1use anyhow::{Result, anyhow};
2use fs::Fs;
3use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};
4use futures::{Stream, TryFutureExt, stream};
5use gpui::{AnyView, App, AsyncApp, Context, Task};
6use http_client::HttpClient;
7use language_model::{
8 AuthenticateError, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,
9 LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
10 LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,
11 LanguageModelRequestTool, LanguageModelToolChoice, LanguageModelToolUse,
12 LanguageModelToolUseId, MessageContent, RateLimiter, Role, StopReason, TokenUsage,
13};
14use menu;
15use ollama::{
16 ChatMessage, ChatOptions, ChatRequest, ChatResponseDelta, OLLAMA_API_URL, OllamaFunctionCall,
17 OllamaFunctionTool, OllamaToolCall, get_models, show_model, stream_chat_completion,
18};
19pub use settings::OllamaAvailableModel as AvailableModel;
20use settings::{Settings, SettingsStore, update_settings_file};
21use std::pin::Pin;
22use std::sync::LazyLock;
23use std::sync::atomic::{AtomicU64, Ordering};
24use std::{collections::HashMap, sync::Arc};
25use ui::{ButtonLike, ElevationIndex, List, Tooltip, prelude::*};
26use ui_input::SingleLineInput;
27use zed_env_vars::{EnvVar, env_var};
28
29use crate::AllLanguageModelSettings;
30use crate::api_key::ApiKeyState;
31use crate::ui::InstructionListItem;
32
33const OLLAMA_DOWNLOAD_URL: &str = "https://ollama.com/download";
34const OLLAMA_LIBRARY_URL: &str = "https://ollama.com/library";
35const OLLAMA_SITE: &str = "https://ollama.com/";
36
37const PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("ollama");
38const PROVIDER_NAME: LanguageModelProviderName = LanguageModelProviderName::new("Ollama");
39
40const API_KEY_ENV_VAR_NAME: &str = "OLLAMA_API_KEY";
41static API_KEY_ENV_VAR: LazyLock<EnvVar> = env_var!(API_KEY_ENV_VAR_NAME);
42
43#[derive(Default, Debug, Clone, PartialEq)]
44pub struct OllamaSettings {
45 pub api_url: String,
46 pub available_models: Vec<AvailableModel>,
47}
48
49pub struct OllamaLanguageModelProvider {
50 http_client: Arc<dyn HttpClient>,
51 state: gpui::Entity<State>,
52}
53
54pub struct State {
55 api_key_state: ApiKeyState,
56 http_client: Arc<dyn HttpClient>,
57 fetched_models: Vec<ollama::Model>,
58 fetch_model_task: Option<Task<Result<()>>>,
59}
60
61impl State {
62 fn is_authenticated(&self) -> bool {
63 !self.fetched_models.is_empty()
64 }
65
66 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
67 let api_url = OllamaLanguageModelProvider::api_url(cx);
68 let task = self
69 .api_key_state
70 .store(api_url, api_key, |this| &mut this.api_key_state, cx);
71
72 self.fetched_models.clear();
73 cx.spawn(async move |this, cx| {
74 let result = task.await;
75 this.update(cx, |this, cx| this.restart_fetch_models_task(cx))
76 .ok();
77 result
78 })
79 }
80
81 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
82 let api_url = OllamaLanguageModelProvider::api_url(cx);
83 let task = self.api_key_state.load_if_needed(
84 api_url,
85 &API_KEY_ENV_VAR,
86 |this| &mut this.api_key_state,
87 cx,
88 );
89
90 // Always try to fetch models - if no API key is needed (local Ollama), it will work
91 // If API key is needed and provided, it will work
92 // If API key is needed and not provided, it will fail gracefully
93 cx.spawn(async move |this, cx| {
94 let result = task.await;
95 this.update(cx, |this, cx| this.restart_fetch_models_task(cx))
96 .ok();
97 result
98 })
99 }
100
101 fn fetch_models(&mut self, cx: &mut Context<Self>) -> Task<Result<()>> {
102 let http_client = Arc::clone(&self.http_client);
103 let api_url = OllamaLanguageModelProvider::api_url(cx);
104 let api_key = self.api_key_state.key(&api_url);
105
106 // As a proxy for the server being "authenticated", we'll check if its up by fetching the models
107 cx.spawn(async move |this, cx| {
108 let models =
109 get_models(http_client.as_ref(), &api_url, api_key.as_deref(), None).await?;
110
111 let tasks = models
112 .into_iter()
113 // Since there is no metadata from the Ollama API
114 // indicating which models are embedding models,
115 // simply filter out models with "-embed" in their name
116 .filter(|model| !model.name.contains("-embed"))
117 .map(|model| {
118 let http_client = Arc::clone(&http_client);
119 let api_url = api_url.clone();
120 let api_key = api_key.clone();
121 async move {
122 let name = model.name.as_str();
123 let capabilities =
124 show_model(http_client.as_ref(), &api_url, api_key.as_deref(), name)
125 .await?;
126 let ollama_model = ollama::Model::new(
127 name,
128 None,
129 None,
130 Some(capabilities.supports_tools()),
131 Some(capabilities.supports_vision()),
132 Some(capabilities.supports_thinking()),
133 );
134 Ok(ollama_model)
135 }
136 });
137
138 // Rate-limit capability fetches
139 // since there is an arbitrary number of models available
140 let mut ollama_models: Vec<_> = futures::stream::iter(tasks)
141 .buffer_unordered(5)
142 .collect::<Vec<Result<_>>>()
143 .await
144 .into_iter()
145 .collect::<Result<Vec<_>>>()?;
146
147 ollama_models.sort_by(|a, b| a.name.cmp(&b.name));
148
149 this.update(cx, |this, cx| {
150 this.fetched_models = ollama_models;
151 cx.notify();
152 })
153 })
154 }
155
156 fn restart_fetch_models_task(&mut self, cx: &mut Context<Self>) {
157 let task = self.fetch_models(cx);
158 self.fetch_model_task.replace(task);
159 }
160}
161
162impl OllamaLanguageModelProvider {
163 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
164 let this = Self {
165 http_client: http_client.clone(),
166 state: cx.new(|cx| {
167 cx.observe_global::<SettingsStore>({
168 let mut last_settings = OllamaLanguageModelProvider::settings(cx).clone();
169 move |this: &mut State, cx| {
170 let current_settings = OllamaLanguageModelProvider::settings(cx);
171 let settings_changed = current_settings != &last_settings;
172 if settings_changed {
173 let url_changed = last_settings.api_url != current_settings.api_url;
174 last_settings = current_settings.clone();
175 if url_changed {
176 this.fetched_models.clear();
177 this.authenticate(cx).detach();
178 }
179 cx.notify();
180 }
181 }
182 })
183 .detach();
184
185 State {
186 http_client,
187 fetched_models: Default::default(),
188 fetch_model_task: None,
189 api_key_state: ApiKeyState::new(Self::api_url(cx)),
190 }
191 }),
192 };
193 this
194 }
195
196 fn settings(cx: &App) -> &OllamaSettings {
197 &AllLanguageModelSettings::get_global(cx).ollama
198 }
199
200 fn api_url(cx: &App) -> SharedString {
201 let api_url = &Self::settings(cx).api_url;
202 if api_url.is_empty() {
203 OLLAMA_API_URL.into()
204 } else {
205 SharedString::new(api_url.as_str())
206 }
207 }
208}
209
210impl LanguageModelProviderState for OllamaLanguageModelProvider {
211 type ObservableEntity = State;
212
213 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
214 Some(self.state.clone())
215 }
216}
217
218impl LanguageModelProvider for OllamaLanguageModelProvider {
219 fn id(&self) -> LanguageModelProviderId {
220 PROVIDER_ID
221 }
222
223 fn name(&self) -> LanguageModelProviderName {
224 PROVIDER_NAME
225 }
226
227 fn icon(&self) -> IconName {
228 IconName::AiOllama
229 }
230
231 fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
232 // We shouldn't try to select default model, because it might lead to a load call for an unloaded model.
233 // In a constrained environment where user might not have enough resources it'll be a bad UX to select something
234 // to load by default.
235 None
236 }
237
238 fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
239 // See explanation for default_model.
240 None
241 }
242
243 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
244 let mut models: HashMap<String, ollama::Model> = HashMap::new();
245
246 // Add models from the Ollama API
247 for model in self.state.read(cx).fetched_models.iter() {
248 models.insert(model.name.clone(), model.clone());
249 }
250
251 // Override with available models from settings
252 for setting_model in &OllamaLanguageModelProvider::settings(cx).available_models {
253 let setting_base = setting_model.name.split(':').next().unwrap();
254 if let Some(model) = models
255 .values_mut()
256 .find(|m| m.name.split(':').next().unwrap() == setting_base)
257 {
258 model.max_tokens = setting_model.max_tokens;
259 model.display_name = setting_model.display_name.clone();
260 model.keep_alive = setting_model.keep_alive.clone();
261 model.supports_tools = setting_model.supports_tools;
262 model.supports_vision = setting_model.supports_images;
263 model.supports_thinking = setting_model.supports_thinking;
264 } else {
265 models.insert(
266 setting_model.name.clone(),
267 ollama::Model {
268 name: setting_model.name.clone(),
269 display_name: setting_model.display_name.clone(),
270 max_tokens: setting_model.max_tokens,
271 keep_alive: setting_model.keep_alive.clone(),
272 supports_tools: setting_model.supports_tools,
273 supports_vision: setting_model.supports_images,
274 supports_thinking: setting_model.supports_thinking,
275 },
276 );
277 }
278 }
279
280 let mut models = models
281 .into_values()
282 .map(|model| {
283 Arc::new(OllamaLanguageModel {
284 id: LanguageModelId::from(model.name.clone()),
285 model,
286 http_client: self.http_client.clone(),
287 request_limiter: RateLimiter::new(4),
288 state: self.state.clone(),
289 }) as Arc<dyn LanguageModel>
290 })
291 .collect::<Vec<_>>();
292 models.sort_by_key(|model| model.name());
293 models
294 }
295
296 fn is_authenticated(&self, cx: &App) -> bool {
297 self.state.read(cx).is_authenticated()
298 }
299
300 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
301 self.state.update(cx, |state, cx| state.authenticate(cx))
302 }
303
304 fn configuration_view(
305 &self,
306 _target_agent: language_model::ConfigurationViewTargetAgent,
307 window: &mut Window,
308 cx: &mut App,
309 ) -> AnyView {
310 let state = self.state.clone();
311 cx.new(|cx| ConfigurationView::new(state, window, cx))
312 .into()
313 }
314
315 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
316 self.state
317 .update(cx, |state, cx| state.set_api_key(None, cx))
318 }
319}
320
321pub struct OllamaLanguageModel {
322 id: LanguageModelId,
323 model: ollama::Model,
324 http_client: Arc<dyn HttpClient>,
325 request_limiter: RateLimiter,
326 state: gpui::Entity<State>,
327}
328
329impl OllamaLanguageModel {
330 fn to_ollama_request(&self, request: LanguageModelRequest) -> ChatRequest {
331 let supports_vision = self.model.supports_vision.unwrap_or(false);
332
333 let mut messages = Vec::with_capacity(request.messages.len());
334
335 for mut msg in request.messages.into_iter() {
336 let images = if supports_vision {
337 msg.content
338 .iter()
339 .filter_map(|content| match content {
340 MessageContent::Image(image) => Some(image.source.to_string()),
341 _ => None,
342 })
343 .collect::<Vec<String>>()
344 } else {
345 vec![]
346 };
347
348 match msg.role {
349 Role::User => {
350 for tool_result in msg
351 .content
352 .extract_if(.., |x| matches!(x, MessageContent::ToolResult(..)))
353 {
354 match tool_result {
355 MessageContent::ToolResult(tool_result) => {
356 messages.push(ChatMessage::Tool {
357 tool_name: tool_result.tool_name.to_string(),
358 content: tool_result.content.to_str().unwrap_or("").to_string(),
359 })
360 }
361 _ => unreachable!("Only tool result should be extracted"),
362 }
363 }
364 if !msg.content.is_empty() {
365 messages.push(ChatMessage::User {
366 content: msg.string_contents(),
367 images: if images.is_empty() {
368 None
369 } else {
370 Some(images)
371 },
372 })
373 }
374 }
375 Role::Assistant => {
376 let content = msg.string_contents();
377 let mut thinking = None;
378 let mut tool_calls = Vec::new();
379 for content in msg.content.into_iter() {
380 match content {
381 MessageContent::Thinking { text, .. } if !text.is_empty() => {
382 thinking = Some(text)
383 }
384 MessageContent::ToolUse(tool_use) => {
385 tool_calls.push(OllamaToolCall::Function(OllamaFunctionCall {
386 name: tool_use.name.to_string(),
387 arguments: tool_use.input,
388 }));
389 }
390 _ => (),
391 }
392 }
393 messages.push(ChatMessage::Assistant {
394 content,
395 tool_calls: Some(tool_calls),
396 images: if images.is_empty() {
397 None
398 } else {
399 Some(images)
400 },
401 thinking,
402 })
403 }
404 Role::System => messages.push(ChatMessage::System {
405 content: msg.string_contents(),
406 }),
407 }
408 }
409 ChatRequest {
410 model: self.model.name.clone(),
411 messages,
412 keep_alive: self.model.keep_alive.clone().unwrap_or_default(),
413 stream: true,
414 options: Some(ChatOptions {
415 num_ctx: Some(self.model.max_tokens),
416 stop: Some(request.stop),
417 temperature: request.temperature.or(Some(1.0)),
418 ..Default::default()
419 }),
420 think: self
421 .model
422 .supports_thinking
423 .map(|supports_thinking| supports_thinking && request.thinking_allowed),
424 tools: if self.model.supports_tools.unwrap_or(false) {
425 request.tools.into_iter().map(tool_into_ollama).collect()
426 } else {
427 vec![]
428 },
429 }
430 }
431}
432
433impl LanguageModel for OllamaLanguageModel {
434 fn id(&self) -> LanguageModelId {
435 self.id.clone()
436 }
437
438 fn name(&self) -> LanguageModelName {
439 LanguageModelName::from(self.model.display_name().to_string())
440 }
441
442 fn provider_id(&self) -> LanguageModelProviderId {
443 PROVIDER_ID
444 }
445
446 fn provider_name(&self) -> LanguageModelProviderName {
447 PROVIDER_NAME
448 }
449
450 fn supports_tools(&self) -> bool {
451 self.model.supports_tools.unwrap_or(false)
452 }
453
454 fn supports_images(&self) -> bool {
455 self.model.supports_vision.unwrap_or(false)
456 }
457
458 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
459 match choice {
460 LanguageModelToolChoice::Auto => false,
461 LanguageModelToolChoice::Any => false,
462 LanguageModelToolChoice::None => false,
463 }
464 }
465
466 fn telemetry_id(&self) -> String {
467 format!("ollama/{}", self.model.id())
468 }
469
470 fn max_token_count(&self) -> u64 {
471 self.model.max_token_count()
472 }
473
474 fn count_tokens(
475 &self,
476 request: LanguageModelRequest,
477 _cx: &App,
478 ) -> BoxFuture<'static, Result<u64>> {
479 // There is no endpoint for this _yet_ in Ollama
480 // see: https://github.com/ollama/ollama/issues/1716 and https://github.com/ollama/ollama/issues/3582
481 let token_count = request
482 .messages
483 .iter()
484 .map(|msg| msg.string_contents().chars().count())
485 .sum::<usize>()
486 / 4;
487
488 async move { Ok(token_count as u64) }.boxed()
489 }
490
491 fn stream_completion(
492 &self,
493 request: LanguageModelRequest,
494 cx: &AsyncApp,
495 ) -> BoxFuture<
496 'static,
497 Result<
498 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
499 LanguageModelCompletionError,
500 >,
501 > {
502 let request = self.to_ollama_request(request);
503
504 let http_client = self.http_client.clone();
505 let Ok((api_key, api_url)) = self.state.read_with(cx, |state, cx| {
506 let api_url = OllamaLanguageModelProvider::api_url(cx);
507 (state.api_key_state.key(&api_url), api_url)
508 }) else {
509 return futures::future::ready(Err(anyhow!("App state dropped").into())).boxed();
510 };
511
512 let future = self.request_limiter.stream(async move {
513 let stream =
514 stream_chat_completion(http_client.as_ref(), &api_url, api_key.as_deref(), request)
515 .await?;
516 let stream = map_to_language_model_completion_events(stream);
517 Ok(stream)
518 });
519
520 future.map_ok(|f| f.boxed()).boxed()
521 }
522}
523
524fn map_to_language_model_completion_events(
525 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
526) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
527 // Used for creating unique tool use ids
528 static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
529
530 struct State {
531 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
532 used_tools: bool,
533 }
534
535 // We need to create a ToolUse and Stop event from a single
536 // response from the original stream
537 let stream = stream::unfold(
538 State {
539 stream,
540 used_tools: false,
541 },
542 async move |mut state| {
543 let response = state.stream.next().await?;
544
545 let delta = match response {
546 Ok(delta) => delta,
547 Err(e) => {
548 let event = Err(LanguageModelCompletionError::from(anyhow!(e)));
549 return Some((vec![event], state));
550 }
551 };
552
553 let mut events = Vec::new();
554
555 match delta.message {
556 ChatMessage::User { content, images: _ } => {
557 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
558 }
559 ChatMessage::System { content } => {
560 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
561 }
562 ChatMessage::Tool { content, .. } => {
563 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
564 }
565 ChatMessage::Assistant {
566 content,
567 tool_calls,
568 images: _,
569 thinking,
570 } => {
571 if let Some(text) = thinking {
572 events.push(Ok(LanguageModelCompletionEvent::Thinking {
573 text,
574 signature: None,
575 }));
576 }
577
578 if let Some(tool_call) = tool_calls.and_then(|v| v.into_iter().next()) {
579 match tool_call {
580 OllamaToolCall::Function(function) => {
581 let tool_id = format!(
582 "{}-{}",
583 &function.name,
584 TOOL_CALL_COUNTER.fetch_add(1, Ordering::Relaxed)
585 );
586 let event =
587 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
588 id: LanguageModelToolUseId::from(tool_id),
589 name: Arc::from(function.name),
590 raw_input: function.arguments.to_string(),
591 input: function.arguments,
592 is_input_complete: true,
593 });
594 events.push(Ok(event));
595 state.used_tools = true;
596 }
597 }
598 } else if !content.is_empty() {
599 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
600 }
601 }
602 };
603
604 if delta.done {
605 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
606 input_tokens: delta.prompt_eval_count.unwrap_or(0),
607 output_tokens: delta.eval_count.unwrap_or(0),
608 cache_creation_input_tokens: 0,
609 cache_read_input_tokens: 0,
610 })));
611 if state.used_tools {
612 state.used_tools = false;
613 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
614 } else {
615 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
616 }
617 }
618
619 Some((events, state))
620 },
621 );
622
623 stream.flat_map(futures::stream::iter)
624}
625
626struct ConfigurationView {
627 api_key_editor: gpui::Entity<SingleLineInput>,
628 api_url_editor: gpui::Entity<SingleLineInput>,
629 state: gpui::Entity<State>,
630}
631
632impl ConfigurationView {
633 pub fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
634 let api_key_editor =
635 cx.new(|cx| SingleLineInput::new(window, cx, "63e02e...").label("API key"));
636
637 let api_url_editor = cx.new(|cx| {
638 let input = SingleLineInput::new(window, cx, OLLAMA_API_URL).label("API URL");
639 input.set_text(OllamaLanguageModelProvider::api_url(cx), window, cx);
640 input
641 });
642
643 cx.observe(&state, |_, _, cx| {
644 cx.notify();
645 })
646 .detach();
647
648 Self {
649 api_key_editor,
650 api_url_editor,
651 state,
652 }
653 }
654
655 fn retry_connection(&self, cx: &mut App) {
656 self.state
657 .update(cx, |state, cx| state.restart_fetch_models_task(cx));
658 }
659
660 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
661 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
662 if api_key.is_empty() {
663 return;
664 }
665
666 // url changes can cause the editor to be displayed again
667 self.api_key_editor
668 .update(cx, |input, cx| input.set_text("", window, cx));
669
670 let state = self.state.clone();
671 cx.spawn_in(window, async move |_, cx| {
672 state
673 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))?
674 .await
675 })
676 .detach_and_log_err(cx);
677 }
678
679 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
680 self.api_key_editor
681 .update(cx, |input, cx| input.set_text("", window, cx));
682
683 let state = self.state.clone();
684 cx.spawn_in(window, async move |_, cx| {
685 state
686 .update(cx, |state, cx| state.set_api_key(None, cx))?
687 .await
688 })
689 .detach_and_log_err(cx);
690
691 cx.notify();
692 }
693
694 fn save_api_url(&mut self, cx: &mut Context<Self>) {
695 let api_url = self.api_url_editor.read(cx).text(cx).trim().to_string();
696 let current_url = OllamaLanguageModelProvider::api_url(cx);
697 if !api_url.is_empty() && &api_url != ¤t_url {
698 let fs = <dyn Fs>::global(cx);
699 update_settings_file(fs, cx, move |settings, _| {
700 settings
701 .language_models
702 .get_or_insert_default()
703 .ollama
704 .get_or_insert_default()
705 .api_url = Some(api_url);
706 });
707 }
708 }
709
710 fn reset_api_url(&mut self, window: &mut Window, cx: &mut Context<Self>) {
711 self.api_url_editor
712 .update(cx, |input, cx| input.set_text("", window, cx));
713 let fs = <dyn Fs>::global(cx);
714 update_settings_file(fs, cx, |settings, _cx| {
715 if let Some(settings) = settings
716 .language_models
717 .as_mut()
718 .and_then(|models| models.ollama.as_mut())
719 {
720 settings.api_url = Some(OLLAMA_API_URL.into());
721 }
722 });
723 cx.notify();
724 }
725
726 fn render_instructions() -> Div {
727 v_flex()
728 .gap_2()
729 .child(Label::new(
730 "Run LLMs locally on your machine with Ollama, or connect to an Ollama server. \
731 Can provide access to Llama, Mistral, Gemma, and hundreds of other models.",
732 ))
733 .child(Label::new("To use local Ollama:"))
734 .child(
735 List::new()
736 .child(InstructionListItem::new(
737 "Download and install Ollama from",
738 Some("ollama.com"),
739 Some("https://ollama.com/download"),
740 ))
741 .child(InstructionListItem::text_only(
742 "Start Ollama and download a model: `ollama run gpt-oss:20b`",
743 ))
744 .child(InstructionListItem::text_only(
745 "Click 'Connect' below to start using Ollama in Zed",
746 )),
747 )
748 .child(Label::new(
749 "Alternatively, you can connect to an Ollama server by specifying its \
750 URL and API key (may not be required):",
751 ))
752 }
753
754 fn render_api_key_editor(&self, cx: &Context<Self>) -> Div {
755 let state = self.state.read(cx);
756 let env_var_set = state.api_key_state.is_from_env_var();
757
758 if !state.api_key_state.has_key() {
759 v_flex()
760 .on_action(cx.listener(Self::save_api_key))
761 .child(self.api_key_editor.clone())
762 .child(
763 Label::new(
764 format!("You can also assign the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed.")
765 )
766 .size(LabelSize::Small)
767 .color(Color::Muted),
768 )
769 } else {
770 h_flex()
771 .p_3()
772 .justify_between()
773 .rounded_md()
774 .border_1()
775 .border_color(cx.theme().colors().border)
776 .bg(cx.theme().colors().elevated_surface_background)
777 .child(
778 h_flex()
779 .gap_2()
780 .child(Icon::new(IconName::Check).color(Color::Success))
781 .child(
782 Label::new(
783 if env_var_set {
784 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable.")
785 } else {
786 "API key configured".to_string()
787 }
788 )
789 )
790 )
791 .child(
792 Button::new("reset-api-key", "Reset API Key")
793 .label_size(LabelSize::Small)
794 .icon(IconName::Undo)
795 .icon_size(IconSize::Small)
796 .icon_position(IconPosition::Start)
797 .layer(ElevationIndex::ModalSurface)
798 .when(env_var_set, |this| {
799 this.tooltip(Tooltip::text(format!("To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable.")))
800 })
801 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx))),
802 )
803 }
804 }
805
806 fn render_api_url_editor(&self, cx: &Context<Self>) -> Div {
807 let api_url = OllamaLanguageModelProvider::api_url(cx);
808 let custom_api_url_set = api_url != OLLAMA_API_URL;
809
810 if custom_api_url_set {
811 h_flex()
812 .p_3()
813 .justify_between()
814 .rounded_md()
815 .border_1()
816 .border_color(cx.theme().colors().border)
817 .bg(cx.theme().colors().elevated_surface_background)
818 .child(
819 h_flex()
820 .gap_2()
821 .child(Icon::new(IconName::Check).color(Color::Success))
822 .child(v_flex().gap_1().child(Label::new(api_url))),
823 )
824 .child(
825 Button::new("reset-api-url", "Reset API URL")
826 .label_size(LabelSize::Small)
827 .icon(IconName::Undo)
828 .icon_size(IconSize::Small)
829 .icon_position(IconPosition::Start)
830 .layer(ElevationIndex::ModalSurface)
831 .on_click(
832 cx.listener(|this, _, window, cx| this.reset_api_url(window, cx)),
833 ),
834 )
835 } else {
836 v_flex()
837 .on_action(cx.listener(|this, _: &menu::Confirm, _window, cx| {
838 this.save_api_url(cx);
839 cx.notify();
840 }))
841 .gap_2()
842 .child(self.api_url_editor.clone())
843 }
844 }
845}
846
847impl Render for ConfigurationView {
848 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
849 let is_authenticated = self.state.read(cx).is_authenticated();
850
851 v_flex()
852 .gap_2()
853 .child(Self::render_instructions())
854 .child(self.render_api_url_editor(cx))
855 .child(self.render_api_key_editor(cx))
856 .child(
857 h_flex()
858 .w_full()
859 .justify_between()
860 .gap_2()
861 .child(
862 h_flex()
863 .w_full()
864 .gap_2()
865 .map(|this| {
866 if is_authenticated {
867 this.child(
868 Button::new("ollama-site", "Ollama")
869 .style(ButtonStyle::Subtle)
870 .icon(IconName::ArrowUpRight)
871 .icon_size(IconSize::XSmall)
872 .icon_color(Color::Muted)
873 .on_click(move |_, _, cx| cx.open_url(OLLAMA_SITE))
874 .into_any_element(),
875 )
876 } else {
877 this.child(
878 Button::new("download_ollama_button", "Download Ollama")
879 .style(ButtonStyle::Subtle)
880 .icon(IconName::ArrowUpRight)
881 .icon_size(IconSize::XSmall)
882 .icon_color(Color::Muted)
883 .on_click(move |_, _, cx| {
884 cx.open_url(OLLAMA_DOWNLOAD_URL)
885 })
886 .into_any_element(),
887 )
888 }
889 })
890 .child(
891 Button::new("view-models", "View All Models")
892 .style(ButtonStyle::Subtle)
893 .icon(IconName::ArrowUpRight)
894 .icon_size(IconSize::XSmall)
895 .icon_color(Color::Muted)
896 .on_click(move |_, _, cx| cx.open_url(OLLAMA_LIBRARY_URL)),
897 ),
898 )
899 .map(|this| {
900 if is_authenticated {
901 this.child(
902 ButtonLike::new("connected")
903 .disabled(true)
904 .cursor_style(gpui::CursorStyle::Arrow)
905 .child(
906 h_flex()
907 .gap_2()
908 .child(Icon::new(IconName::Check).color(Color::Success))
909 .child(Label::new("Connected"))
910 .into_any_element(),
911 ),
912 )
913 } else {
914 this.child(
915 Button::new("retry_ollama_models", "Connect")
916 .icon_position(IconPosition::Start)
917 .icon_size(IconSize::XSmall)
918 .icon(IconName::PlayOutlined)
919 .on_click(
920 cx.listener(move |this, _, _, cx| {
921 this.retry_connection(cx)
922 }),
923 ),
924 )
925 }
926 }),
927 )
928 }
929}
930
931fn tool_into_ollama(tool: LanguageModelRequestTool) -> ollama::OllamaTool {
932 ollama::OllamaTool::Function {
933 function: OllamaFunctionTool {
934 name: tool.name,
935 description: Some(tool.description),
936 parameters: Some(tool.input_schema),
937 },
938 }
939}