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, anyhow};
12use client::Client;
13use futures::FutureExt;
14use futures::{StreamExt, future::BoxFuture, stream::BoxStream};
15use gpui::{AnyElement, AnyView, App, AsyncApp, SharedString, Task, Window};
16use http_client::http::{HeaderMap, HeaderValue};
17use icons::IconName;
18use parking_lot::Mutex;
19use proto::Plan;
20use schemars::JsonSchema;
21use serde::{Deserialize, Serialize, de::DeserializeOwned};
22use std::fmt;
23use std::ops::{Add, Sub};
24use std::str::FromStr as _;
25use std::sync::Arc;
26use thiserror::Error;
27use util::serde::is_default;
28use zed_llm_client::{
29 MODEL_REQUESTS_USAGE_AMOUNT_HEADER_NAME, MODEL_REQUESTS_USAGE_LIMIT_HEADER_NAME, UsageLimit,
30};
31
32pub use crate::model::*;
33pub use crate::rate_limiter::*;
34pub use crate::registry::*;
35pub use crate::request::*;
36pub use crate::role::*;
37pub use crate::telemetry::*;
38
39pub const ZED_CLOUD_PROVIDER_ID: &str = "zed.dev";
40
41pub fn init(client: Arc<Client>, cx: &mut App) {
42 registry::init(cx);
43 RefreshLlmTokenListener::register(client.clone(), cx);
44}
45
46/// The availability of a [`LanguageModel`].
47#[derive(Debug, PartialEq, Eq, Clone, Copy)]
48pub enum LanguageModelAvailability {
49 /// The language model is available to the general public.
50 Public,
51 /// The language model is available to users on the indicated plan.
52 RequiresPlan(Plan),
53}
54
55/// Configuration for caching language model messages.
56#[derive(Clone, Debug, PartialEq, Serialize, Deserialize, JsonSchema)]
57pub struct LanguageModelCacheConfiguration {
58 pub max_cache_anchors: usize,
59 pub should_speculate: bool,
60 pub min_total_token: usize,
61}
62
63/// A completion event from a language model.
64#[derive(Debug, PartialEq, Clone, Serialize, Deserialize)]
65pub enum LanguageModelCompletionEvent {
66 Stop(StopReason),
67 Text(String),
68 Thinking {
69 text: String,
70 signature: Option<String>,
71 },
72 ToolUse(LanguageModelToolUse),
73 StartMessage {
74 message_id: String,
75 },
76 UsageUpdate(TokenUsage),
77}
78
79/// Indicates the format used to define the input schema for a language model tool.
80#[derive(Debug, PartialEq, Eq, Clone, Copy, Hash)]
81pub enum LanguageModelToolSchemaFormat {
82 /// A JSON schema, see https://json-schema.org
83 JsonSchema,
84 /// A subset of an OpenAPI 3.0 schema object supported by Google AI, see https://ai.google.dev/api/caching#Schema
85 JsonSchemaSubset,
86}
87
88#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize)]
89#[serde(rename_all = "snake_case")]
90pub enum StopReason {
91 EndTurn,
92 MaxTokens,
93 ToolUse,
94}
95
96#[derive(Debug, Clone, Copy)]
97pub struct RequestUsage {
98 pub limit: UsageLimit,
99 pub amount: i32,
100}
101
102impl RequestUsage {
103 pub fn from_headers(headers: &HeaderMap<HeaderValue>) -> Result<Self> {
104 let limit = headers
105 .get(MODEL_REQUESTS_USAGE_LIMIT_HEADER_NAME)
106 .ok_or_else(|| anyhow!("missing {MODEL_REQUESTS_USAGE_LIMIT_HEADER_NAME:?} header"))?;
107 let limit = UsageLimit::from_str(limit.to_str()?)?;
108
109 let amount = headers
110 .get(MODEL_REQUESTS_USAGE_AMOUNT_HEADER_NAME)
111 .ok_or_else(|| anyhow!("missing {MODEL_REQUESTS_USAGE_AMOUNT_HEADER_NAME:?} header"))?;
112 let amount = amount.to_str()?.parse::<i32>()?;
113
114 Ok(Self { limit, amount })
115 }
116}
117
118#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize, Default)]
119pub struct TokenUsage {
120 #[serde(default, skip_serializing_if = "is_default")]
121 pub input_tokens: u32,
122 #[serde(default, skip_serializing_if = "is_default")]
123 pub output_tokens: u32,
124 #[serde(default, skip_serializing_if = "is_default")]
125 pub cache_creation_input_tokens: u32,
126 #[serde(default, skip_serializing_if = "is_default")]
127 pub cache_read_input_tokens: u32,
128}
129
130impl TokenUsage {
131 pub fn total_tokens(&self) -> u32 {
132 self.input_tokens
133 + self.output_tokens
134 + self.cache_read_input_tokens
135 + self.cache_creation_input_tokens
136 }
137}
138
139impl Add<TokenUsage> for TokenUsage {
140 type Output = Self;
141
142 fn add(self, other: Self) -> Self {
143 Self {
144 input_tokens: self.input_tokens + other.input_tokens,
145 output_tokens: self.output_tokens + other.output_tokens,
146 cache_creation_input_tokens: self.cache_creation_input_tokens
147 + other.cache_creation_input_tokens,
148 cache_read_input_tokens: self.cache_read_input_tokens + other.cache_read_input_tokens,
149 }
150 }
151}
152
153impl Sub<TokenUsage> for TokenUsage {
154 type Output = Self;
155
156 fn sub(self, other: Self) -> Self {
157 Self {
158 input_tokens: self.input_tokens - other.input_tokens,
159 output_tokens: self.output_tokens - other.output_tokens,
160 cache_creation_input_tokens: self.cache_creation_input_tokens
161 - other.cache_creation_input_tokens,
162 cache_read_input_tokens: self.cache_read_input_tokens - other.cache_read_input_tokens,
163 }
164 }
165}
166
167#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
168pub struct LanguageModelToolUseId(Arc<str>);
169
170impl fmt::Display for LanguageModelToolUseId {
171 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
172 write!(f, "{}", self.0)
173 }
174}
175
176impl<T> From<T> for LanguageModelToolUseId
177where
178 T: Into<Arc<str>>,
179{
180 fn from(value: T) -> Self {
181 Self(value.into())
182 }
183}
184
185#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
186pub struct LanguageModelToolUse {
187 pub id: LanguageModelToolUseId,
188 pub name: Arc<str>,
189 pub input: serde_json::Value,
190 pub is_input_complete: bool,
191}
192
193pub struct LanguageModelTextStream {
194 pub message_id: Option<String>,
195 pub stream: BoxStream<'static, Result<String>>,
196 // Has complete token usage after the stream has finished
197 pub last_token_usage: Arc<Mutex<TokenUsage>>,
198}
199
200impl Default for LanguageModelTextStream {
201 fn default() -> Self {
202 Self {
203 message_id: None,
204 stream: Box::pin(futures::stream::empty()),
205 last_token_usage: Arc::new(Mutex::new(TokenUsage::default())),
206 }
207 }
208}
209
210pub trait LanguageModel: Send + Sync {
211 fn id(&self) -> LanguageModelId;
212 fn name(&self) -> LanguageModelName;
213 fn provider_id(&self) -> LanguageModelProviderId;
214 fn provider_name(&self) -> LanguageModelProviderName;
215 fn telemetry_id(&self) -> String;
216
217 fn api_key(&self, _cx: &App) -> Option<String> {
218 None
219 }
220
221 /// Returns the availability of this language model.
222 fn availability(&self) -> LanguageModelAvailability {
223 LanguageModelAvailability::Public
224 }
225
226 /// Whether this model supports tools.
227 fn supports_tools(&self) -> bool;
228
229 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
230 LanguageModelToolSchemaFormat::JsonSchema
231 }
232
233 fn max_token_count(&self) -> usize;
234 fn max_output_tokens(&self) -> Option<u32> {
235 None
236 }
237
238 fn count_tokens(
239 &self,
240 request: LanguageModelRequest,
241 cx: &App,
242 ) -> BoxFuture<'static, Result<usize>>;
243
244 fn stream_completion(
245 &self,
246 request: LanguageModelRequest,
247 cx: &AsyncApp,
248 ) -> BoxFuture<'static, Result<BoxStream<'static, Result<LanguageModelCompletionEvent>>>>;
249
250 fn stream_completion_with_usage(
251 &self,
252 request: LanguageModelRequest,
253 cx: &AsyncApp,
254 ) -> BoxFuture<
255 'static,
256 Result<(
257 BoxStream<'static, Result<LanguageModelCompletionEvent>>,
258 Option<RequestUsage>,
259 )>,
260 > {
261 self.stream_completion(request, cx)
262 .map(|result| result.map(|stream| (stream, None)))
263 .boxed()
264 }
265
266 fn stream_completion_text(
267 &self,
268 request: LanguageModelRequest,
269 cx: &AsyncApp,
270 ) -> BoxFuture<'static, Result<LanguageModelTextStream>> {
271 self.stream_completion_text_with_usage(request, cx)
272 .map(|result| result.map(|(stream, _usage)| stream))
273 .boxed()
274 }
275
276 fn stream_completion_text_with_usage(
277 &self,
278 request: LanguageModelRequest,
279 cx: &AsyncApp,
280 ) -> BoxFuture<'static, Result<(LanguageModelTextStream, Option<RequestUsage>)>> {
281 let future = self.stream_completion_with_usage(request, cx);
282
283 async move {
284 let (events, usage) = future.await?;
285 let mut events = events.fuse();
286 let mut message_id = None;
287 let mut first_item_text = None;
288 let last_token_usage = Arc::new(Mutex::new(TokenUsage::default()));
289
290 if let Some(first_event) = events.next().await {
291 match first_event {
292 Ok(LanguageModelCompletionEvent::StartMessage { message_id: id }) => {
293 message_id = Some(id.clone());
294 }
295 Ok(LanguageModelCompletionEvent::Text(text)) => {
296 first_item_text = Some(text);
297 }
298 _ => (),
299 }
300 }
301
302 let stream = futures::stream::iter(first_item_text.map(Ok))
303 .chain(events.filter_map({
304 let last_token_usage = last_token_usage.clone();
305 move |result| {
306 let last_token_usage = last_token_usage.clone();
307 async move {
308 match result {
309 Ok(LanguageModelCompletionEvent::StartMessage { .. }) => None,
310 Ok(LanguageModelCompletionEvent::Text(text)) => Some(Ok(text)),
311 Ok(LanguageModelCompletionEvent::Thinking { .. }) => None,
312 Ok(LanguageModelCompletionEvent::Stop(_)) => None,
313 Ok(LanguageModelCompletionEvent::ToolUse(_)) => None,
314 Ok(LanguageModelCompletionEvent::UsageUpdate(token_usage)) => {
315 *last_token_usage.lock() = token_usage;
316 None
317 }
318 Err(err) => Some(Err(err)),
319 }
320 }
321 }
322 }))
323 .boxed();
324
325 Ok((
326 LanguageModelTextStream {
327 message_id,
328 stream,
329 last_token_usage,
330 },
331 usage,
332 ))
333 }
334 .boxed()
335 }
336
337 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
338 None
339 }
340
341 #[cfg(any(test, feature = "test-support"))]
342 fn as_fake(&self) -> &fake_provider::FakeLanguageModel {
343 unimplemented!()
344 }
345}
346
347#[derive(Debug, Error)]
348pub enum LanguageModelKnownError {
349 #[error("Context window limit exceeded ({tokens})")]
350 ContextWindowLimitExceeded { tokens: usize },
351}
352
353pub trait LanguageModelTool: 'static + DeserializeOwned + JsonSchema {
354 fn name() -> String;
355 fn description() -> String;
356}
357
358/// An error that occurred when trying to authenticate the language model provider.
359#[derive(Debug, Error)]
360pub enum AuthenticateError {
361 #[error("credentials not found")]
362 CredentialsNotFound,
363 #[error(transparent)]
364 Other(#[from] anyhow::Error),
365}
366
367pub trait LanguageModelProvider: 'static {
368 fn id(&self) -> LanguageModelProviderId;
369 fn name(&self) -> LanguageModelProviderName;
370 fn icon(&self) -> IconName {
371 IconName::ZedAssistant
372 }
373 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;
374 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;
375 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>>;
376 fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {
377 Vec::new()
378 }
379 fn load_model(&self, _model: Arc<dyn LanguageModel>, _cx: &App) {}
380 fn is_authenticated(&self, cx: &App) -> bool;
381 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>>;
382 fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView;
383 fn must_accept_terms(&self, _cx: &App) -> bool {
384 false
385 }
386 fn render_accept_terms(
387 &self,
388 _view: LanguageModelProviderTosView,
389 _cx: &mut App,
390 ) -> Option<AnyElement> {
391 None
392 }
393 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>>;
394}
395
396#[derive(PartialEq, Eq)]
397pub enum LanguageModelProviderTosView {
398 /// When there are some past interactions in the Agent Panel.
399 ThreadtEmptyState,
400 /// When there are no past interactions in the Agent Panel.
401 ThreadFreshStart,
402 PromptEditorPopup,
403 Configuration,
404}
405
406pub trait LanguageModelProviderState: 'static {
407 type ObservableEntity;
408
409 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>>;
410
411 fn subscribe<T: 'static>(
412 &self,
413 cx: &mut gpui::Context<T>,
414 callback: impl Fn(&mut T, &mut gpui::Context<T>) + 'static,
415 ) -> Option<gpui::Subscription> {
416 let entity = self.observable_entity()?;
417 Some(cx.observe(&entity, move |this, _, cx| {
418 callback(this, cx);
419 }))
420 }
421}
422
423#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd, Serialize, Deserialize)]
424pub struct LanguageModelId(pub SharedString);
425
426#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
427pub struct LanguageModelName(pub SharedString);
428
429#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
430pub struct LanguageModelProviderId(pub SharedString);
431
432#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
433pub struct LanguageModelProviderName(pub SharedString);
434
435impl fmt::Display for LanguageModelProviderId {
436 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
437 write!(f, "{}", self.0)
438 }
439}
440
441impl From<String> for LanguageModelId {
442 fn from(value: String) -> Self {
443 Self(SharedString::from(value))
444 }
445}
446
447impl From<String> for LanguageModelName {
448 fn from(value: String) -> Self {
449 Self(SharedString::from(value))
450 }
451}
452
453impl From<String> for LanguageModelProviderId {
454 fn from(value: String) -> Self {
455 Self(SharedString::from(value))
456 }
457}
458
459impl From<String> for LanguageModelProviderName {
460 fn from(value: String) -> Self {
461 Self(SharedString::from(value))
462 }
463}