(Self::PLACEHOLDER_REGION, cx);
1081 editor
1082 }),
1083 state,
1084 load_credentials_task,
1085 }
1086 }
1087
1088 fn save_credentials(
1089 &mut self,
1090 _: &menu::Confirm,
1091 _window: &mut Window,
1092 cx: &mut Context<Self>,
1093 ) {
1094 let access_key_id = self
1095 .access_key_id_editor
1096 .read(cx)
1097 .text(cx)
1098 .to_string()
1099 .trim()
1100 .to_string();
1101 let secret_access_key = self
1102 .secret_access_key_editor
1103 .read(cx)
1104 .text(cx)
1105 .to_string()
1106 .trim()
1107 .to_string();
1108 let session_token = self
1109 .session_token_editor
1110 .read(cx)
1111 .text(cx)
1112 .to_string()
1113 .trim()
1114 .to_string();
1115 let session_token = if session_token.is_empty() {
1116 None
1117 } else {
1118 Some(session_token)
1119 };
1120 let region = self
1121 .region_editor
1122 .read(cx)
1123 .text(cx)
1124 .to_string()
1125 .trim()
1126 .to_string();
1127 let region = if region.is_empty() {
1128 "us-east-1".to_string()
1129 } else {
1130 region
1131 };
1132
1133 let state = self.state.clone();
1134 cx.spawn(async move |_, cx| {
1135 state
1136 .update(cx, |state, cx| {
1137 let credentials: BedrockCredentials = BedrockCredentials {
1138 region: region.clone(),
1139 access_key_id: access_key_id.clone(),
1140 secret_access_key: secret_access_key.clone(),
1141 session_token: session_token.clone(),
1142 };
1143
1144 state.set_credentials(credentials, cx)
1145 })?
1146 .await
1147 })
1148 .detach_and_log_err(cx);
1149 }
1150
1151 fn reset_credentials(&mut self, window: &mut Window, cx: &mut Context<Self>) {
1152 self.access_key_id_editor
1153 .update(cx, |editor, cx| editor.set_text("", window, cx));
1154 self.secret_access_key_editor
1155 .update(cx, |editor, cx| editor.set_text("", window, cx));
1156 self.session_token_editor
1157 .update(cx, |editor, cx| editor.set_text("", window, cx));
1158 self.region_editor
1159 .update(cx, |editor, cx| editor.set_text("", window, cx));
1160
1161 let state = self.state.clone();
1162 cx.spawn(async move |_, cx| {
1163 state
1164 .update(cx, |state, cx| state.reset_credentials(cx))?
1165 .await
1166 })
1167 .detach_and_log_err(cx);
1168 }
1169
1170 fn make_text_style(&self, cx: &Context<Self>) -> TextStyle {
1171 let settings = ThemeSettings::get_global(cx);
1172 TextStyle {
1173 color: cx.theme().colors().text,
1174 font_family: settings.ui_font.family.clone(),
1175 font_features: settings.ui_font.features.clone(),
1176 font_fallbacks: settings.ui_font.fallbacks.clone(),
1177 font_size: rems(0.875).into(),
1178 font_weight: settings.ui_font.weight,
1179 font_style: FontStyle::Normal,
1180 line_height: relative(1.3),
1181 background_color: None,
1182 underline: None,
1183 strikethrough: None,
1184 white_space: WhiteSpace::Normal,
1185 text_overflow: None,
1186 text_align: Default::default(),
1187 line_clamp: None,
1188 }
1189 }
1190
1191 fn make_input_styles(&self, cx: &Context<Self>) -> Div {
1192 let bg_color = cx.theme().colors().editor_background;
1193 let border_color = cx.theme().colors().border;
1194
1195 h_flex()
1196 .w_full()
1197 .px_2()
1198 .py_1()
1199 .bg(bg_color)
1200 .border_1()
1201 .border_color(border_color)
1202 .rounded_sm()
1203 }
1204
1205 fn should_render_editor(&self, cx: &Context<Self>) -> bool {
1206 self.state.read(cx).is_authenticated()
1207 }
1208}
1209
1210impl Render for ConfigurationView {
1211 fn render(&mut self, _window: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
1212 let env_var_set = self.state.read(cx).credentials_from_env;
1213 let bedrock_settings = self.state.read(cx).settings.as_ref();
1214 let bedrock_method = bedrock_settings
1215 .as_ref()
1216 .and_then(|s| s.authentication_method.clone());
1217
1218 if self.load_credentials_task.is_some() {
1219 return div().child(Label::new("Loading credentials...")).into_any();
1220 }
1221
1222 if self.should_render_editor(cx) {
1223 return h_flex()
1224 .mt_1()
1225 .p_1()
1226 .justify_between()
1227 .rounded_md()
1228 .border_1()
1229 .border_color(cx.theme().colors().border)
1230 .bg(cx.theme().colors().background)
1231 .child(
1232 h_flex()
1233 .gap_1()
1234 .child(Icon::new(IconName::Check).color(Color::Success))
1235 .child(Label::new(if env_var_set {
1236 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.")
1237 } else {
1238 match bedrock_method {
1239 Some(BedrockAuthMethod::Automatic) => "You are using automatic credentials".into(),
1240 Some(BedrockAuthMethod::NamedProfile) => {
1241 "You are using named profile".into()
1242 },
1243 Some(BedrockAuthMethod::SingleSignOn) => "You are using a single sign on profile".into(),
1244 None => "You are using static credentials".into(),
1245 }
1246 })),
1247 )
1248 .child(
1249 Button::new("reset-key", "Reset Key")
1250 .icon(Some(IconName::Trash))
1251 .icon_size(IconSize::Small)
1252 .icon_position(IconPosition::Start)
1253 .disabled(env_var_set || bedrock_method.is_some())
1254 .when(env_var_set, |this| {
1255 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.")))
1256 })
1257 .when(bedrock_method.is_some(), |this| {
1258 this.tooltip(Tooltip::text("You cannot reset credentials as they're being derived, check Zed settings to understand how"))
1259 })
1260 .on_click(cx.listener(|this, _, window, cx| this.reset_credentials(window, cx))),
1261 )
1262 .into_any();
1263 }
1264
1265 v_flex()
1266 .size_full()
1267 .on_action(cx.listener(ConfigurationView::save_credentials))
1268 .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."))
1269 .child(Label::new("But, to access models on AWS, you need to:").mt_1())
1270 .child(
1271 List::new()
1272 .child(
1273 InstructionListItem::new(
1274 "Grant permissions to the strategy you'll use according to the:",
1275 Some("Prerequisites"),
1276 Some("https://docs.aws.amazon.com/bedrock/latest/userguide/inference-prereq.html"),
1277 )
1278 )
1279 .child(
1280 InstructionListItem::new(
1281 "Select the models you would like access to:",
1282 Some("Bedrock Model Catalog"),
1283 Some("https://us-east-1.console.aws.amazon.com/bedrock/home?region=us-east-1#/modelaccess"),
1284 )
1285 )
1286 )
1287 .child(self.render_static_credentials_ui(cx))
1288 .child(self.render_common_fields(cx))
1289 .child(
1290 Label::new(
1291 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."),
1292 )
1293 .size(LabelSize::Small)
1294 .color(Color::Muted)
1295 .my_1(),
1296 )
1297 .child(
1298 Label::new(
1299 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}."),
1300 )
1301 .size(LabelSize::Small)
1302 .color(Color::Muted),
1303 )
1304 .into_any()
1305 }
1306}
1307
1308impl ConfigurationView {
1309 fn render_access_key_id_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1310 let text_style = self.make_text_style(cx);
1311
1312 EditorElement::new(
1313 &self.access_key_id_editor,
1314 EditorStyle {
1315 background: cx.theme().colors().editor_background,
1316 local_player: cx.theme().players().local(),
1317 text: text_style,
1318 ..Default::default()
1319 },
1320 )
1321 }
1322
1323 fn render_secret_key_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1324 let text_style = self.make_text_style(cx);
1325
1326 EditorElement::new(
1327 &self.secret_access_key_editor,
1328 EditorStyle {
1329 background: cx.theme().colors().editor_background,
1330 local_player: cx.theme().players().local(),
1331 text: text_style,
1332 ..Default::default()
1333 },
1334 )
1335 }
1336
1337 fn render_session_token_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1338 let text_style = self.make_text_style(cx);
1339
1340 EditorElement::new(
1341 &self.session_token_editor,
1342 EditorStyle {
1343 background: cx.theme().colors().editor_background,
1344 local_player: cx.theme().players().local(),
1345 text: text_style,
1346 ..Default::default()
1347 },
1348 )
1349 }
1350
1351 fn render_region_editor(&self, cx: &mut Context<Self>) -> impl IntoElement {
1352 let text_style = self.make_text_style(cx);
1353
1354 EditorElement::new(
1355 &self.region_editor,
1356 EditorStyle {
1357 background: cx.theme().colors().editor_background,
1358 local_player: cx.theme().players().local(),
1359 text: text_style,
1360 ..Default::default()
1361 },
1362 )
1363 }
1364
1365 fn render_static_credentials_ui(&self, cx: &mut Context<Self>) -> AnyElement {
1366 v_flex()
1367 .my_2()
1368 .gap_1p5()
1369 .child(
1370 Label::new("Static Keys")
1371 .size(LabelSize::Default)
1372 .weight(FontWeight::BOLD),
1373 )
1374 .child(
1375 Label::new(
1376 "This method uses your AWS access key ID and secret access key directly.",
1377 )
1378 )
1379 .child(
1380 List::new()
1381 .child(InstructionListItem::new(
1382 "Create an IAM user in the AWS console with programmatic access",
1383 Some("IAM Console"),
1384 Some("https://us-east-1.console.aws.amazon.com/iam/home?region=us-east-1#/users"),
1385 ))
1386 .child(InstructionListItem::new(
1387 "Attach the necessary Bedrock permissions to this ",
1388 Some("user"),
1389 Some("https://docs.aws.amazon.com/bedrock/latest/userguide/inference-prereq.html"),
1390 ))
1391 .child(InstructionListItem::text_only(
1392 "Copy the access key ID and secret access key when provided",
1393 ))
1394 .child(InstructionListItem::text_only(
1395 "Enter these credentials below",
1396 )),
1397 )
1398 .child(
1399 v_flex()
1400 .gap_0p5()
1401 .child(Label::new("Access Key ID").size(LabelSize::Small))
1402 .child(
1403 self.make_input_styles(cx)
1404 .child(self.render_access_key_id_editor(cx)),
1405 ),
1406 )
1407 .child(
1408 v_flex()
1409 .gap_0p5()
1410 .child(Label::new("Secret Access Key").size(LabelSize::Small))
1411 .child(self.make_input_styles(cx).child(self.render_secret_key_editor(cx))),
1412 )
1413 .child(
1414 v_flex()
1415 .gap_0p5()
1416 .child(Label::new("Session Token (Optional)").size(LabelSize::Small))
1417 .child(
1418 self.make_input_styles(cx)
1419 .child(self.render_session_token_editor(cx)),
1420 ),
1421 )
1422 .into_any_element()
1423 }
1424
1425 fn render_common_fields(&self, cx: &mut Context<Self>) -> AnyElement {
1426 v_flex()
1427 .gap_0p5()
1428 .child(Label::new("Region").size(LabelSize::Small))
1429 .child(
1430 self.make_input_styles(cx)
1431 .child(self.render_region_editor(cx)),
1432 )
1433 .into_any_element()
1434 }
1435}