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, 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
207impl LanguageModelProviderState for OllamaLanguageModelProvider {
208 type ObservableEntity = State;
209
210 fn observable_entity(&self) -> Option<Entity<Self::ObservableEntity>> {
211 Some(self.state.clone())
212 }
213}
214
215impl LanguageModelProvider for OllamaLanguageModelProvider {
216 fn id(&self) -> LanguageModelProviderId {
217 PROVIDER_ID
218 }
219
220 fn name(&self) -> LanguageModelProviderName {
221 PROVIDER_NAME
222 }
223
224 fn icon(&self) -> IconName {
225 IconName::AiOllama
226 }
227
228 fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
229 // We shouldn't try to select default model, because it might lead to a load call for an unloaded model.
230 // In a constrained environment where user might not have enough resources it'll be a bad UX to select something
231 // to load by default.
232 None
233 }
234
235 fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
236 // See explanation for default_model.
237 None
238 }
239
240 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
241 let mut models: HashMap<String, ollama::Model> = HashMap::new();
242 let settings = OllamaLanguageModelProvider::settings(cx);
243
244 // Add models from the Ollama API
245 if settings.auto_discover {
246 for model in self.state.read(cx).fetched_models.iter() {
247 models.insert(model.name.clone(), model.clone());
248 }
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: 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 {
386 id: Some(tool_use.id.to_string()),
387 function: OllamaFunctionCall {
388 name: tool_use.name.to_string(),
389 arguments: tool_use.input,
390 },
391 });
392 }
393 _ => (),
394 }
395 }
396 messages.push(ChatMessage::Assistant {
397 content,
398 tool_calls: Some(tool_calls),
399 images: if images.is_empty() {
400 None
401 } else {
402 Some(images)
403 },
404 thinking,
405 })
406 }
407 Role::System => messages.push(ChatMessage::System {
408 content: msg.string_contents(),
409 }),
410 }
411 }
412 ChatRequest {
413 model: self.model.name.clone(),
414 messages,
415 keep_alive: self.model.keep_alive.clone().unwrap_or_default(),
416 stream: true,
417 options: Some(ChatOptions {
418 num_ctx: Some(self.model.max_tokens),
419 stop: Some(request.stop),
420 temperature: request.temperature.or(Some(1.0)),
421 ..Default::default()
422 }),
423 think: self
424 .model
425 .supports_thinking
426 .map(|supports_thinking| supports_thinking && request.thinking_allowed),
427 tools: if self.model.supports_tools.unwrap_or(false) {
428 request.tools.into_iter().map(tool_into_ollama).collect()
429 } else {
430 vec![]
431 },
432 }
433 }
434}
435
436impl LanguageModel for OllamaLanguageModel {
437 fn id(&self) -> LanguageModelId {
438 self.id.clone()
439 }
440
441 fn name(&self) -> LanguageModelName {
442 LanguageModelName::from(self.model.display_name().to_string())
443 }
444
445 fn provider_id(&self) -> LanguageModelProviderId {
446 PROVIDER_ID
447 }
448
449 fn provider_name(&self) -> LanguageModelProviderName {
450 PROVIDER_NAME
451 }
452
453 fn supports_tools(&self) -> bool {
454 self.model.supports_tools.unwrap_or(false)
455 }
456
457 fn supports_images(&self) -> bool {
458 self.model.supports_vision.unwrap_or(false)
459 }
460
461 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
462 match choice {
463 LanguageModelToolChoice::Auto => false,
464 LanguageModelToolChoice::Any => false,
465 LanguageModelToolChoice::None => false,
466 }
467 }
468
469 fn telemetry_id(&self) -> String {
470 format!("ollama/{}", self.model.id())
471 }
472
473 fn max_token_count(&self) -> u64 {
474 self.model.max_token_count()
475 }
476
477 fn count_tokens(
478 &self,
479 request: LanguageModelRequest,
480 _cx: &App,
481 ) -> BoxFuture<'static, Result<u64>> {
482 // There is no endpoint for this _yet_ in Ollama
483 // see: https://github.com/ollama/ollama/issues/1716 and https://github.com/ollama/ollama/issues/3582
484 let token_count = request
485 .messages
486 .iter()
487 .map(|msg| msg.string_contents().chars().count())
488 .sum::<usize>()
489 / 4;
490
491 async move { Ok(token_count as u64) }.boxed()
492 }
493
494 fn stream_completion(
495 &self,
496 request: LanguageModelRequest,
497 cx: &AsyncApp,
498 ) -> BoxFuture<
499 'static,
500 Result<
501 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
502 LanguageModelCompletionError,
503 >,
504 > {
505 let request = self.to_ollama_request(request);
506
507 let http_client = self.http_client.clone();
508 let Ok((api_key, api_url)) = self.state.read_with(cx, |state, cx| {
509 let api_url = OllamaLanguageModelProvider::api_url(cx);
510 (state.api_key_state.key(&api_url), api_url)
511 }) else {
512 return futures::future::ready(Err(anyhow!("App state dropped").into())).boxed();
513 };
514
515 let future = self.request_limiter.stream(async move {
516 let stream =
517 stream_chat_completion(http_client.as_ref(), &api_url, api_key.as_deref(), request)
518 .await?;
519 let stream = map_to_language_model_completion_events(stream);
520 Ok(stream)
521 });
522
523 future.map_ok(|f| f.boxed()).boxed()
524 }
525}
526
527fn map_to_language_model_completion_events(
528 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
529) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
530 // Used for creating unique tool use ids
531 static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
532
533 struct State {
534 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
535 used_tools: bool,
536 }
537
538 // We need to create a ToolUse and Stop event from a single
539 // response from the original stream
540 let stream = stream::unfold(
541 State {
542 stream,
543 used_tools: false,
544 },
545 async move |mut state| {
546 let response = state.stream.next().await?;
547
548 let delta = match response {
549 Ok(delta) => delta,
550 Err(e) => {
551 let event = Err(LanguageModelCompletionError::from(anyhow!(e)));
552 return Some((vec![event], state));
553 }
554 };
555
556 let mut events = Vec::new();
557
558 match delta.message {
559 ChatMessage::User { content, images: _ } => {
560 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
561 }
562 ChatMessage::System { content } => {
563 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
564 }
565 ChatMessage::Tool { content, .. } => {
566 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
567 }
568 ChatMessage::Assistant {
569 content,
570 tool_calls,
571 images: _,
572 thinking,
573 } => {
574 if let Some(text) = thinking {
575 events.push(Ok(LanguageModelCompletionEvent::Thinking {
576 text,
577 signature: None,
578 }));
579 }
580
581 if let Some(tool_call) = tool_calls.and_then(|v| v.into_iter().next()) {
582 let OllamaToolCall { id, function } = tool_call;
583 let id = id.unwrap_or_else(|| {
584 format!(
585 "{}-{}",
586 &function.name,
587 TOOL_CALL_COUNTER.fetch_add(1, Ordering::Relaxed)
588 )
589 });
590 let event = LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
591 id: LanguageModelToolUseId::from(id),
592 name: Arc::from(function.name),
593 raw_input: function.arguments.to_string(),
594 input: function.arguments,
595 is_input_complete: true,
596 thought_signature: None,
597 });
598 events.push(Ok(event));
599 state.used_tools = true;
600 } else if !content.is_empty() {
601 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
602 }
603 }
604 };
605
606 if delta.done {
607 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
608 input_tokens: delta.prompt_eval_count.unwrap_or(0),
609 output_tokens: delta.eval_count.unwrap_or(0),
610 cache_creation_input_tokens: 0,
611 cache_read_input_tokens: 0,
612 })));
613 if state.used_tools {
614 state.used_tools = false;
615 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
616 } else {
617 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
618 }
619 }
620
621 Some((events, state))
622 },
623 );
624
625 stream.flat_map(futures::stream::iter)
626}
627
628struct ConfigurationView {
629 api_key_editor: Entity<InputField>,
630 api_url_editor: Entity<InputField>,
631 state: Entity<State>,
632}
633
634impl ConfigurationView {
635 pub fn new(state: Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
636 let api_key_editor = cx.new(|cx| InputField::new(window, cx, "63e02e...").label("API key"));
637
638 let api_url_editor = cx.new(|cx| {
639 let input = InputField::new(window, cx, OLLAMA_API_URL).label("API URL");
640 input.set_text(OllamaLanguageModelProvider::api_url(cx), window, cx);
641 input
642 });
643
644 cx.observe(&state, |_, _, cx| {
645 cx.notify();
646 })
647 .detach();
648
649 Self {
650 api_key_editor,
651 api_url_editor,
652 state,
653 }
654 }
655
656 fn retry_connection(&self, cx: &mut App) {
657 self.state
658 .update(cx, |state, cx| state.restart_fetch_models_task(cx));
659 }
660
661 fn save_api_key(&mut self, _: &menu::Confirm, window: &mut Window, cx: &mut Context<Self>) {
662 let api_key = self.api_key_editor.read(cx).text(cx).trim().to_string();
663 if api_key.is_empty() {
664 return;
665 }
666
667 // url changes can cause the editor to be displayed again
668 self.api_key_editor
669 .update(cx, |input, cx| input.set_text("", window, cx));
670
671 let state = self.state.clone();
672 cx.spawn_in(window, async move |_, cx| {
673 state
674 .update(cx, |state, cx| state.set_api_key(Some(api_key), cx))?
675 .await
676 })
677 .detach_and_log_err(cx);
678 }
679
680 fn reset_api_key(&mut self, window: &mut Window, cx: &mut Context<Self>) {
681 self.api_key_editor
682 .update(cx, |input, cx| input.set_text("", window, cx));
683
684 let state = self.state.clone();
685 cx.spawn_in(window, async move |_, cx| {
686 state
687 .update(cx, |state, cx| state.set_api_key(None, cx))?
688 .await
689 })
690 .detach_and_log_err(cx);
691
692 cx.notify();
693 }
694
695 fn save_api_url(&mut self, cx: &mut Context<Self>) {
696 let api_url = self.api_url_editor.read(cx).text(cx).trim().to_string();
697 let current_url = OllamaLanguageModelProvider::api_url(cx);
698 if !api_url.is_empty() && &api_url != ¤t_url {
699 let fs = <dyn Fs>::global(cx);
700 update_settings_file(fs, cx, move |settings, _| {
701 settings
702 .language_models
703 .get_or_insert_default()
704 .ollama
705 .get_or_insert_default()
706 .api_url = Some(api_url);
707 });
708 }
709 }
710
711 fn reset_api_url(&mut self, window: &mut Window, cx: &mut Context<Self>) {
712 self.api_url_editor
713 .update(cx, |input, cx| input.set_text("", window, cx));
714 let fs = <dyn Fs>::global(cx);
715 update_settings_file(fs, cx, |settings, _cx| {
716 if let Some(settings) = settings
717 .language_models
718 .as_mut()
719 .and_then(|models| models.ollama.as_mut())
720 {
721 settings.api_url = Some(OLLAMA_API_URL.into());
722 }
723 });
724 cx.notify();
725 }
726
727 fn render_instructions(cx: &mut Context<Self>) -> Div {
728 v_flex()
729 .gap_2()
730 .child(Label::new(
731 "Run LLMs locally on your machine with Ollama, or connect to an Ollama server. \
732 Can provide access to Llama, Mistral, Gemma, and hundreds of other models.",
733 ))
734 .child(Label::new("To use local Ollama:"))
735 .child(
736 List::new()
737 .child(
738 ListBulletItem::new("")
739 .child(Label::new("Download and install Ollama from"))
740 .child(ButtonLink::new("ollama.com", "https://ollama.com/download")),
741 )
742 .child(
743 ListBulletItem::new("")
744 .child(Label::new("Start Ollama and download a model:"))
745 .child(Label::new("ollama run gpt-oss:20b").inline_code(cx)),
746 )
747 .child(ListBulletItem::new(
748 "Click 'Connect' below to start using Ollama in Zed",
749 )),
750 )
751 .child(Label::new(
752 "Alternatively, you can connect to an Ollama server by specifying its \
753 URL and API key (may not be required):",
754 ))
755 }
756
757 fn render_api_key_editor(&self, cx: &Context<Self>) -> impl IntoElement {
758 let state = self.state.read(cx);
759 let env_var_set = state.api_key_state.is_from_env_var();
760 let configured_card_label = if env_var_set {
761 format!("API key set in {API_KEY_ENV_VAR_NAME} environment variable.")
762 } else {
763 "API key configured".to_string()
764 };
765
766 if !state.api_key_state.has_key() {
767 v_flex()
768 .on_action(cx.listener(Self::save_api_key))
769 .child(self.api_key_editor.clone())
770 .child(
771 Label::new(
772 format!("You can also assign the {API_KEY_ENV_VAR_NAME} environment variable and restart Zed.")
773 )
774 .size(LabelSize::Small)
775 .color(Color::Muted),
776 )
777 .into_any_element()
778 } else {
779 ConfiguredApiCard::new(configured_card_label)
780 .disabled(env_var_set)
781 .on_click(cx.listener(|this, _, window, cx| this.reset_api_key(window, cx)))
782 .when(env_var_set, |this| {
783 this.tooltip_label(format!("To reset your API key, unset the {API_KEY_ENV_VAR_NAME} environment variable."))
784 })
785 .into_any_element()
786 }
787 }
788
789 fn render_api_url_editor(&self, cx: &Context<Self>) -> Div {
790 let api_url = OllamaLanguageModelProvider::api_url(cx);
791 let custom_api_url_set = api_url != OLLAMA_API_URL;
792
793 if custom_api_url_set {
794 h_flex()
795 .p_3()
796 .justify_between()
797 .rounded_md()
798 .border_1()
799 .border_color(cx.theme().colors().border)
800 .bg(cx.theme().colors().elevated_surface_background)
801 .child(
802 h_flex()
803 .gap_2()
804 .child(Icon::new(IconName::Check).color(Color::Success))
805 .child(v_flex().gap_1().child(Label::new(api_url))),
806 )
807 .child(
808 Button::new("reset-api-url", "Reset API URL")
809 .label_size(LabelSize::Small)
810 .icon(IconName::Undo)
811 .icon_size(IconSize::Small)
812 .icon_position(IconPosition::Start)
813 .layer(ElevationIndex::ModalSurface)
814 .on_click(
815 cx.listener(|this, _, window, cx| this.reset_api_url(window, cx)),
816 ),
817 )
818 } else {
819 v_flex()
820 .on_action(cx.listener(|this, _: &menu::Confirm, _window, cx| {
821 this.save_api_url(cx);
822 cx.notify();
823 }))
824 .gap_2()
825 .child(self.api_url_editor.clone())
826 }
827 }
828}
829
830impl Render for ConfigurationView {
831 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
832 let is_authenticated = self.state.read(cx).is_authenticated();
833
834 v_flex()
835 .gap_2()
836 .child(Self::render_instructions(cx))
837 .child(self.render_api_url_editor(cx))
838 .child(self.render_api_key_editor(cx))
839 .child(
840 h_flex()
841 .w_full()
842 .justify_between()
843 .gap_2()
844 .child(
845 h_flex()
846 .w_full()
847 .gap_2()
848 .map(|this| {
849 if is_authenticated {
850 this.child(
851 Button::new("ollama-site", "Ollama")
852 .style(ButtonStyle::Subtle)
853 .icon(IconName::ArrowUpRight)
854 .icon_size(IconSize::XSmall)
855 .icon_color(Color::Muted)
856 .on_click(move |_, _, cx| cx.open_url(OLLAMA_SITE))
857 .into_any_element(),
858 )
859 } else {
860 this.child(
861 Button::new("download_ollama_button", "Download Ollama")
862 .style(ButtonStyle::Subtle)
863 .icon(IconName::ArrowUpRight)
864 .icon_size(IconSize::XSmall)
865 .icon_color(Color::Muted)
866 .on_click(move |_, _, cx| {
867 cx.open_url(OLLAMA_DOWNLOAD_URL)
868 })
869 .into_any_element(),
870 )
871 }
872 })
873 .child(
874 Button::new("view-models", "View All Models")
875 .style(ButtonStyle::Subtle)
876 .icon(IconName::ArrowUpRight)
877 .icon_size(IconSize::XSmall)
878 .icon_color(Color::Muted)
879 .on_click(move |_, _, cx| cx.open_url(OLLAMA_LIBRARY_URL)),
880 ),
881 )
882 .map(|this| {
883 if is_authenticated {
884 this.child(
885 ButtonLike::new("connected")
886 .disabled(true)
887 .cursor_style(CursorStyle::Arrow)
888 .child(
889 h_flex()
890 .gap_2()
891 .child(Icon::new(IconName::Check).color(Color::Success))
892 .child(Label::new("Connected"))
893 .into_any_element(),
894 )
895 .child(
896 IconButton::new("refresh-models", IconName::RotateCcw)
897 .tooltip(Tooltip::text("Refresh Models"))
898 .on_click(cx.listener(|this, _, _, cx| {
899 this.state.update(cx, |state, _| {
900 state.fetched_models.clear();
901 });
902 this.retry_connection(cx);
903 })),
904 ),
905 )
906 } else {
907 this.child(
908 Button::new("retry_ollama_models", "Connect")
909 .icon_position(IconPosition::Start)
910 .icon_size(IconSize::XSmall)
911 .icon(IconName::PlayOutlined)
912 .on_click(
913 cx.listener(move |this, _, _, cx| {
914 this.retry_connection(cx)
915 }),
916 ),
917 )
918 }
919 }),
920 )
921 }
922}
923
924fn tool_into_ollama(tool: LanguageModelRequestTool) -> ollama::OllamaTool {
925 ollama::OllamaTool::Function {
926 function: OllamaFunctionTool {
927 name: tool.name,
928 description: Some(tool.description),
929 parameters: Some(tool.input_schema),
930 },
931 }
932}