ollama.rs

  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, cx: &App) -> Option<Arc<dyn LanguageModel>> {
196        self.provided_models(cx).into_iter().next()
197    }
198
199    fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>> {
200        self.default_model(cx)
201    }
202
203    fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>> {
204        let mut models: HashMap<String, ollama::Model> = HashMap::new();
205
206        // Add models from the Ollama API
207        for model in self.state.read(cx).available_models.iter() {
208            models.insert(model.name.clone(), model.clone());
209        }
210
211        // Override with available models from settings
212        for model in AllLanguageModelSettings::get_global(cx)
213            .ollama
214            .available_models
215            .iter()
216        {
217            models.insert(
218                model.name.clone(),
219                ollama::Model {
220                    name: model.name.clone(),
221                    display_name: model.display_name.clone(),
222                    max_tokens: model.max_tokens,
223                    keep_alive: model.keep_alive.clone(),
224                    supports_tools: model.supports_tools,
225                    supports_vision: model.supports_images,
226                    supports_thinking: model.supports_thinking,
227                },
228            );
229        }
230
231        let mut models = models
232            .into_values()
233            .map(|model| {
234                Arc::new(OllamaLanguageModel {
235                    id: LanguageModelId::from(model.name.clone()),
236                    model: model.clone(),
237                    http_client: self.http_client.clone(),
238                    request_limiter: RateLimiter::new(4),
239                }) as Arc<dyn LanguageModel>
240            })
241            .collect::<Vec<_>>();
242        models.sort_by_key(|model| model.name());
243        models
244    }
245
246    fn is_authenticated(&self, cx: &App) -> bool {
247        self.state.read(cx).is_authenticated()
248    }
249
250    fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>> {
251        self.state.update(cx, |state, cx| state.authenticate(cx))
252    }
253
254    fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView {
255        let state = self.state.clone();
256        cx.new(|cx| ConfigurationView::new(state, window, cx))
257            .into()
258    }
259
260    fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>> {
261        self.state.update(cx, |state, cx| state.fetch_models(cx))
262    }
263}
264
265pub struct OllamaLanguageModel {
266    id: LanguageModelId,
267    model: ollama::Model,
268    http_client: Arc<dyn HttpClient>,
269    request_limiter: RateLimiter,
270}
271
272impl OllamaLanguageModel {
273    fn to_ollama_request(&self, request: LanguageModelRequest) -> ChatRequest {
274        let supports_vision = self.model.supports_vision.unwrap_or(false);
275
276        ChatRequest {
277            model: self.model.name.clone(),
278            messages: request
279                .messages
280                .into_iter()
281                .map(|msg| {
282                    let images = if supports_vision {
283                        msg.content
284                            .iter()
285                            .filter_map(|content| match content {
286                                MessageContent::Image(image) => Some(image.source.to_string()),
287                                _ => None,
288                            })
289                            .collect::<Vec<String>>()
290                    } else {
291                        vec![]
292                    };
293
294                    match msg.role {
295                        Role::User => ChatMessage::User {
296                            content: msg.string_contents(),
297                            images: if images.is_empty() {
298                                None
299                            } else {
300                                Some(images)
301                            },
302                        },
303                        Role::Assistant => {
304                            let content = msg.string_contents();
305                            let thinking =
306                                msg.content.into_iter().find_map(|content| match content {
307                                    MessageContent::Thinking { text, .. } if !text.is_empty() => {
308                                        Some(text)
309                                    }
310                                    _ => None,
311                                });
312                            ChatMessage::Assistant {
313                                content,
314                                tool_calls: None,
315                                images: if images.is_empty() {
316                                    None
317                                } else {
318                                    Some(images)
319                                },
320                                thinking,
321                            }
322                        }
323                        Role::System => ChatMessage::System {
324                            content: msg.string_contents(),
325                        },
326                    }
327                })
328                .collect(),
329            keep_alive: self.model.keep_alive.clone().unwrap_or_default(),
330            stream: true,
331            options: Some(ChatOptions {
332                num_ctx: Some(self.model.max_tokens),
333                stop: Some(request.stop),
334                temperature: request.temperature.or(Some(1.0)),
335                ..Default::default()
336            }),
337            think: self
338                .model
339                .supports_thinking
340                .map(|supports_thinking| supports_thinking && request.thinking_allowed),
341            tools: request.tools.into_iter().map(tool_into_ollama).collect(),
342        }
343    }
344}
345
346impl LanguageModel for OllamaLanguageModel {
347    fn id(&self) -> LanguageModelId {
348        self.id.clone()
349    }
350
351    fn name(&self) -> LanguageModelName {
352        LanguageModelName::from(self.model.display_name().to_string())
353    }
354
355    fn provider_id(&self) -> LanguageModelProviderId {
356        PROVIDER_ID
357    }
358
359    fn provider_name(&self) -> LanguageModelProviderName {
360        PROVIDER_NAME
361    }
362
363    fn supports_tools(&self) -> bool {
364        self.model.supports_tools.unwrap_or(false)
365    }
366
367    fn supports_images(&self) -> bool {
368        self.model.supports_vision.unwrap_or(false)
369    }
370
371    fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool {
372        match choice {
373            LanguageModelToolChoice::Auto => false,
374            LanguageModelToolChoice::Any => false,
375            LanguageModelToolChoice::None => false,
376        }
377    }
378
379    fn telemetry_id(&self) -> String {
380        format!("ollama/{}", self.model.id())
381    }
382
383    fn max_token_count(&self) -> u64 {
384        self.model.max_token_count()
385    }
386
387    fn count_tokens(
388        &self,
389        request: LanguageModelRequest,
390        _cx: &App,
391    ) -> BoxFuture<'static, Result<u64>> {
392        // There is no endpoint for this _yet_ in Ollama
393        // see: https://github.com/ollama/ollama/issues/1716 and https://github.com/ollama/ollama/issues/3582
394        let token_count = request
395            .messages
396            .iter()
397            .map(|msg| msg.string_contents().chars().count())
398            .sum::<usize>()
399            / 4;
400
401        async move { Ok(token_count as u64) }.boxed()
402    }
403
404    fn stream_completion(
405        &self,
406        request: LanguageModelRequest,
407        cx: &AsyncApp,
408    ) -> BoxFuture<
409        'static,
410        Result<
411            BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
412            LanguageModelCompletionError,
413        >,
414    > {
415        let request = self.to_ollama_request(request);
416
417        let http_client = self.http_client.clone();
418        let Ok(api_url) = cx.update(|cx| {
419            let settings = &AllLanguageModelSettings::get_global(cx).ollama;
420            settings.api_url.clone()
421        }) else {
422            return futures::future::ready(Err(anyhow!("App state dropped").into())).boxed();
423        };
424
425        let future = self.request_limiter.stream(async move {
426            let stream = stream_chat_completion(http_client.as_ref(), &api_url, request).await?;
427            let stream = map_to_language_model_completion_events(stream);
428            Ok(stream)
429        });
430
431        future.map_ok(|f| f.boxed()).boxed()
432    }
433}
434
435fn map_to_language_model_completion_events(
436    stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
437) -> impl Stream<Item = Result<LanguageModelCompletionEvent, LanguageModelCompletionError>> {
438    // Used for creating unique tool use ids
439    static TOOL_CALL_COUNTER: AtomicU64 = AtomicU64::new(0);
440
441    struct State {
442        stream: Pin<Box<dyn Stream<Item = anyhow::Result<ChatResponseDelta>> + Send>>,
443        used_tools: bool,
444    }
445
446    // We need to create a ToolUse and Stop event from a single
447    // response from the original stream
448    let stream = stream::unfold(
449        State {
450            stream,
451            used_tools: false,
452        },
453        async move |mut state| {
454            let response = state.stream.next().await?;
455
456            let delta = match response {
457                Ok(delta) => delta,
458                Err(e) => {
459                    let event = Err(LanguageModelCompletionError::from(anyhow!(e)));
460                    return Some((vec![event], state));
461                }
462            };
463
464            let mut events = Vec::new();
465
466            match delta.message {
467                ChatMessage::User { content, images: _ } => {
468                    events.push(Ok(LanguageModelCompletionEvent::Text(content)));
469                }
470                ChatMessage::System { content } => {
471                    events.push(Ok(LanguageModelCompletionEvent::Text(content)));
472                }
473                ChatMessage::Assistant {
474                    content,
475                    tool_calls,
476                    images: _,
477                    thinking,
478                } => {
479                    if let Some(text) = thinking {
480                        events.push(Ok(LanguageModelCompletionEvent::Thinking {
481                            text,
482                            signature: None,
483                        }));
484                    }
485
486                    if let Some(tool_call) = tool_calls.and_then(|v| v.into_iter().next()) {
487                        match tool_call {
488                            OllamaToolCall::Function(function) => {
489                                let tool_id = format!(
490                                    "{}-{}",
491                                    &function.name,
492                                    TOOL_CALL_COUNTER.fetch_add(1, Ordering::Relaxed)
493                                );
494                                let event =
495                                    LanguageModelCompletionEvent::ToolUse(LanguageModelToolUse {
496                                        id: LanguageModelToolUseId::from(tool_id),
497                                        name: Arc::from(function.name),
498                                        raw_input: function.arguments.to_string(),
499                                        input: function.arguments,
500                                        is_input_complete: true,
501                                    });
502                                events.push(Ok(event));
503                                state.used_tools = true;
504                            }
505                        }
506                    } else if !content.is_empty() {
507                        events.push(Ok(LanguageModelCompletionEvent::Text(content)));
508                    }
509                }
510            };
511
512            if delta.done {
513                events.push(Ok(LanguageModelCompletionEvent::UsageUpdate(TokenUsage {
514                    input_tokens: delta.prompt_eval_count.unwrap_or(0),
515                    output_tokens: delta.eval_count.unwrap_or(0),
516                    cache_creation_input_tokens: 0,
517                    cache_read_input_tokens: 0,
518                })));
519                if state.used_tools {
520                    state.used_tools = false;
521                    events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::ToolUse)));
522                } else {
523                    events.push(Ok(LanguageModelCompletionEvent::Stop(StopReason::EndTurn)));
524                }
525            }
526
527            Some((events, state))
528        },
529    );
530
531    stream.flat_map(futures::stream::iter)
532}
533
534struct ConfigurationView {
535    state: gpui::Entity<State>,
536    loading_models_task: Option<Task<()>>,
537}
538
539impl ConfigurationView {
540    pub fn new(state: gpui::Entity<State>, window: &mut Window, cx: &mut Context<Self>) -> Self {
541        let loading_models_task = Some(cx.spawn_in(window, {
542            let state = state.clone();
543            async move |this, cx| {
544                if let Some(task) = state
545                    .update(cx, |state, cx| state.authenticate(cx))
546                    .log_err()
547                {
548                    task.await.log_err();
549                }
550                this.update(cx, |this, cx| {
551                    this.loading_models_task = None;
552                    cx.notify();
553                })
554                .log_err();
555            }
556        }));
557
558        Self {
559            state,
560            loading_models_task,
561        }
562    }
563
564    fn retry_connection(&self, cx: &mut App) {
565        self.state
566            .update(cx, |state, cx| state.fetch_models(cx))
567            .detach_and_log_err(cx);
568    }
569}
570
571impl Render for ConfigurationView {
572    fn render(&mut self, _: &mut Window, cx: &mut Context<Self>) -> impl IntoElement {
573        let is_authenticated = self.state.read(cx).is_authenticated();
574
575        let ollama_intro =
576            "Get up & running with Llama 3.3, Mistral, Gemma 2, and other LLMs with Ollama.";
577
578        if self.loading_models_task.is_some() {
579            div().child(Label::new("Loading models...")).into_any()
580        } else {
581            v_flex()
582                .gap_2()
583                .child(
584                    v_flex().gap_1().child(Label::new(ollama_intro)).child(
585                        List::new()
586                            .child(InstructionListItem::text_only("Ollama must be running with at least one model installed to use it in the assistant."))
587                            .child(InstructionListItem::text_only(
588                                "Once installed, try `ollama run llama3.2`",
589                            )),
590                    ),
591                )
592                .child(
593                    h_flex()
594                        .w_full()
595                        .justify_between()
596                        .gap_2()
597                        .child(
598                            h_flex()
599                                .w_full()
600                                .gap_2()
601                                .map(|this| {
602                                    if is_authenticated {
603                                        this.child(
604                                            Button::new("ollama-site", "Ollama")
605                                                .style(ButtonStyle::Subtle)
606                                                .icon(IconName::ArrowUpRight)
607                                                .icon_size(IconSize::XSmall)
608                                                .icon_color(Color::Muted)
609                                                .on_click(move |_, _, cx| cx.open_url(OLLAMA_SITE))
610                                                .into_any_element(),
611                                        )
612                                    } else {
613                                        this.child(
614                                            Button::new(
615                                                "download_ollama_button",
616                                                "Download Ollama",
617                                            )
618                                            .style(ButtonStyle::Subtle)
619                                            .icon(IconName::ArrowUpRight)
620                                            .icon_size(IconSize::XSmall)
621                                            .icon_color(Color::Muted)
622                                            .on_click(move |_, _, cx| {
623                                                cx.open_url(OLLAMA_DOWNLOAD_URL)
624                                            })
625                                            .into_any_element(),
626                                        )
627                                    }
628                                })
629                                .child(
630                                    Button::new("view-models", "All Models")
631                                        .style(ButtonStyle::Subtle)
632                                        .icon(IconName::ArrowUpRight)
633                                        .icon_size(IconSize::XSmall)
634                                        .icon_color(Color::Muted)
635                                        .on_click(move |_, _, cx| cx.open_url(OLLAMA_LIBRARY_URL)),
636                                ),
637                        )
638                        .map(|this| {
639                            if is_authenticated {
640                                this.child(
641                                    ButtonLike::new("connected")
642                                        .disabled(true)
643                                        .cursor_style(gpui::CursorStyle::Arrow)
644                                        .child(
645                                            h_flex()
646                                                .gap_2()
647                                                .child(Indicator::dot().color(Color::Success))
648                                                .child(Label::new("Connected"))
649                                                .into_any_element(),
650                                        ),
651                                )
652                            } else {
653                                this.child(
654                                    Button::new("retry_ollama_models", "Connect")
655                                        .icon_position(IconPosition::Start)
656                                        .icon_size(IconSize::XSmall)
657                                        .icon(IconName::Play)
658                                        .on_click(cx.listener(move |this, _, _, cx| {
659                                            this.retry_connection(cx)
660                                        })),
661                                )
662                            }
663                        })
664                )
665                .into_any()
666        }
667    }
668}
669
670fn tool_into_ollama(tool: LanguageModelRequestTool) -> ollama::OllamaTool {
671    ollama::OllamaTool::Function {
672        function: OllamaFunctionTool {
673            name: tool.name,
674            description: Some(tool.description),
675            parameters: Some(tool.input_schema),
676        },
677    }
678}