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::{Context as _, Result};
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 schemars::JsonSchema;
20use serde::{Deserialize, Serialize, de::DeserializeOwned};
21use std::fmt;
22use std::ops::{Add, Sub};
23use std::str::FromStr as _;
24use std::sync::Arc;
25use thiserror::Error;
26use util::serde::is_default;
27use zed_llm_client::{
28 CompletionRequestStatus, MODEL_REQUESTS_USAGE_AMOUNT_HEADER_NAME,
29 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 init_settings(cx);
43 RefreshLlmTokenListener::register(client.clone(), cx);
44}
45
46pub fn init_settings(cx: &mut App) {
47 registry::init(cx);
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 StatusUpdate(CompletionRequestStatus),
62 Stop(StopReason),
63 Text(String),
64 Thinking {
65 text: String,
66 signature: Option<String>,
67 },
68 ToolUse(LanguageModelToolUse),
69 StartMessage {
70 message_id: String,
71 },
72 UsageUpdate(TokenUsage),
73}
74
75#[derive(Error, Debug)]
76pub enum LanguageModelCompletionError {
77 #[error("received bad input JSON")]
78 BadInputJson {
79 id: LanguageModelToolUseId,
80 tool_name: Arc<str>,
81 raw_input: Arc<str>,
82 json_parse_error: String,
83 },
84 #[error(transparent)]
85 Other(#[from] anyhow::Error),
86}
87
88/// Indicates the format used to define the input schema for a language model tool.
89#[derive(Debug, PartialEq, Eq, Clone, Copy, Hash)]
90pub enum LanguageModelToolSchemaFormat {
91 /// A JSON schema, see https://json-schema.org
92 JsonSchema,
93 /// A subset of an OpenAPI 3.0 schema object supported by Google AI, see https://ai.google.dev/api/caching#Schema
94 JsonSchemaSubset,
95}
96
97#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize)]
98#[serde(rename_all = "snake_case")]
99pub enum StopReason {
100 EndTurn,
101 MaxTokens,
102 ToolUse,
103}
104
105#[derive(Debug, Clone, Copy)]
106pub struct RequestUsage {
107 pub limit: UsageLimit,
108 pub amount: i32,
109}
110
111impl RequestUsage {
112 pub fn from_headers(headers: &HeaderMap<HeaderValue>) -> Result<Self> {
113 let limit = headers
114 .get(MODEL_REQUESTS_USAGE_LIMIT_HEADER_NAME)
115 .with_context(|| {
116 format!("missing {MODEL_REQUESTS_USAGE_LIMIT_HEADER_NAME:?} header")
117 })?;
118 let limit = UsageLimit::from_str(limit.to_str()?)?;
119
120 let amount = headers
121 .get(MODEL_REQUESTS_USAGE_AMOUNT_HEADER_NAME)
122 .with_context(|| {
123 format!("missing {MODEL_REQUESTS_USAGE_AMOUNT_HEADER_NAME:?} header")
124 })?;
125 let amount = amount.to_str()?.parse::<i32>()?;
126
127 Ok(Self { limit, amount })
128 }
129}
130
131#[derive(Debug, PartialEq, Clone, Copy, Serialize, Deserialize, Default)]
132pub struct TokenUsage {
133 #[serde(default, skip_serializing_if = "is_default")]
134 pub input_tokens: u32,
135 #[serde(default, skip_serializing_if = "is_default")]
136 pub output_tokens: u32,
137 #[serde(default, skip_serializing_if = "is_default")]
138 pub cache_creation_input_tokens: u32,
139 #[serde(default, skip_serializing_if = "is_default")]
140 pub cache_read_input_tokens: u32,
141}
142
143impl TokenUsage {
144 pub fn total_tokens(&self) -> u32 {
145 self.input_tokens
146 + self.output_tokens
147 + self.cache_read_input_tokens
148 + self.cache_creation_input_tokens
149 }
150}
151
152impl Add<TokenUsage> for TokenUsage {
153 type Output = Self;
154
155 fn add(self, other: Self) -> Self {
156 Self {
157 input_tokens: self.input_tokens + other.input_tokens,
158 output_tokens: self.output_tokens + other.output_tokens,
159 cache_creation_input_tokens: self.cache_creation_input_tokens
160 + other.cache_creation_input_tokens,
161 cache_read_input_tokens: self.cache_read_input_tokens + other.cache_read_input_tokens,
162 }
163 }
164}
165
166impl Sub<TokenUsage> for TokenUsage {
167 type Output = Self;
168
169 fn sub(self, other: Self) -> Self {
170 Self {
171 input_tokens: self.input_tokens - other.input_tokens,
172 output_tokens: self.output_tokens - other.output_tokens,
173 cache_creation_input_tokens: self.cache_creation_input_tokens
174 - other.cache_creation_input_tokens,
175 cache_read_input_tokens: self.cache_read_input_tokens - other.cache_read_input_tokens,
176 }
177 }
178}
179
180#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
181pub struct LanguageModelToolUseId(Arc<str>);
182
183impl fmt::Display for LanguageModelToolUseId {
184 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
185 write!(f, "{}", self.0)
186 }
187}
188
189impl<T> From<T> for LanguageModelToolUseId
190where
191 T: Into<Arc<str>>,
192{
193 fn from(value: T) -> Self {
194 Self(value.into())
195 }
196}
197
198#[derive(Debug, PartialEq, Eq, Hash, Clone, Serialize, Deserialize)]
199pub struct LanguageModelToolUse {
200 pub id: LanguageModelToolUseId,
201 pub name: Arc<str>,
202 pub raw_input: String,
203 pub input: serde_json::Value,
204 pub is_input_complete: bool,
205}
206
207pub struct LanguageModelTextStream {
208 pub message_id: Option<String>,
209 pub stream: BoxStream<'static, Result<String, LanguageModelCompletionError>>,
210 // Has complete token usage after the stream has finished
211 pub last_token_usage: Arc<Mutex<TokenUsage>>,
212}
213
214impl Default for LanguageModelTextStream {
215 fn default() -> Self {
216 Self {
217 message_id: None,
218 stream: Box::pin(futures::stream::empty()),
219 last_token_usage: Arc::new(Mutex::new(TokenUsage::default())),
220 }
221 }
222}
223
224pub trait LanguageModel: Send + Sync {
225 fn id(&self) -> LanguageModelId;
226 fn name(&self) -> LanguageModelName;
227 fn provider_id(&self) -> LanguageModelProviderId;
228 fn provider_name(&self) -> LanguageModelProviderName;
229 fn telemetry_id(&self) -> String;
230
231 fn api_key(&self, _cx: &App) -> Option<String> {
232 None
233 }
234
235 /// Whether this model supports images
236 fn supports_images(&self) -> bool;
237
238 /// Whether this model supports tools.
239 fn supports_tools(&self) -> bool;
240
241 /// Whether this model supports choosing which tool to use.
242 fn supports_tool_choice(&self, choice: LanguageModelToolChoice) -> bool;
243
244 /// Returns whether this model supports "max mode";
245 fn supports_max_mode(&self) -> bool {
246 if self.provider_id().0 != ZED_CLOUD_PROVIDER_ID {
247 return false;
248 }
249
250 const MAX_MODE_CAPABLE_MODELS: &[CloudModel] = &[
251 CloudModel::Anthropic(anthropic::Model::ClaudeOpus4),
252 CloudModel::Anthropic(anthropic::Model::ClaudeOpus4Thinking),
253 CloudModel::Anthropic(anthropic::Model::ClaudeSonnet4),
254 CloudModel::Anthropic(anthropic::Model::ClaudeSonnet4Thinking),
255 CloudModel::Anthropic(anthropic::Model::Claude3_7Sonnet),
256 CloudModel::Anthropic(anthropic::Model::Claude3_7SonnetThinking),
257 ];
258
259 for model in MAX_MODE_CAPABLE_MODELS {
260 if self.id().0 == model.id() {
261 return true;
262 }
263 }
264
265 false
266 }
267
268 fn tool_input_format(&self) -> LanguageModelToolSchemaFormat {
269 LanguageModelToolSchemaFormat::JsonSchema
270 }
271
272 fn max_token_count(&self) -> usize;
273 fn max_output_tokens(&self) -> Option<u32> {
274 None
275 }
276
277 fn count_tokens(
278 &self,
279 request: LanguageModelRequest,
280 cx: &App,
281 ) -> BoxFuture<'static, Result<usize>>;
282
283 fn stream_completion(
284 &self,
285 request: LanguageModelRequest,
286 cx: &AsyncApp,
287 ) -> BoxFuture<
288 'static,
289 Result<
290 BoxStream<'static, Result<LanguageModelCompletionEvent, LanguageModelCompletionError>>,
291 >,
292 >;
293
294 fn stream_completion_text(
295 &self,
296 request: LanguageModelRequest,
297 cx: &AsyncApp,
298 ) -> BoxFuture<'static, Result<LanguageModelTextStream>> {
299 let future = self.stream_completion(request, cx);
300
301 async move {
302 let events = future.await?;
303 let mut events = events.fuse();
304 let mut message_id = None;
305 let mut first_item_text = None;
306 let last_token_usage = Arc::new(Mutex::new(TokenUsage::default()));
307
308 if let Some(first_event) = events.next().await {
309 match first_event {
310 Ok(LanguageModelCompletionEvent::StartMessage { message_id: id }) => {
311 message_id = Some(id.clone());
312 }
313 Ok(LanguageModelCompletionEvent::Text(text)) => {
314 first_item_text = Some(text);
315 }
316 _ => (),
317 }
318 }
319
320 let stream = futures::stream::iter(first_item_text.map(Ok))
321 .chain(events.filter_map({
322 let last_token_usage = last_token_usage.clone();
323 move |result| {
324 let last_token_usage = last_token_usage.clone();
325 async move {
326 match result {
327 Ok(LanguageModelCompletionEvent::StatusUpdate { .. }) => None,
328 Ok(LanguageModelCompletionEvent::StartMessage { .. }) => None,
329 Ok(LanguageModelCompletionEvent::Text(text)) => Some(Ok(text)),
330 Ok(LanguageModelCompletionEvent::Thinking { .. }) => None,
331 Ok(LanguageModelCompletionEvent::Stop(_)) => None,
332 Ok(LanguageModelCompletionEvent::ToolUse(_)) => None,
333 Ok(LanguageModelCompletionEvent::UsageUpdate(token_usage)) => {
334 *last_token_usage.lock() = token_usage;
335 None
336 }
337 Err(err) => Some(Err(err)),
338 }
339 }
340 }
341 }))
342 .boxed();
343
344 Ok(LanguageModelTextStream {
345 message_id,
346 stream,
347 last_token_usage,
348 })
349 }
350 .boxed()
351 }
352
353 fn cache_configuration(&self) -> Option<LanguageModelCacheConfiguration> {
354 None
355 }
356
357 #[cfg(any(test, feature = "test-support"))]
358 fn as_fake(&self) -> &fake_provider::FakeLanguageModel {
359 unimplemented!()
360 }
361}
362
363#[derive(Debug, Error)]
364pub enum LanguageModelKnownError {
365 #[error("Context window limit exceeded ({tokens})")]
366 ContextWindowLimitExceeded { tokens: usize },
367}
368
369pub trait LanguageModelTool: 'static + DeserializeOwned + JsonSchema {
370 fn name() -> String;
371 fn description() -> String;
372}
373
374/// An error that occurred when trying to authenticate the language model provider.
375#[derive(Debug, Error)]
376pub enum AuthenticateError {
377 #[error("credentials not found")]
378 CredentialsNotFound,
379 #[error(transparent)]
380 Other(#[from] anyhow::Error),
381}
382
383pub trait LanguageModelProvider: 'static {
384 fn id(&self) -> LanguageModelProviderId;
385 fn name(&self) -> LanguageModelProviderName;
386 fn icon(&self) -> IconName {
387 IconName::ZedAssistant
388 }
389 fn default_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;
390 fn default_fast_model(&self, cx: &App) -> Option<Arc<dyn LanguageModel>>;
391 fn provided_models(&self, cx: &App) -> Vec<Arc<dyn LanguageModel>>;
392 fn recommended_models(&self, _cx: &App) -> Vec<Arc<dyn LanguageModel>> {
393 Vec::new()
394 }
395 fn load_model(&self, _model: Arc<dyn LanguageModel>, _cx: &App) {}
396 fn is_authenticated(&self, cx: &App) -> bool;
397 fn authenticate(&self, cx: &mut App) -> Task<Result<(), AuthenticateError>>;
398 fn configuration_view(&self, window: &mut Window, cx: &mut App) -> AnyView;
399 fn must_accept_terms(&self, _cx: &App) -> bool {
400 false
401 }
402 fn render_accept_terms(
403 &self,
404 _view: LanguageModelProviderTosView,
405 _cx: &mut App,
406 ) -> Option<AnyElement> {
407 None
408 }
409 fn reset_credentials(&self, cx: &mut App) -> Task<Result<()>>;
410}
411
412#[derive(PartialEq, Eq)]
413pub enum LanguageModelProviderTosView {
414 /// When there are some past interactions in the Agent Panel.
415 ThreadtEmptyState,
416 /// When there are no past interactions in the Agent Panel.
417 ThreadFreshStart,
418 PromptEditorPopup,
419 Configuration,
420}
421
422pub trait LanguageModelProviderState: 'static {
423 type ObservableEntity;
424
425 fn observable_entity(&self) -> Option<gpui::Entity<Self::ObservableEntity>>;
426
427 fn subscribe<T: 'static>(
428 &self,
429 cx: &mut gpui::Context<T>,
430 callback: impl Fn(&mut T, &mut gpui::Context<T>) + 'static,
431 ) -> Option<gpui::Subscription> {
432 let entity = self.observable_entity()?;
433 Some(cx.observe(&entity, move |this, _, cx| {
434 callback(this, cx);
435 }))
436 }
437}
438
439#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd, Serialize, Deserialize)]
440pub struct LanguageModelId(pub SharedString);
441
442#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
443pub struct LanguageModelName(pub SharedString);
444
445#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
446pub struct LanguageModelProviderId(pub SharedString);
447
448#[derive(Clone, Eq, PartialEq, Hash, Debug, Ord, PartialOrd)]
449pub struct LanguageModelProviderName(pub SharedString);
450
451impl fmt::Display for LanguageModelProviderId {
452 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
453 write!(f, "{}", self.0)
454 }
455}
456
457impl From<String> for LanguageModelId {
458 fn from(value: String) -> Self {
459 Self(SharedString::from(value))
460 }
461}
462
463impl From<String> for LanguageModelName {
464 fn from(value: String) -> Self {
465 Self(SharedString::from(value))
466 }
467}
468
469impl From<String> for LanguageModelProviderId {
470 fn from(value: String) -> Self {
471 Self(SharedString::from(value))
472 }
473}
474
475impl From<String> for LanguageModelProviderName {
476 fn from(value: String) -> Self {
477 Self(SharedString::from(value))
478 }
479}