1mod model;
2mod rate_limiter;
3mod registry;
4mod request;
5mod role;
6mod telemetry;
7
8#[cfg(any(test, feature = "test-support"))]
9pub mod fake_provider;
10
11use anyhow::Result;
12use client::Client;
13use futures::FutureExt;
14use futures::{StreamExt, future::BoxFuture, stream::BoxStream};
15use gpui::{AnyElement, AnyView, App, AsyncApp, SharedString, Task, Window};
16use icons::IconName;
17use parking_lot::Mutex;
18use proto::Plan;
19use schemars::JsonSchema;
20use serde::{Deserialize, Serialize, de::DeserializeOwned};
21use std::fmt;
22use std::ops::{Add, Sub};
23use std::sync::Arc;
24use thiserror::Error;
25use util::serde::is_default;
26
27pub use crate::model::*;
28pub use crate::rate_limiter::*;
29pub use crate::registry::*;
30pub use crate::request::*;
31pub use crate::role::*;
32pub use crate::telemetry::*;
33
34pub const ZED_CLOUD_PROVIDER_ID: &str = "zed.dev";
35
36pub fn init(client: Arc<Client>, cx: &mut App) {
37 registry::init(cx);
38 RefreshLlmTokenListener::register(client.clone(), cx);
39}
40
41/// The availability of a [`LanguageModel`].
42#[derive(Debug, PartialEq, Eq, Clone, Copy)]
43pub enum LanguageModelAvailability {
44 /// The language model is available to the general public.
45 Public,
46 /// The language model is available to users on the indicated plan.
47 RequiresPlan(Plan),
48}
49
50/// Configuration for caching language model messages.
51#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
52pub struct LanguageModelCacheConfiguration {
53 pub max_cache_anchors: usize,
54 pub should_speculate: bool,
55 pub min_total_token: usize,
56}
57
58/// A completion event from a language model.
59#[derive(Debug, PartialEq, Clone, Serialize, Deserialize)]
60pub enum LanguageModelCompletionEvent {
61 Stop(StopReason),
62 Text(String),
63 Thinking(String),
64 ToolUse(LanguageModelToolUse),
65 StartMessage { message_id: String },
66 UsageUpdate(TokenUsage),
67}
68
69/// Indicates the format used to define the input schema for a language model tool.
70#[derive(Debug, PartialEq, Eq, Clone, Copy, Hash)]
71pub enum LanguageModelToolSchemaFormat {
72 /// A JSON schema, see https://json-schema.org
73 JsonSchema,
74 /// A subset of an OpenAPI 3.0 schema object supported by Google AI, see https://ai.google.dev/api/caching#Schema
75 JsonSchemaSubset,
76}
77
78#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize)]
79#[serde(rename_all = "snake_case")]
80pub enum StopReason {
81 EndTurn,
82 MaxTokens,
83 ToolUse,
84}
85
86#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize, Default)]
87pub struct TokenUsage {
88 #[serde(default, skip_serializing_if = "is_default")]
89 pub input_tokens: u32,
90 #[serde(default, skip_serializing_if = "is_default")]
91 pub output_tokens: u32,
92 #[serde(default, skip_serializing_if = "is_default")]
93 pub cache_creation_input_tokens: u32,
94 #[serde(default, skip_serializing_if = "is_default")]
95 pub cache_read_input_tokens: u32,
96}
97
98impl TokenUsage {
99 pub fn total_tokens(&self) -> u32 {
100 self.input_tokens + self.output_tokens
101 }
102}
103
104impl Add<TokenUsage> for TokenUsage {
105 type Output = Self;
106
107 fn add(self, other: Self) -> Self {
108 Self {
109 input_tokens: self.input_tokens + other.input_tokens,
110 output_tokens: self.output_tokens + other.output_tokens,
111 cache_creation_input_tokens: self.cache_creation_input_tokens
112 + other.cache_creation_input_tokens,
113 cache_read_input_tokens: self.cache_read_input_tokens + other.cache_read_input_tokens,
114 }
115 }
116}
117
118impl Sub<TokenUsage> for TokenUsage {
119 type Output = Self;
120
121 fn sub(self, other: Self) -> Self {
122 Self {
123 input_tokens: self.input_tokens - other.input_tokens,
124 output_tokens: self.output_tokens - other.output_tokens,
125 cache_creation_input_tokens: self.cache_creation_input_tokens
126 - other.cache_creation_input_tokens,
127 cache_read_input_tokens: self.cache_read_input_tokens - other.cache_read_input_tokens,
128 }
129 }
130}
131
132#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
133pub struct LanguageModelToolUseId(Arc<str>);
134
135impl fmt::Display for LanguageModelToolUseId {
136 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
137 write!(f, "{}", self.0)
138 }
139}
140
141impl<T> From<T> for LanguageModelToolUseId
142where
143 T: Into<Arc<str>>,
144{
145 fn from(value: T) -> Self {
146 Self(value.into())
147 }
148}
149
150#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
151pub struct LanguageModelToolUse {
152 pub id: LanguageModelToolUseId,
153 pub name: Arc<str>,
154 pub input: serde_json::Value,
155}
156
157pub struct LanguageModelTextStream {
158 pub message_id: Option<String>,
159 pub stream: BoxStream<'static, Result<String>>,
160 // Has complete token usage after the stream has finished
161 pub last_token_usage: Arc<Mutex<TokenUsage>>,
162}
163
164impl Default for LanguageModelTextStream {
165 fn default() -> Self {
166 Self {
167 message_id: None,
168 stream: Box::pin(futures::stream::empty()),
169 last_token_usage: Arc::new(Mutex::new(TokenUsage::default())),
170 }
171 }
172}
173
174pub trait LanguageModel: Send + Sync {
175 fn id(&self) -> LanguageModelId;
176 fn name(&self) -> LanguageModelName;
177 fn provider_id(&self) -> LanguageModelProviderId;
178 fn provider_name(&self) -> LanguageModelProviderName;
179 fn telemetry_id(&self) -> String;
180
181 fn api_key(&self, _cx: &App) -> Option<String> {
182 None
183 }
184
185 /// Returns the availability of this language model.
186 fn availability(&self) -> LanguageModelAvailability {
187 LanguageModelAvailability::Public
188 }
189
190 /// Whether this model supports tools.
191 fn supports_tools(&self) -> bool;
192
193 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
194 LanguageModelToolSchemaFormat::JsonSchema
195 }
196
197 fn max_token_count(&self) -> usize;
198 fn max_output_tokens(&self) -> Option<u32> {
199 None
200 }
201
202 fn count_tokens(
203 &self,
204 request: LanguageModelRequest,
205 cx: &App,
206 ) -> BoxFuture<'static, Result<usize>>;
207
208 fn stream_completion(
209 &self,
210 request: LanguageModelRequest,
211 cx: &AsyncApp,
212 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>>;
213
214 fn stream_completion_text(
215 &self,
216 request: LanguageModelRequest,
217 cx: &AsyncApp,
218 ) -> BoxFuture<'static, Result<LanguageModelTextStream>> {
219 let events = self.stream_completion(request, cx);
220
221 async move {
222 let mut events = events.await?.fuse();
223 let mut message_id = None;
224 let mut first_item_text = None;
225 let last_token_usage = Arc::new(Mutex::new(TokenUsage::default()));
226
227 if let Some(first_event) = events.next().await {
228 match first_event {
229 Ok(LanguageModelCompletionEvent::StartMessage { message_id: id }) => {
230 message_id = Some(id.clone());
231 }
232 Ok(LanguageModelCompletionEvent::Text(text)) => {
233 first_item_text = Some(text);
234 }
235 _ => (),
236 }
237 }
238
239 let stream = futures::stream::iter(first_item_text.map(Ok))
240 .chain(events.filter_map({
241 let last_token_usage = last_token_usage.clone();
242 move |result| {
243 let last_token_usage = last_token_usage.clone();
244 async move {
245 match result {
246 Ok(LanguageModelCompletionEvent::StartMessage { .. }) => None,
247 Ok(LanguageModelCompletionEvent::Text(text)) => Some(Ok(text)),
248 Ok(LanguageModelCompletionEvent::Thinking(_)) => None,
249 Ok(LanguageModelCompletionEvent::Stop(_)) => None,
250 Ok(LanguageModelCompletionEvent::ToolUse(_)) => None,
251 Ok(LanguageModelCompletionEvent::UsageUpdate(token_usage)) => {
252 *last_token_usage.lock() = token_usage;
253 None
254 }
255 Err(err) => Some(Err(err)),
256 }
257 }
258 }
259 }))
260 .boxed();
261
262 Ok(LanguageModelTextStream {
263 message_id,
264 stream,
265 last_token_usage,
266 })
267 }
268 .boxed()
269 }
270
271 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
272 None
273 }
274
275 #[cfg(any(test, feature = "test-support"))]
276 fn as_fake(&self) -> &fake_provider::FakeLanguageModel {
277 unimplemented!()
278 }
279}
280
281#[derive(Debug, Error)]
282pub enum LanguageModelKnownError {
283 #[error("Context window limit exceeded ({tokens})")]
284 ContextWindowLimitExceeded { tokens: usize },
285}
286
287pub trait LanguageModelTool: 'static + DeserializeOwned + JsonSchema {
288 fn name() -> String;
289 fn description() -> String;
290}
291
292/// An error that occurred when trying to authenticate the language model provider.
293#[derive(Debug, Error)]
294pub enum AuthenticateError {
295 #[error("credentials not found")]
296 CredentialsNotFound,
297 #[error(transparent)]
298 Other(#[from] anyhow::Error),
299}
300
301pub trait LanguageModelProvider: 'static {
302 fn id(&self) -> LanguageModelProviderId;
303 fn name(&self) -> LanguageModelProviderName;
304 fn icon(&self) -> IconName {
305 IconName::ZedAssistant
306 }
307 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;
308 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>>;
309 fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {
310 Vec::new()
311 }
312 fn load_model(&self, _model: Arc<dyn LanguageModel>, _cx: &App) {}
313 fn is_authenticated(&self, cx: &App) -> bool;
314 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>>;
315 fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView;
316 fn must_accept_terms(&self, _cx: &App) -> bool {
317 false
318 }
319 fn render_accept_terms(
320 &self,
321 _view: LanguageModelProviderTosView,
322 _cx: &mut App,
323 ) -> Option<AnyElement> {
324 None
325 }
326 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>>;
327}
328
329#[derive(PartialEq, Eq)]
330pub enum LanguageModelProviderTosView {
331 /// When there are some past interactions in the Agent Panel.
332 ThreadtEmptyState,
333 /// When there are no past interactions in the Agent Panel.
334 ThreadFreshStart,
335 PromptEditorPopup,
336 Configuration,
337}
338
339pub trait LanguageModelProviderState: 'static {
340 type ObservableEntity;
341
342 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>>;
343
344 fn subscribe<T: 'static>(
345 &self,
346 cx: &mut gpui::Context<T>,
347 callback: impl Fn(&mut T, &mut gpui::Context<T>) + 'static,
348 ) -> Option<gpui::Subscription> {
349 let entity = self.observable_entity()?;
350 Some(cx.observe(&entity, move |this, _, cx| {
351 callback(this, cx);
352 }))
353 }
354}
355
356#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd, Serialize, Deserialize)]
357pub struct LanguageModelId(pub SharedString);
358
359#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
360pub struct LanguageModelName(pub SharedString);
361
362#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
363pub struct LanguageModelProviderId(pub SharedString);
364
365#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
366pub struct LanguageModelProviderName(pub SharedString);
367
368impl fmt::Display for LanguageModelProviderId {
369 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
370 write!(f, "{}", self.0)
371 }
372}
373
374impl From<String> for LanguageModelId {
375 fn from(value: String) -> Self {
376 Self(SharedString::from(value))
377 }
378}
379
380impl From<String> for LanguageModelName {
381 fn from(value: String) -> Self {
382 Self(SharedString::from(value))
383 }
384}
385
386impl From<String> for LanguageModelProviderId {
387 fn from(value: String) -> Self {
388 Self(SharedString::from(value))
389 }
390}
391
392impl From<String> for LanguageModelProviderName {
393 fn from(value: String) -> Self {
394 Self(SharedString::from(value))
395 }
396}