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, CursorStyle, Entity, Task};
6use http_client::HttpClient;
7use language_model::{
8 ApiKeyState, AuthenticateError, EnvVar, IconOrSvg, LanguageModel, LanguageModelCompletionError,
9 LanguageModelCompletionEvent, LanguageModelId, LanguageModelName, LanguageModelProvider,
10 LanguageModelProviderId, LanguageModelProviderName, LanguageModelProviderState,
11 LanguageModelRequest, LanguageModelRequestTool, LanguageModelToolChoice, LanguageModelToolUse,
12 LanguageModelToolUseId, MessageContent, RateLimiter, Role, StopReason, TokenUsage, env_var,
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::{
26 ButtonLike, ButtonLink, ConfiguredApiCard, ElevationIndex, List, ListBulletItem, Tooltip,
27 prelude::*,
28};
29use ui_input::InputField;
30
31use crate::AllLanguageModelSettings;
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 auto_discover: bool,
47 pub available_models: Vec<AvailableModel>,
48}
49
50pub struct OllamaLanguageModelProvider {
51 http_client: Arc<dyn HttpClient>,
52 state: Entity<State>,
53}
54
55pub struct State {
56 api_key_state: ApiKeyState,
57 http_client: Arc<dyn HttpClient>,
58 fetched_models: Vec<ollama::Model>,
59 fetch_model_task: Option<Task<Result<()>>>,
60}
61
62impl State {
63 fn is_authenticated(&self) -> bool {
64 !self.fetched_models.is_empty()
65 }
66
67 fn set_api_key(&mut self, api_key: Option<String>, cx: &mut Context<Self>) -> Task<Result<()>> {
68 let api_url = OllamaLanguageModelProvider::api_url(cx);
69 let task = self
70 .api_key_state
71 .store(api_url, api_key, |this| &mut this.api_key_state, cx);
72
73 self.fetched_models.clear();
74 cx.spawn(async move |this, cx| {
75 let result = task.await;
76 this.update(cx, |this, cx| this.restart_fetch_models_task(cx))
77 .ok();
78 result
79 })
80 }
81
82 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
83 let api_url = OllamaLanguageModelProvider::api_url(cx);
84 let task = self
85 .api_key_state
86 .load_if_needed(api_url, |this| &mut this.api_key_state, cx);
87
88 // Always try to fetch models - if no API key is needed (local Ollama), it will work
89 // If API key is needed and provided, it will work
90 // If API key is needed and not provided, it will fail gracefully
91 cx.spawn(async move |this, cx| {
92 let result = task.await;
93 this.update(cx, |this, cx| this.restart_fetch_models_task(cx))
94 .ok();
95 result
96 })
97 }
98
99 fn fetch_models(&mut self, cx: &mut Context<Self>) -> Task<Result<()>> {
100 let http_client = Arc::clone(&self.http_client);
101 let api_url = OllamaLanguageModelProvider::api_url(cx);
102 let api_key = self.api_key_state.key(&api_url);
103
104 // As a proxy for the server being "authenticated", we'll check if its up by fetching the models
105 cx.spawn(async move |this, cx| {
106 let models = get_models(http_client.as_ref(), &api_url, api_key.as_deref()).await?;
107
108 let tasks = models
109 .into_iter()
110 // Since there is no metadata from the Ollama API
111 // indicating which models are embedding models,
112 // simply filter out models with "-embed" in their name
113 .filter(|model| !model.name.contains("-embed"))
114 .map(|model| {
115 let http_client = Arc::clone(&http_client);
116 let api_url = api_url.clone();
117 let api_key = api_key.clone();
118 async move {
119 let name = model.name.as_str();
120 let model =
121 show_model(http_client.as_ref(), &api_url, api_key.as_deref(), name)
122 .await?;
123 let ollama_model = ollama::Model::new(
124 name,
125 None,
126 model.context_length,
127 Some(model.supports_tools()),
128 Some(model.supports_vision()),
129 Some(model.supports_thinking()),
130 );
131 Ok(ollama_model)
132 }
133 });
134
135 // Rate-limit capability fetches
136 // since there is an arbitrary number of models available
137 let mut ollama_models: Vec<_> = futures::stream::iter(tasks)
138 .buffer_unordered(5)
139 .collect::<Vec<Result<_>>>()
140 .await
141 .into_iter()
142 .collect::<Result<Vec<_>>>()?;
143
144 ollama_models.sort_by(|a, b| a.name.cmp(&b.name));
145
146 this.update(cx, |this, cx| {
147 this.fetched_models = ollama_models;
148 cx.notify();
149 })
150 })
151 }
152
153 fn restart_fetch_models_task(&mut self, cx: &mut Context<Self>) {
154 let task = self.fetch_models(cx);
155 self.fetch_model_task.replace(task);
156 }
157}
158
159impl OllamaLanguageModelProvider {
160 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
161 let this = Self {
162 http_client: http_client.clone(),
163 state: cx.new(|cx| {
164 cx.observe_global::<SettingsStore>({
165 let mut last_settings = OllamaLanguageModelProvider::settings(cx).clone();
166 move |this: &mut State, cx| {
167 let current_settings = OllamaLanguageModelProvider::settings(cx);
168 let settings_changed = current_settings != &last_settings;
169 if settings_changed {
170 let url_changed = last_settings.api_url != current_settings.api_url;
171 last_settings = current_settings.clone();
172 if url_changed {
173 this.fetched_models.clear();
174 this.authenticate(cx).detach();
175 }
176 cx.notify();
177 }
178 }
179 })
180 .detach();
181
182 State {
183 http_client,
184 fetched_models: Default::default(),
185 fetch_model_task: None,
186 api_key_state: ApiKeyState::new(Self::api_url(cx), (*API_KEY_ENV_VAR).clone()),
187 }
188 }),
189 };
190 this
191 }
192
193 fn settings(cx: &App) -> &OllamaSettings {
194 &AllLanguageModelSettings::get_global(cx).ollama
195 }
196
197 fn api_url(cx: &App) -> SharedString {
198 let api_url = &Self::settings(cx).api_url;
199 if api_url.is_empty() {
200 OLLAMA_API_URL.into()
201 } else {
202 SharedString::new(api_url.as_str())
203 }
204 }
205
206 fn has_custom_url(cx: &App) -> bool {
207 Self::settings(cx).api_url != OLLAMA_API_URL
208 }
209}
210
211impl LanguageModelProviderState for OllamaLanguageModelProvider {
212 type ObservableEntity = State;
213
214 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
215 Some(self.state.clone())
216 }
217}
218
219impl LanguageModelProvider for OllamaLanguageModelProvider {
220 fn id(&self) -> LanguageModelProviderId {
221 PROVIDER_ID
222 }
223
224 fn name(&self) -> LanguageModelProviderName {
225 PROVIDER_NAME
226 }
227
228 fn icon(&self) -> IconOrSvg {
229 IconOrSvg::Icon(IconName::AiOllama)
230 }
231
232 fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
233 // We shouldn't try to select default model, because it might lead to a load call for an unloaded model.
234 // In a constrained environment where user might not have enough resources it'll be a bad UX to select something
235 // to load by default.
236 None
237 }
238
239 fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
240 // See explanation for default_model.
241 None
242 }
243
244 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
245 let mut models: HashMap<String, ollama::Model> = HashMap::new();
246 let settings = OllamaLanguageModelProvider::settings(cx);
247
248 // Add models from the Ollama API
249 if settings.auto_discover {
250 for model in self.state.read(cx).fetched_models.iter() {
251 models.insert(model.name.clone(), model.clone());
252 }
253 }
254
255 // Override with available models from settings
256 merge_settings_into_models(&mut models, &settings.available_models);
257
258 let mut models = models
259 .into_values()
260 .map(|model| {
261 Arc::new(OllamaLanguageModel {
262 id: LanguageModelId::from(model.name.clone()),
263 model,
264 http_client: self.http_client.clone(),
265 request_limiter: RateLimiter::new(4),
266 state: self.state.clone(),
267 }) as Arc<dyn LanguageModel>
268 })
269 .collect::<Vec<_>>();
270 models.sort_by_key(|model| model.name());
271 models
272 }
273
274 fn is_authenticated(&self, cx: &App) -> bool {
275 self.state.read(cx).is_authenticated()
276 }
277
278 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
279 self.state.update(cx, |state, cx| state.authenticate(cx))
280 }
281
282 fn configuration_view(
283 &self,
284 _target_agent: language_model::ConfigurationViewTargetAgent,
285 window: &mut Window,
286 cx: &mut App,
287 ) -> AnyView {
288 let state = self.state.clone();
289 cx.new(|cx| ConfigurationView::new(state, window, cx))
290 .into()
291 }
292
293 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
294 self.state
295 .update(cx, |state, cx| state.set_api_key(None, cx))
296 }
297}
298
299pub struct OllamaLanguageModel {
300 id: LanguageModelId,
301 model: ollama::Model,
302 http_client: Arc<dyn HttpClient>,
303 request_limiter: RateLimiter,
304 state: Entity<State>,
305}
306
307impl OllamaLanguageModel {
308 fn to_ollama_request(&self, request: LanguageModelRequest) -> ChatRequest {
309 let supports_vision = self.model.supports_vision.unwrap_or(false);
310
311 let mut messages = Vec::with_capacity(request.messages.len());
312
313 for mut msg in request.messages.into_iter() {
314 let images = if supports_vision {
315 msg.content
316 .iter()
317 .filter_map(|content| match content {
318 MessageContent::Image(image) => Some(image.source.to_string()),
319 _ => None,
320 })
321 .collect::<Vec<String>>()
322 } else {
323 vec![]
324 };
325
326 match msg.role {
327 Role::User => {
328 for tool_result in msg
329 .content
330 .extract_if(.., |x| matches!(x, MessageContent::ToolResult(..)))
331 {
332 match tool_result {
333 MessageContent::ToolResult(tool_result) => {
334 messages.push(ChatMessage::Tool {
335 tool_name: tool_result.tool_name.to_string(),
336 content: tool_result.content.to_str().unwrap_or("").to_string(),
337 })
338 }
339 _ => unreachable!("Only tool result should be extracted"),
340 }
341 }
342 if !msg.content.is_empty() {
343 messages.push(ChatMessage::User {
344 content: msg.string_contents(),
345 images: if images.is_empty() {
346 None
347 } else {
348 Some(images)
349 },
350 })
351 }
352 }
353 Role::Assistant => {
354 let content = msg.string_contents();
355 let mut thinking = None;
356 let mut tool_calls = Vec::new();
357 for content in msg.content.into_iter() {
358 match content {
359 MessageContent::Thinking { text, .. } if !text.is_empty() => {
360 thinking = Some(text)
361 }
362 MessageContent::ToolUse(tool_use) => {
363 tool_calls.push(OllamaToolCall {
364 id: Some(tool_use.id.to_string()),
365 function: OllamaFunctionCall {
366 name: tool_use.name.to_string(),
367 arguments: tool_use.input,
368 },
369 });
370 }
371 _ => (),
372 }
373 }
374 messages.push(ChatMessage::Assistant {
375 content,
376 tool_calls: Some(tool_calls),
377 images: if images.is_empty() {
378 None
379 } else {
380 Some(images)
381 },
382 thinking,
383 })
384 }
385 Role::System => messages.push(ChatMessage::System {
386 content: msg.string_contents(),
387 }),
388 }
389 }
390 ChatRequest {
391 model: self.model.name.clone(),
392 messages,
393 keep_alive: self.model.keep_alive.clone().unwrap_or_default(),
394 stream: true,
395 options: Some(ChatOptions {
396 num_ctx: Some(self.model.max_tokens),
397 stop: Some(request.stop),
398 temperature: request.temperature.or(Some(1.0)),
399 ..Default::default()
400 }),
401 think: self
402 .model
403 .supports_thinking
404 .map(|supports_thinking| supports_thinking && request.thinking_allowed),
405 tools: if self.model.supports_tools.unwrap_or(false) {
406 request.tools.into_iter().map(tool_into_ollama).collect()
407 } else {
408 vec![]
409 },
410 }
411 }
412}
413
414impl LanguageModel for OllamaLanguageModel {
415 fn id(&self) -> LanguageModelId {
416 self.id.clone()
417 }
418
419 fn name(&self) -> LanguageModelName {
420 LanguageModelName::from(self.model.display_name().to_string())
421 }
422
423 fn provider_id(&self) -> LanguageModelProviderId {
424 PROVIDER_ID
425 }
426
427 fn provider_name(&self) -> LanguageModelProviderName {
428 PROVIDER_NAME
429 }
430
431 fn supports_tools(&self) -> bool {
432 self.model.supports_tools.unwrap_or(false)
433 }
434
435 fn supports_images(&self) -> bool {
436 self.model.supports_vision.unwrap_or(false)
437 }
438
439 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
440 match choice {
441 LanguageModelToolChoice::Auto => false,
442 LanguageModelToolChoice::Any => false,
443 LanguageModelToolChoice::None => false,
444 }
445 }
446
447 fn telemetry_id(&self) -> String {
448 format!("ollama/{}", self.model.id())
449 }
450
451 fn max_token_count(&self) -> u64 {
452 self.model.max_token_count()
453 }
454
455 fn count_tokens(
456 &self,
457 request: LanguageModelRequest,
458 _cx: &App,
459 ) -> BoxFuture<'static, Result<u64>> {
460 // There is no endpoint for this _yet_ in Ollama
461 // see: https://github.com/ollama/ollama/issues/1716 and https://github.com/ollama/ollama/issues/3582
462 let token_count = request
463 .messages
464 .iter()
465 .map(|msg| msg.string_contents().chars().count())
466 .sum::<usize>()
467 / 4;
468
469 async move { Ok(token_count as u64) }.boxed()
470 }
471
472 fn stream_completion(
473 &self,
474 request: LanguageModelRequest,
475 cx: &AsyncApp,
476 ) -> BoxFuture<
477 'static,
478 Result<
479 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
480 LanguageModelCompletionError,
481 >,
482 > {
483 let request = self.to_ollama_request(request);
484
485 let http_client = self.http_client.clone();
486 let (api_key, api_url) = self.state.read_with(cx, |state, cx| {
487 let api_url = OllamaLanguageModelProvider::api_url(cx);
488 (state.api_key_state.key(&api_url), api_url)
489 });
490
491 let future = self.request_limiter.stream(async move {
492 let stream =
493 stream_chat_completion(http_client.as_ref(), &api_url, api_key.as_deref(), request)
494 .await?;
495 let stream = map_to_language_model_completion_events(stream);
496 Ok(stream)
497 });
498
499 future.map_ok(|f| f.boxed()).boxed()
500 }
501}
502
503fn map_to_language_model_completion_events(
504 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
505) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
506 // Used for creating unique tool use ids
507 static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
508
509 struct State {
510 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
511 used_tools: bool,
512 }
513
514 // We need to create a ToolUse and Stop event from a single
515 // response from the original stream
516 let stream = stream::unfold(
517 State {
518 stream,
519 used_tools: false,
520 },
521 async move |mut state| {
522 let response = state.stream.next().await?;
523
524 let delta = match response {
525 Ok(delta) => delta,
526 Err(e) => {
527 let event = Err(LanguageModelCompletionError::from(anyhow!(e)));
528 return Some((vec![event], state));
529 }
530 };
531
532 let mut events = Vec::new();
533
534 match delta.message {
535 ChatMessage::User { content, images: _ } => {
536 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
537 }
538 ChatMessage::System { content } => {
539 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
540 }
541 ChatMessage::Tool { content, .. } => {
542 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
543 }
544 ChatMessage::Assistant {
545 content,
546 tool_calls,
547 images: _,
548 thinking,
549 } => {
550 if let Some(text) = thinking {
551 events.push(Ok(LanguageModelCompletionEvent::Thinking {
552 text,
553 signature: None,
554 }));
555 }
556
557 if let Some(tool_call) = tool_calls.and_then(|v| v.into_iter().next()) {
558 let OllamaToolCall { id, function } = tool_call;
559 let id = id.unwrap_or_else(|| {
560 format!(
561 "{}-{}",
562 &function.name,
563 TOOL_CALL_COUNTER.fetch_add(1, Ordering::Relaxed)
564 )
565 });
566 let event = LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
567 id: LanguageModelToolUseId::from(id),
568 name: Arc::from(function.name),
569 raw_input: function.arguments.to_string(),
570 input: function.arguments,
571 is_input_complete: true,
572 thought_signature: None,
573 });
574 events.push(Ok(event));
575 state.used_tools = true;
576 } else if !content.is_empty() {
577 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
578 }
579 }
580 };
581
582 if delta.done {
583 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
584 input_tokens: delta.prompt_eval_count.unwrap_or(0),
585 output_tokens: delta.eval_count.unwrap_or(0),
586 cache_creation_input_tokens: 0,
587 cache_read_input_tokens: 0,
588 })));
589 if state.used_tools {
590 state.used_tools = false;
591 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
592 } else {
593 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
594 }
595 }
596
597 Some((events, state))
598 },
599 );
600
601 stream.flat_map(futures::stream::iter)
602}
603
604struct ConfigurationView {
605 api_key_editor: Entity<InputField>,
606 api_url_editor: Entity<InputField>,
607 state: Entity<State>,
608}
609
610impl ConfigurationView {
611 pub fn new(state: Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
612 let api_key_editor = cx.new(|cx| InputField::new(window, cx, "63e02e...").label("API key"));
613
614 let api_url_editor = cx.new(|cx| {
615 let input = InputField::new(window, cx, OLLAMA_API_URL).label("API URL");
616 input.set_text(OllamaLanguageModelProvider::api_url(cx), window, cx);
617 input
618 });
619
620 cx.observe(&state, |_, _, cx| {
621 cx.notify();
622 })
623 .detach();
624
625 Self {
626 api_key_editor,
627 api_url_editor,
628 state,
629 }
630 }
631
632 fn retry_connection(&mut self, window: &mut Window, cx: &mut Context<Self>) {
633 let has_api_url = OllamaLanguageModelProvider::has_custom_url(cx);
634 let has_api_key = self
635 .state
636 .read_with(cx, |state, _| state.api_key_state.has_key());
637 if !has_api_url {
638 self.save_api_url(cx);
639 }
640 if !has_api_key {
641 self.save_api_key(&Default::default(), window, cx);
642 }
643
644 self.state.update(cx, |state, cx| {
645 state.restart_fetch_models_task(cx);
646 });
647 }
648
649 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
650 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
651 if api_key.is_empty() {
652 return;
653 }
654
655 // url changes can cause the editor to be displayed again
656 self.api_key_editor
657 .update(cx, |input, cx| input.set_text("", window, cx));
658
659 let state = self.state.clone();
660 cx.spawn_in(window, async move |_, cx| {
661 state
662 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))
663 .await
664 })
665 .detach_and_log_err(cx);
666 }
667
668 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
669 self.api_key_editor
670 .update(cx, |input, cx| input.set_text("", window, cx));
671
672 let state = self.state.clone();
673 cx.spawn_in(window, async move |_, cx| {
674 state
675 .update(cx, |state, cx| state.set_api_key(None, cx))
676 .await
677 })
678 .detach_and_log_err(cx);
679
680 cx.notify();
681 }
682
683 fn save_api_url(&self, cx: &mut Context<Self>) {
684 let api_url = self.api_url_editor.read(cx).text(cx).trim().to_string();
685 let current_url = OllamaLanguageModelProvider::api_url(cx);
686 if !api_url.is_empty() && &api_url != ¤t_url {
687 let fs = <dyn Fs>::global(cx);
688 update_settings_file(fs, cx, move |settings, _| {
689 settings
690 .language_models
691 .get_or_insert_default()
692 .ollama
693 .get_or_insert_default()
694 .api_url = Some(api_url);
695 });
696 }
697 }
698
699 fn reset_api_url(&mut self, window: &mut Window, cx: &mut Context<Self>) {
700 self.api_url_editor
701 .update(cx, |input, cx| input.set_text("", window, cx));
702 let fs = <dyn Fs>::global(cx);
703 update_settings_file(fs, cx, |settings, _cx| {
704 if let Some(settings) = settings
705 .language_models
706 .as_mut()
707 .and_then(|models| models.ollama.as_mut())
708 {
709 settings.api_url = Some(OLLAMA_API_URL.into());
710 }
711 });
712 cx.notify();
713 }
714
715 fn render_instructions(cx: &mut Context<Self>) -> Div {
716 v_flex()
717 .gap_2()
718 .child(Label::new(
719 "Run LLMs locally on your machine with Ollama, or connect to an Ollama server. \
720 Can provide access to Llama, Mistral, Gemma, and hundreds of other models.",
721 ))
722 .child(Label::new("To use local Ollama:"))
723 .child(
724 List::new()
725 .child(
726 ListBulletItem::new("")
727 .child(Label::new("Download and install Ollama from"))
728 .child(ButtonLink::new("ollama.com", "https://ollama.com/download")),
729 )
730 .child(
731 ListBulletItem::new("")
732 .child(Label::new("Start Ollama and download a model:"))
733 .child(Label::new("ollama run gpt-oss:20b").inline_code(cx)),
734 )
735 .child(ListBulletItem::new(
736 "Click 'Connect' below to start using Ollama in Zed",
737 )),
738 )
739 .child(Label::new(
740 "Alternatively, you can connect to an Ollama server by specifying its \
741 URL and API key (may not be required):",
742 ))
743 }
744
745 fn render_api_key_editor(&self, cx: &Context<Self>) -> impl IntoElement {
746 let state = self.state.read(cx);
747 let env_var_set = state.api_key_state.is_from_env_var();
748 let configured_card_label = if env_var_set {
749 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable.")
750 } else {
751 "API key configured".to_string()
752 };
753
754 if !state.api_key_state.has_key() {
755 v_flex()
756 .on_action(cx.listener(Self::save_api_key))
757 .child(self.api_key_editor.clone())
758 .child(
759 Label::new(
760 format!("You can also set the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed.")
761 )
762 .size(LabelSize::Small)
763 .color(Color::Muted),
764 )
765 .into_any_element()
766 } else {
767 ConfiguredApiCard::new(configured_card_label)
768 .disabled(env_var_set)
769 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
770 .when(env_var_set, |this| {
771 this.tooltip_label(format!("To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable."))
772 })
773 .into_any_element()
774 }
775 }
776
777 fn render_api_url_editor(&self, cx: &Context<Self>) -> Div {
778 let api_url = OllamaLanguageModelProvider::api_url(cx);
779 let custom_api_url_set = api_url != OLLAMA_API_URL;
780
781 if custom_api_url_set {
782 h_flex()
783 .p_3()
784 .justify_between()
785 .rounded_md()
786 .border_1()
787 .border_color(cx.theme().colors().border)
788 .bg(cx.theme().colors().elevated_surface_background)
789 .child(
790 h_flex()
791 .gap_2()
792 .child(Icon::new(IconName::Check).color(Color::Success))
793 .child(v_flex().gap_1().child(Label::new(api_url))),
794 )
795 .child(
796 Button::new("reset-api-url", "Reset API URL")
797 .label_size(LabelSize::Small)
798 .icon(IconName::Undo)
799 .icon_size(IconSize::Small)
800 .icon_position(IconPosition::Start)
801 .layer(ElevationIndex::ModalSurface)
802 .on_click(
803 cx.listener(|this, _, window, cx| this.reset_api_url(window, cx)),
804 ),
805 )
806 } else {
807 v_flex()
808 .on_action(cx.listener(|this, _: &menu::Confirm, _window, cx| {
809 this.save_api_url(cx);
810 cx.notify();
811 }))
812 .gap_2()
813 .child(self.api_url_editor.clone())
814 }
815 }
816}
817
818impl Render for ConfigurationView {
819 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
820 let is_authenticated = self.state.read(cx).is_authenticated();
821
822 v_flex()
823 .gap_2()
824 .child(Self::render_instructions(cx))
825 .child(self.render_api_url_editor(cx))
826 .child(self.render_api_key_editor(cx))
827 .child(
828 h_flex()
829 .w_full()
830 .justify_between()
831 .gap_2()
832 .child(
833 h_flex()
834 .w_full()
835 .gap_2()
836 .map(|this| {
837 if is_authenticated {
838 this.child(
839 Button::new("ollama-site", "Ollama")
840 .style(ButtonStyle::Subtle)
841 .icon(IconName::ArrowUpRight)
842 .icon_size(IconSize::XSmall)
843 .icon_color(Color::Muted)
844 .on_click(move |_, _, cx| cx.open_url(OLLAMA_SITE))
845 .into_any_element(),
846 )
847 } else {
848 this.child(
849 Button::new("download_ollama_button", "Download Ollama")
850 .style(ButtonStyle::Subtle)
851 .icon(IconName::ArrowUpRight)
852 .icon_size(IconSize::XSmall)
853 .icon_color(Color::Muted)
854 .on_click(move |_, _, cx| {
855 cx.open_url(OLLAMA_DOWNLOAD_URL)
856 })
857 .into_any_element(),
858 )
859 }
860 })
861 .child(
862 Button::new("view-models", "View All Models")
863 .style(ButtonStyle::Subtle)
864 .icon(IconName::ArrowUpRight)
865 .icon_size(IconSize::XSmall)
866 .icon_color(Color::Muted)
867 .on_click(move |_, _, cx| cx.open_url(OLLAMA_LIBRARY_URL)),
868 ),
869 )
870 .map(|this| {
871 if is_authenticated {
872 this.child(
873 ButtonLike::new("connected")
874 .disabled(true)
875 .cursor_style(CursorStyle::Arrow)
876 .child(
877 h_flex()
878 .gap_2()
879 .child(Icon::new(IconName::Check).color(Color::Success))
880 .child(Label::new("Connected"))
881 .into_any_element(),
882 )
883 .child(
884 IconButton::new("refresh-models", IconName::RotateCcw)
885 .tooltip(Tooltip::text("Refresh Models"))
886 .on_click(cx.listener(|this, _, window, cx| {
887 this.state.update(cx, |state, _| {
888 state.fetched_models.clear();
889 });
890 this.retry_connection(window, cx);
891 })),
892 ),
893 )
894 } else {
895 this.child(
896 Button::new("retry_ollama_models", "Connect")
897 .icon_position(IconPosition::Start)
898 .icon_size(IconSize::XSmall)
899 .icon(IconName::PlayOutlined)
900 .on_click(cx.listener(move |this, _, window, cx| {
901 this.retry_connection(window, cx)
902 })),
903 )
904 }
905 }),
906 )
907 }
908}
909
910fn merge_settings_into_models(
911 models: &mut HashMap<String, ollama::Model>,
912 available_models: &[AvailableModel],
913) {
914 for setting_model in available_models {
915 if let Some(model) = models.get_mut(&setting_model.name) {
916 model.max_tokens = setting_model.max_tokens;
917 model.display_name = setting_model.display_name.clone();
918 model.keep_alive = setting_model.keep_alive.clone();
919 model.supports_tools = setting_model.supports_tools;
920 model.supports_vision = setting_model.supports_images;
921 model.supports_thinking = setting_model.supports_thinking;
922 } else {
923 models.insert(
924 setting_model.name.clone(),
925 ollama::Model {
926 name: setting_model.name.clone(),
927 display_name: setting_model.display_name.clone(),
928 max_tokens: setting_model.max_tokens,
929 keep_alive: setting_model.keep_alive.clone(),
930 supports_tools: setting_model.supports_tools,
931 supports_vision: setting_model.supports_images,
932 supports_thinking: setting_model.supports_thinking,
933 },
934 );
935 }
936 }
937}
938
939fn tool_into_ollama(tool: LanguageModelRequestTool) -> ollama::OllamaTool {
940 ollama::OllamaTool::Function {
941 function: OllamaFunctionTool {
942 name: tool.name,
943 description: Some(tool.description),
944 parameters: Some(tool.input_schema),
945 },
946 }
947}
948
949#[cfg(test)]
950mod tests {
951 use super::*;
952
953 #[test]
954 fn test_merge_settings_preserves_display_names_for_similar_models() {
955 // Regression test for https://github.com/zed-industries/zed/issues/43646
956 // When multiple models share the same base name (e.g., qwen2.5-coder:1.5b and qwen2.5-coder:3b),
957 // each model should get its own display_name from settings, not a random one.
958
959 let mut models: HashMap<String, ollama::Model> = HashMap::new();
960 models.insert(
961 "qwen2.5-coder:1.5b".to_string(),
962 ollama::Model {
963 name: "qwen2.5-coder:1.5b".to_string(),
964 display_name: None,
965 max_tokens: 4096,
966 keep_alive: None,
967 supports_tools: None,
968 supports_vision: None,
969 supports_thinking: None,
970 },
971 );
972 models.insert(
973 "qwen2.5-coder:3b".to_string(),
974 ollama::Model {
975 name: "qwen2.5-coder:3b".to_string(),
976 display_name: None,
977 max_tokens: 4096,
978 keep_alive: None,
979 supports_tools: None,
980 supports_vision: None,
981 supports_thinking: None,
982 },
983 );
984
985 let available_models = vec![
986 AvailableModel {
987 name: "qwen2.5-coder:1.5b".to_string(),
988 display_name: Some("QWEN2.5 Coder 1.5B".to_string()),
989 max_tokens: 5000,
990 keep_alive: None,
991 supports_tools: Some(true),
992 supports_images: None,
993 supports_thinking: None,
994 },
995 AvailableModel {
996 name: "qwen2.5-coder:3b".to_string(),
997 display_name: Some("QWEN2.5 Coder 3B".to_string()),
998 max_tokens: 6000,
999 keep_alive: None,
1000 supports_tools: Some(true),
1001 supports_images: None,
1002 supports_thinking: None,
1003 },
1004 ];
1005
1006 merge_settings_into_models(&mut models, &available_models);
1007
1008 let model_1_5b = models
1009 .get("qwen2.5-coder:1.5b")
1010 .expect("1.5b model missing");
1011 let model_3b = models.get("qwen2.5-coder:3b").expect("3b model missing");
1012
1013 assert_eq!(
1014 model_1_5b.display_name,
1015 Some("QWEN2.5 Coder 1.5B".to_string()),
1016 "1.5b model should have its own display_name"
1017 );
1018 assert_eq!(model_1_5b.max_tokens, 5000);
1019
1020 assert_eq!(
1021 model_3b.display_name,
1022 Some("QWEN2.5 Coder 3B".to_string()),
1023 "3b model should have its own display_name"
1024 );
1025 assert_eq!(model_3b.max_tokens, 6000);
1026 }
1027}