1use anyhow::{Result, anyhow};
2use futures::{FutureExt, StreamExt, future::BoxFuture, stream::BoxStream};
3use futures::{Stream, TryFutureExt, stream};
4use gpui::{AnyView, App, AsyncApp, Context, Subscription, Task};
5use http_client::HttpClient;
6use language_model::{
7 AuthenticateError, LanguageModel, LanguageModelCompletionError, LanguageModelCompletionEvent,
8 LanguageModelId, LanguageModelName, LanguageModelProvider, LanguageModelProviderId,
9 LanguageModelProviderName, LanguageModelProviderState, LanguageModelRequest,
10 LanguageModelRequestTool, LanguageModelToolChoice, LanguageModelToolUse,
11 LanguageModelToolUseId, MessageContent, RateLimiter, Role, StopReason, TokenUsage,
12};
13use ollama::{
14 ChatMessage, ChatOptions, ChatRequest, ChatResponseDelta, KeepAlive, OllamaFunctionTool,
15 OllamaToolCall, get_models, show_model, stream_chat_completion,
16};
17use schemars::JsonSchema;
18use serde::{Deserialize, Serialize};
19use settings::{Settings, SettingsStore};
20use std::pin::Pin;
21use std::sync::atomic::{AtomicU64, Ordering};
22use std::{collections::HashMap, sync::Arc};
23use ui::{ButtonLike, Indicator, List, prelude::*};
24use util::ResultExt;
25
26use crate::AllLanguageModelSettings;
27use crate::ui::InstructionListItem;
28
29const OLLAMA_DOWNLOAD_URL: &str = "https://ollama.com/download";
30const OLLAMA_LIBRARY_URL: &str = "https://ollama.com/library";
31const OLLAMA_SITE: &str = "https://ollama.com/";
32
33const PROVIDER_ID: LanguageModelProviderId = LanguageModelProviderId::new("ollama");
34const PROVIDER_NAME: LanguageModelProviderName = LanguageModelProviderName::new("Ollama");
35
36#[derive(Default, Debug, Clone, PartialEq)]
37pub struct OllamaSettings {
38 pub api_url: String,
39 pub available_models: Vec<AvailableModel>,
40}
41
42#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
43pub struct AvailableModel {
44 /// The model name in the Ollama API (e.g. "llama3.2:latest")
45 pub name: String,
46 /// The model's name in Zed's UI, such as in the model selector dropdown menu in the assistant panel.
47 pub display_name: Option<String>,
48 /// The Context Length parameter to the model (aka num_ctx or n_ctx)
49 pub max_tokens: u64,
50 /// The number of seconds to keep the connection open after the last request
51 pub keep_alive: Option<KeepAlive>,
52 /// Whether the model supports tools
53 pub supports_tools: Option<bool>,
54 /// Whether the model supports vision
55 pub supports_images: Option<bool>,
56 /// Whether to enable think mode
57 pub supports_thinking: Option<bool>,
58}
59
60pub struct OllamaLanguageModelProvider {
61 http_client: Arc<dyn HttpClient>,
62 state: gpui::Entity<State>,
63}
64
65pub struct State {
66 http_client: Arc<dyn HttpClient>,
67 available_models: Vec<ollama::Model>,
68 fetch_model_task: Option<Task<Result<()>>>,
69 _subscription: Subscription,
70}
71
72impl State {
73 fn is_authenticated(&self) -> bool {
74 !self.available_models.is_empty()
75 }
76
77 fn fetch_models(&mut self, cx: &mut Context<Self>) -> Task<Result<()>> {
78 let settings = &AllLanguageModelSettings::get_global(cx).ollama;
79 let http_client = Arc::clone(&self.http_client);
80 let api_url = settings.api_url.clone();
81
82 // As a proxy for the server being "authenticated", we'll check if its up by fetching the models
83 cx.spawn(async move |this, cx| {
84 let models = get_models(http_client.as_ref(), &api_url, None).await?;
85
86 let tasks = models
87 .into_iter()
88 // Since there is no metadata from the Ollama API
89 // indicating which models are embedding models,
90 // simply filter out models with "-embed" in their name
91 .filter(|model| !model.name.contains("-embed"))
92 .map(|model| {
93 let http_client = Arc::clone(&http_client);
94 let api_url = api_url.clone();
95 async move {
96 let name = model.name.as_str();
97 let capabilities = show_model(http_client.as_ref(), &api_url, name).await?;
98 let ollama_model = ollama::Model::new(
99 name,
100 None,
101 None,
102 Some(capabilities.supports_tools()),
103 Some(capabilities.supports_vision()),
104 Some(capabilities.supports_thinking()),
105 );
106 Ok(ollama_model)
107 }
108 });
109
110 // Rate-limit capability fetches
111 // since there is an arbitrary number of models available
112 let mut ollama_models: Vec<_> = futures::stream::iter(tasks)
113 .buffer_unordered(5)
114 .collect::<Vec<Result<_>>>()
115 .await
116 .into_iter()
117 .collect::<Result<Vec<_>>>()?;
118
119 ollama_models.sort_by(|a, b| a.name.cmp(&b.name));
120
121 this.update(cx, |this, cx| {
122 this.available_models = ollama_models;
123 cx.notify();
124 })
125 })
126 }
127
128 fn restart_fetch_models_task(&mut self, cx: &mut Context<Self>) {
129 let task = self.fetch_models(cx);
130 self.fetch_model_task.replace(task);
131 }
132
133 fn authenticate(&mut self, cx: &mut Context<Self>) -> Task<Result<(), AuthenticateError>> {
134 if self.is_authenticated() {
135 return Task::ready(Ok(()));
136 }
137
138 let fetch_models_task = self.fetch_models(cx);
139 cx.spawn(async move |_this, _cx| Ok(fetch_models_task.await?))
140 }
141}
142
143impl OllamaLanguageModelProvider {
144 pub fn new(http_client: Arc<dyn HttpClient>, cx: &mut App) -> Self {
145 let this = Self {
146 http_client: http_client.clone(),
147 state: cx.new(|cx| {
148 let subscription = cx.observe_global::<SettingsStore>({
149 let mut settings = AllLanguageModelSettings::get_global(cx).ollama.clone();
150 move |this: &mut State, cx| {
151 let new_settings = &AllLanguageModelSettings::get_global(cx).ollama;
152 if &settings != new_settings {
153 settings = new_settings.clone();
154 this.restart_fetch_models_task(cx);
155 cx.notify();
156 }
157 }
158 });
159
160 State {
161 http_client,
162 available_models: Default::default(),
163 fetch_model_task: None,
164 _subscription: subscription,
165 }
166 }),
167 };
168 this.state
169 .update(cx, |state, cx| state.restart_fetch_models_task(cx));
170 this
171 }
172}
173
174impl LanguageModelProviderState for OllamaLanguageModelProvider {
175 type ObservableEntity = State;
176
177 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>> {
178 Some(self.state.clone())
179 }
180}
181
182impl LanguageModelProvider for OllamaLanguageModelProvider {
183 fn id(&self) -> LanguageModelProviderId {
184 PROVIDER_ID
185 }
186
187 fn name(&self) -> LanguageModelProviderName {
188 PROVIDER_NAME
189 }
190
191 fn icon(&self) -> IconName {
192 IconName::AiOllama
193 }
194
195 fn default_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
196 // We shouldn't try to select default model, because it might lead to a load call for an unloaded model.
197 // In a constrained environment where user might not have enough resources it'll be a bad UX to select something
198 // to load by default.
199 None
200 }
201
202 fn default_fast_model(&self, _: &App) -> Option<Arc<dyn LanguageModel>> {
203 // See explanation for default_model.
204 None
205 }
206
207 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
208 let mut models: HashMap<String, ollama::Model> = HashMap::new();
209
210 // Add models from the Ollama API
211 for model in self.state.read(cx).available_models.iter() {
212 models.insert(model.name.clone(), model.clone());
213 }
214
215 // Override with available models from settings
216 for model in AllLanguageModelSettings::get_global(cx)
217 .ollama
218 .available_models
219 .iter()
220 {
221 models.insert(
222 model.name.clone(),
223 ollama::Model {
224 name: model.name.clone(),
225 display_name: model.display_name.clone(),
226 max_tokens: model.max_tokens,
227 keep_alive: model.keep_alive.clone(),
228 supports_tools: model.supports_tools,
229 supports_vision: model.supports_images,
230 supports_thinking: model.supports_thinking,
231 },
232 );
233 }
234
235 let mut models = models
236 .into_values()
237 .map(|model| {
238 Arc::new(OllamaLanguageModel {
239 id: LanguageModelId::from(model.name.clone()),
240 model: model.clone(),
241 http_client: self.http_client.clone(),
242 request_limiter: RateLimiter::new(4),
243 }) as Arc<dyn LanguageModel>
244 })
245 .collect::<Vec<_>>();
246 models.sort_by_key(|model| model.name());
247 models
248 }
249
250 fn is_authenticated(&self, cx: &App) -> bool {
251 self.state.read(cx).is_authenticated()
252 }
253
254 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
255 self.state.update(cx, |state, cx| state.authenticate(cx))
256 }
257
258 fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView {
259 let state = self.state.clone();
260 cx.new(|cx| ConfigurationView::new(state, window, cx))
261 .into()
262 }
263
264 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
265 self.state.update(cx, |state, cx| state.fetch_models(cx))
266 }
267}
268
269pub struct OllamaLanguageModel {
270 id: LanguageModelId,
271 model: ollama::Model,
272 http_client: Arc<dyn HttpClient>,
273 request_limiter: RateLimiter,
274}
275
276impl OllamaLanguageModel {
277 fn to_ollama_request(&self, request: LanguageModelRequest) -> ChatRequest {
278 let supports_vision = self.model.supports_vision.unwrap_or(false);
279
280 ChatRequest {
281 model: self.model.name.clone(),
282 messages: request
283 .messages
284 .into_iter()
285 .map(|msg| {
286 let images = if supports_vision {
287 msg.content
288 .iter()
289 .filter_map(|content| match content {
290 MessageContent::Image(image) => Some(image.source.to_string()),
291 _ => None,
292 })
293 .collect::<Vec<String>>()
294 } else {
295 vec![]
296 };
297
298 match msg.role {
299 Role::User => ChatMessage::User {
300 content: msg.string_contents(),
301 images: if images.is_empty() {
302 None
303 } else {
304 Some(images)
305 },
306 },
307 Role::Assistant => {
308 let content = msg.string_contents();
309 let thinking =
310 msg.content.into_iter().find_map(|content| match content {
311 MessageContent::Thinking { text, .. } if !text.is_empty() => {
312 Some(text)
313 }
314 _ => None,
315 });
316 ChatMessage::Assistant {
317 content,
318 tool_calls: None,
319 images: if images.is_empty() {
320 None
321 } else {
322 Some(images)
323 },
324 thinking,
325 }
326 }
327 Role::System => ChatMessage::System {
328 content: msg.string_contents(),
329 },
330 }
331 })
332 .collect(),
333 keep_alive: self.model.keep_alive.clone().unwrap_or_default(),
334 stream: true,
335 options: Some(ChatOptions {
336 num_ctx: Some(self.model.max_tokens),
337 stop: Some(request.stop),
338 temperature: request.temperature.or(Some(1.0)),
339 ..Default::default()
340 }),
341 think: self
342 .model
343 .supports_thinking
344 .map(|supports_thinking| supports_thinking && request.thinking_allowed),
345 tools: request.tools.into_iter().map(tool_into_ollama).collect(),
346 }
347 }
348}
349
350impl LanguageModel for OllamaLanguageModel {
351 fn id(&self) -> LanguageModelId {
352 self.id.clone()
353 }
354
355 fn name(&self) -> LanguageModelName {
356 LanguageModelName::from(self.model.display_name().to_string())
357 }
358
359 fn provider_id(&self) -> LanguageModelProviderId {
360 PROVIDER_ID
361 }
362
363 fn provider_name(&self) -> LanguageModelProviderName {
364 PROVIDER_NAME
365 }
366
367 fn supports_tools(&self) -> bool {
368 self.model.supports_tools.unwrap_or(false)
369 }
370
371 fn supports_images(&self) -> bool {
372 self.model.supports_vision.unwrap_or(false)
373 }
374
375 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
376 match choice {
377 LanguageModelToolChoice::Auto => false,
378 LanguageModelToolChoice::Any => false,
379 LanguageModelToolChoice::None => false,
380 }
381 }
382
383 fn telemetry_id(&self) -> String {
384 format!("ollama/{}", self.model.id())
385 }
386
387 fn max_token_count(&self) -> u64 {
388 self.model.max_token_count()
389 }
390
391 fn count_tokens(
392 &self,
393 request: LanguageModelRequest,
394 _cx: &App,
395 ) -> BoxFuture<'static, Result<u64>> {
396 // There is no endpoint for this _yet_ in Ollama
397 // see: https://github.com/ollama/ollama/issues/1716 and https://github.com/ollama/ollama/issues/3582
398 let token_count = request
399 .messages
400 .iter()
401 .map(|msg| msg.string_contents().chars().count())
402 .sum::<usize>()
403 / 4;
404
405 async move { Ok(token_count as u64) }.boxed()
406 }
407
408 fn stream_completion(
409 &self,
410 request: LanguageModelRequest,
411 cx: &AsyncApp,
412 ) -> BoxFuture<
413 'static,
414 Result<
415 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
416 LanguageModelCompletionError,
417 >,
418 > {
419 let request = self.to_ollama_request(request);
420
421 let http_client = self.http_client.clone();
422 let Ok(api_url) = cx.update(|cx| {
423 let settings = &AllLanguageModelSettings::get_global(cx).ollama;
424 settings.api_url.clone()
425 }) else {
426 return futures::future::ready(Err(anyhow!("App state dropped").into())).boxed();
427 };
428
429 let future = self.request_limiter.stream(async move {
430 let stream = stream_chat_completion(http_client.as_ref(), &api_url, request).await?;
431 let stream = map_to_language_model_completion_events(stream);
432 Ok(stream)
433 });
434
435 future.map_ok(|f| f.boxed()).boxed()
436 }
437}
438
439fn map_to_language_model_completion_events(
440 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
441) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
442 // Used for creating unique tool use ids
443 static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
444
445 struct State {
446 stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
447 used_tools: bool,
448 }
449
450 // We need to create a ToolUse and Stop event from a single
451 // response from the original stream
452 let stream = stream::unfold(
453 State {
454 stream,
455 used_tools: false,
456 },
457 async move |mut state| {
458 let response = state.stream.next().await?;
459
460 let delta = match response {
461 Ok(delta) => delta,
462 Err(e) => {
463 let event = Err(LanguageModelCompletionError::from(anyhow!(e)));
464 return Some((vec![event], state));
465 }
466 };
467
468 let mut events = Vec::new();
469
470 match delta.message {
471 ChatMessage::User { content, images: _ } => {
472 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
473 }
474 ChatMessage::System { content } => {
475 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
476 }
477 ChatMessage::Assistant {
478 content,
479 tool_calls,
480 images: _,
481 thinking,
482 } => {
483 if let Some(text) = thinking {
484 events.push(Ok(LanguageModelCompletionEvent::Thinking {
485 text,
486 signature: None,
487 }));
488 }
489
490 if let Some(tool_call) = tool_calls.and_then(|v| v.into_iter().next()) {
491 match tool_call {
492 OllamaToolCall::Function(function) => {
493 let tool_id = format!(
494 "{}-{}",
495 &function.name,
496 TOOL_CALL_COUNTER.fetch_add(1, Ordering::Relaxed)
497 );
498 let event =
499 LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
500 id: LanguageModelToolUseId::from(tool_id),
501 name: Arc::from(function.name),
502 raw_input: function.arguments.to_string(),
503 input: function.arguments,
504 is_input_complete: true,
505 });
506 events.push(Ok(event));
507 state.used_tools = true;
508 }
509 }
510 } else if !content.is_empty() {
511 events.push(Ok(LanguageModelCompletionEvent::Text(content)));
512 }
513 }
514 };
515
516 if delta.done {
517 events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
518 input_tokens: delta.prompt_eval_count.unwrap_or(0),
519 output_tokens: delta.eval_count.unwrap_or(0),
520 cache_creation_input_tokens: 0,
521 cache_read_input_tokens: 0,
522 })));
523 if state.used_tools {
524 state.used_tools = false;
525 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
526 } else {
527 events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
528 }
529 }
530
531 Some((events, state))
532 },
533 );
534
535 stream.flat_map(futures::stream::iter)
536}
537
538struct ConfigurationView {
539 state: gpui::Entity<State>,
540 loading_models_task: Option<Task<()>>,
541}
542
543impl ConfigurationView {
544 pub fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
545 let loading_models_task = Some(cx.spawn_in(window, {
546 let state = state.clone();
547 async move |this, cx| {
548 if let Some(task) = state
549 .update(cx, |state, cx| state.authenticate(cx))
550 .log_err()
551 {
552 task.await.log_err();
553 }
554 this.update(cx, |this, cx| {
555 this.loading_models_task = None;
556 cx.notify();
557 })
558 .log_err();
559 }
560 }));
561
562 Self {
563 state,
564 loading_models_task,
565 }
566 }
567
568 fn retry_connection(&self, cx: &mut App) {
569 self.state
570 .update(cx, |state, cx| state.fetch_models(cx))
571 .detach_and_log_err(cx);
572 }
573}
574
575impl Render for ConfigurationView {
576 fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
577 let is_authenticated = self.state.read(cx).is_authenticated();
578
579 let ollama_intro =
580 "Get up & running with Llama 3.3, Mistral, Gemma 2, and other LLMs with Ollama.";
581
582 if self.loading_models_task.is_some() {
583 div().child(Label::new("Loading models...")).into_any()
584 } else {
585 v_flex()
586 .gap_2()
587 .child(
588 v_flex().gap_1().child(Label::new(ollama_intro)).child(
589 List::new()
590 .child(InstructionListItem::text_only("Ollama must be running with at least one model installed to use it in the assistant."))
591 .child(InstructionListItem::text_only(
592 "Once installed, try `ollama run llama3.2`",
593 )),
594 ),
595 )
596 .child(
597 h_flex()
598 .w_full()
599 .justify_between()
600 .gap_2()
601 .child(
602 h_flex()
603 .w_full()
604 .gap_2()
605 .map(|this| {
606 if is_authenticated {
607 this.child(
608 Button::new("ollama-site", "Ollama")
609 .style(ButtonStyle::Subtle)
610 .icon(IconName::ArrowUpRight)
611 .icon_size(IconSize::Small)
612 .icon_color(Color::Muted)
613 .on_click(move |_, _, cx| cx.open_url(OLLAMA_SITE))
614 .into_any_element(),
615 )
616 } else {
617 this.child(
618 Button::new(
619 "download_ollama_button",
620 "Download Ollama",
621 )
622 .style(ButtonStyle::Subtle)
623 .icon(IconName::ArrowUpRight)
624 .icon_size(IconSize::Small)
625 .icon_color(Color::Muted)
626 .on_click(move |_, _, cx| {
627 cx.open_url(OLLAMA_DOWNLOAD_URL)
628 })
629 .into_any_element(),
630 )
631 }
632 })
633 .child(
634 Button::new("view-models", "View All Models")
635 .style(ButtonStyle::Subtle)
636 .icon(IconName::ArrowUpRight)
637 .icon_size(IconSize::Small)
638 .icon_color(Color::Muted)
639 .on_click(move |_, _, cx| cx.open_url(OLLAMA_LIBRARY_URL)),
640 ),
641 )
642 .map(|this| {
643 if is_authenticated {
644 this.child(
645 ButtonLike::new("connected")
646 .disabled(true)
647 .cursor_style(gpui::CursorStyle::Arrow)
648 .child(
649 h_flex()
650 .gap_2()
651 .child(Indicator::dot().color(Color::Success))
652 .child(Label::new("Connected"))
653 .into_any_element(),
654 ),
655 )
656 } else {
657 this.child(
658 Button::new("retry_ollama_models", "Connect")
659 .icon_position(IconPosition::Start)
660 .icon_size(IconSize::XSmall)
661 .icon(IconName::PlayFilled)
662 .on_click(cx.listener(move |this, _, _, cx| {
663 this.retry_connection(cx)
664 })),
665 )
666 }
667 })
668 )
669 .into_any()
670 }
671 }
672}
673
674fn tool_into_ollama(tool: LanguageModelRequestTool) -> ollama::OllamaTool {
675 ollama::OllamaTool::Function {
676 function: OllamaFunctionTool {
677 name: tool.name,
678 description: Some(tool.description),
679 parameters: Some(tool.input_schema),
680 },
681 }
682}