1mod authorization;
2pub mod db;
3mod telemetry;
4mod token;
5
6use crate::{
7 api::CloudflareIpCountryHeader, build_clickhouse_client, executor::Executor, Config, Error,
8 Result,
9};
10use anyhow::{anyhow, Context as _};
11use authorization::authorize_access_to_language_model;
12use axum::{
13 body::Body,
14 http::{self, HeaderName, HeaderValue, Request, StatusCode},
15 middleware::{self, Next},
16 response::{IntoResponse, Response},
17 routing::post,
18 Extension, Json, Router, TypedHeader,
19};
20use chrono::{DateTime, Duration, Utc};
21use db::{ActiveUserCount, LlmDatabase};
22use futures::{Stream, StreamExt as _};
23use http_client::IsahcHttpClient;
24use rpc::{
25 proto::Plan, LanguageModelProvider, PerformCompletionParams, EXPIRED_LLM_TOKEN_HEADER_NAME,
26};
27use std::{
28 pin::Pin,
29 sync::Arc,
30 task::{Context, Poll},
31};
32use telemetry::{report_llm_usage, LlmUsageEventRow};
33use tokio::sync::RwLock;
34use util::ResultExt;
35
36pub use token::*;
37
38pub struct LlmState {
39 pub config: Config,
40 pub executor: Executor,
41 pub db: Arc<LlmDatabase>,
42 pub http_client: IsahcHttpClient,
43 pub clickhouse_client: Option<clickhouse::Client>,
44 active_user_count: RwLock<Option<(DateTime<Utc>, ActiveUserCount)>>,
45}
46
47const ACTIVE_USER_COUNT_CACHE_DURATION: Duration = Duration::seconds(30);
48
49impl LlmState {
50 pub async fn new(config: Config, executor: Executor) -> Result<Arc<Self>> {
51 let database_url = config
52 .llm_database_url
53 .as_ref()
54 .ok_or_else(|| anyhow!("missing LLM_DATABASE_URL"))?;
55 let max_connections = config
56 .llm_database_max_connections
57 .ok_or_else(|| anyhow!("missing LLM_DATABASE_MAX_CONNECTIONS"))?;
58
59 let mut db_options = db::ConnectOptions::new(database_url);
60 db_options.max_connections(max_connections);
61 let mut db = LlmDatabase::new(db_options, executor.clone()).await?;
62 db.initialize().await?;
63
64 let db = Arc::new(db);
65
66 let user_agent = format!("Zed Server/{}", env!("CARGO_PKG_VERSION"));
67 let http_client = IsahcHttpClient::builder()
68 .default_header("User-Agent", user_agent)
69 .build()
70 .context("failed to construct http client")?;
71
72 let initial_active_user_count =
73 Some((Utc::now(), db.get_active_user_count(Utc::now()).await?));
74
75 let this = Self {
76 executor,
77 db,
78 http_client,
79 clickhouse_client: config
80 .clickhouse_url
81 .as_ref()
82 .and_then(|_| build_clickhouse_client(&config).log_err()),
83 active_user_count: RwLock::new(initial_active_user_count),
84 config,
85 };
86
87 Ok(Arc::new(this))
88 }
89
90 pub async fn get_active_user_count(&self) -> Result<ActiveUserCount> {
91 let now = Utc::now();
92
93 if let Some((last_updated, count)) = self.active_user_count.read().await.as_ref() {
94 if now - *last_updated < ACTIVE_USER_COUNT_CACHE_DURATION {
95 return Ok(*count);
96 }
97 }
98
99 let mut cache = self.active_user_count.write().await;
100 let new_count = self.db.get_active_user_count(now).await?;
101 *cache = Some((now, new_count));
102 Ok(new_count)
103 }
104}
105
106pub fn routes() -> Router<(), Body> {
107 Router::new()
108 .route("/completion", post(perform_completion))
109 .layer(middleware::from_fn(validate_api_token))
110}
111
112async fn validate_api_token<B>(mut req: Request<B>, next: Next<B>) -> impl IntoResponse {
113 let token = req
114 .headers()
115 .get(http::header::AUTHORIZATION)
116 .and_then(|header| header.to_str().ok())
117 .ok_or_else(|| {
118 Error::http(
119 StatusCode::BAD_REQUEST,
120 "missing authorization header".to_string(),
121 )
122 })?
123 .strip_prefix("Bearer ")
124 .ok_or_else(|| {
125 Error::http(
126 StatusCode::BAD_REQUEST,
127 "invalid authorization header".to_string(),
128 )
129 })?;
130
131 let state = req.extensions().get::<Arc<LlmState>>().unwrap();
132 match LlmTokenClaims::validate(&token, &state.config) {
133 Ok(claims) => {
134 req.extensions_mut().insert(claims);
135 Ok::<_, Error>(next.run(req).await.into_response())
136 }
137 Err(ValidateLlmTokenError::Expired) => Err(Error::Http(
138 StatusCode::UNAUTHORIZED,
139 "unauthorized".to_string(),
140 [(
141 HeaderName::from_static(EXPIRED_LLM_TOKEN_HEADER_NAME),
142 HeaderValue::from_static("true"),
143 )]
144 .into_iter()
145 .collect(),
146 )),
147 Err(_err) => Err(Error::http(
148 StatusCode::UNAUTHORIZED,
149 "unauthorized".to_string(),
150 )),
151 }
152}
153
154async fn perform_completion(
155 Extension(state): Extension<Arc<LlmState>>,
156 Extension(claims): Extension<LlmTokenClaims>,
157 country_code_header: Option<TypedHeader<CloudflareIpCountryHeader>>,
158 Json(params): Json<PerformCompletionParams>,
159) -> Result<impl IntoResponse> {
160 let model = normalize_model_name(params.provider, params.model);
161
162 authorize_access_to_language_model(
163 &state.config,
164 &claims,
165 country_code_header.map(|header| header.to_string()),
166 params.provider,
167 &model,
168 )?;
169
170 check_usage_limit(&state, params.provider, &model, &claims).await?;
171
172 let stream = match params.provider {
173 LanguageModelProvider::Anthropic => {
174 let api_key = state
175 .config
176 .anthropic_api_key
177 .as_ref()
178 .context("no Anthropic AI API key configured on the server")?;
179
180 let mut request: anthropic::Request =
181 serde_json::from_str(¶ms.provider_request.get())?;
182
183 // Parse the model, throw away the version that was included, and then set a specific
184 // version that we control on the server.
185 // Right now, we use the version that's defined in `model.id()`, but we will likely
186 // want to change this code once a new version of an Anthropic model is released,
187 // so that users can use the new version, without having to update Zed.
188 request.model = match anthropic::Model::from_id(&request.model) {
189 Ok(model) => model.id().to_string(),
190 Err(_) => request.model,
191 };
192
193 let chunks = anthropic::stream_completion(
194 &state.http_client,
195 anthropic::ANTHROPIC_API_URL,
196 api_key,
197 request,
198 None,
199 )
200 .await?;
201
202 chunks
203 .map(move |event| {
204 let chunk = event?;
205 let (input_tokens, output_tokens) = match &chunk {
206 anthropic::Event::MessageStart {
207 message: anthropic::Response { usage, .. },
208 }
209 | anthropic::Event::MessageDelta { usage, .. } => (
210 usage.input_tokens.unwrap_or(0) as usize,
211 usage.output_tokens.unwrap_or(0) as usize,
212 ),
213 _ => (0, 0),
214 };
215
216 anyhow::Ok((
217 serde_json::to_vec(&chunk).unwrap(),
218 input_tokens,
219 output_tokens,
220 ))
221 })
222 .boxed()
223 }
224 LanguageModelProvider::OpenAi => {
225 let api_key = state
226 .config
227 .openai_api_key
228 .as_ref()
229 .context("no OpenAI API key configured on the server")?;
230 let chunks = open_ai::stream_completion(
231 &state.http_client,
232 open_ai::OPEN_AI_API_URL,
233 api_key,
234 serde_json::from_str(¶ms.provider_request.get())?,
235 None,
236 )
237 .await?;
238
239 chunks
240 .map(|event| {
241 event.map(|chunk| {
242 let input_tokens =
243 chunk.usage.as_ref().map_or(0, |u| u.prompt_tokens) as usize;
244 let output_tokens =
245 chunk.usage.as_ref().map_or(0, |u| u.completion_tokens) as usize;
246 (
247 serde_json::to_vec(&chunk).unwrap(),
248 input_tokens,
249 output_tokens,
250 )
251 })
252 })
253 .boxed()
254 }
255 LanguageModelProvider::Google => {
256 let api_key = state
257 .config
258 .google_ai_api_key
259 .as_ref()
260 .context("no Google AI API key configured on the server")?;
261 let chunks = google_ai::stream_generate_content(
262 &state.http_client,
263 google_ai::API_URL,
264 api_key,
265 serde_json::from_str(¶ms.provider_request.get())?,
266 )
267 .await?;
268
269 chunks
270 .map(|event| {
271 event.map(|chunk| {
272 // TODO - implement token counting for Google AI
273 let input_tokens = 0;
274 let output_tokens = 0;
275 (
276 serde_json::to_vec(&chunk).unwrap(),
277 input_tokens,
278 output_tokens,
279 )
280 })
281 })
282 .boxed()
283 }
284 LanguageModelProvider::Zed => {
285 let api_key = state
286 .config
287 .qwen2_7b_api_key
288 .as_ref()
289 .context("no Qwen2-7B API key configured on the server")?;
290 let api_url = state
291 .config
292 .qwen2_7b_api_url
293 .as_ref()
294 .context("no Qwen2-7B URL configured on the server")?;
295 let chunks = open_ai::stream_completion(
296 &state.http_client,
297 &api_url,
298 api_key,
299 serde_json::from_str(¶ms.provider_request.get())?,
300 None,
301 )
302 .await?;
303
304 chunks
305 .map(|event| {
306 event.map(|chunk| {
307 let input_tokens =
308 chunk.usage.as_ref().map_or(0, |u| u.prompt_tokens) as usize;
309 let output_tokens =
310 chunk.usage.as_ref().map_or(0, |u| u.completion_tokens) as usize;
311 (
312 serde_json::to_vec(&chunk).unwrap(),
313 input_tokens,
314 output_tokens,
315 )
316 })
317 })
318 .boxed()
319 }
320 };
321
322 Ok(Response::new(Body::wrap_stream(TokenCountingStream {
323 state,
324 claims,
325 provider: params.provider,
326 model,
327 input_tokens: 0,
328 output_tokens: 0,
329 inner_stream: stream,
330 })))
331}
332
333fn normalize_model_name(provider: LanguageModelProvider, name: String) -> String {
334 let prefixes: &[_] = match provider {
335 LanguageModelProvider::Anthropic => &[
336 "claude-3-5-sonnet",
337 "claude-3-haiku",
338 "claude-3-opus",
339 "claude-3-sonnet",
340 ],
341 LanguageModelProvider::OpenAi => &[
342 "gpt-3.5-turbo",
343 "gpt-4-turbo-preview",
344 "gpt-4o-mini",
345 "gpt-4o",
346 "gpt-4",
347 ],
348 LanguageModelProvider::Google => &[],
349 LanguageModelProvider::Zed => &[],
350 };
351
352 if let Some(prefix) = prefixes
353 .iter()
354 .filter(|&&prefix| name.starts_with(prefix))
355 .max_by_key(|&&prefix| prefix.len())
356 {
357 prefix.to_string()
358 } else {
359 name
360 }
361}
362
363async fn check_usage_limit(
364 state: &Arc<LlmState>,
365 provider: LanguageModelProvider,
366 model_name: &str,
367 claims: &LlmTokenClaims,
368) -> Result<()> {
369 let model = state.db.model(provider, model_name)?;
370 let usage = state
371 .db
372 .get_usage(claims.user_id as i32, provider, model_name, Utc::now())
373 .await?;
374
375 let active_users = state.get_active_user_count().await?;
376
377 let per_user_max_requests_per_minute =
378 model.max_requests_per_minute as usize / active_users.users_in_recent_minutes.max(1);
379 let per_user_max_tokens_per_minute =
380 model.max_tokens_per_minute as usize / active_users.users_in_recent_minutes.max(1);
381 let per_user_max_tokens_per_day =
382 model.max_tokens_per_day as usize / active_users.users_in_recent_days.max(1);
383
384 let checks = [
385 (
386 usage.requests_this_minute,
387 per_user_max_requests_per_minute,
388 "requests per minute",
389 ),
390 (
391 usage.tokens_this_minute,
392 per_user_max_tokens_per_minute,
393 "tokens per minute",
394 ),
395 (
396 usage.tokens_this_day,
397 per_user_max_tokens_per_day,
398 "tokens per day",
399 ),
400 ];
401
402 for (usage, limit, resource) in checks {
403 // Temporarily bypass rate-limiting for staff members.
404 if claims.is_staff {
405 continue;
406 }
407
408 if usage > limit {
409 return Err(Error::http(
410 StatusCode::TOO_MANY_REQUESTS,
411 format!("Rate limit exceeded. Maximum {} reached.", resource),
412 ));
413 }
414 }
415
416 Ok(())
417}
418
419struct TokenCountingStream<S> {
420 state: Arc<LlmState>,
421 claims: LlmTokenClaims,
422 provider: LanguageModelProvider,
423 model: String,
424 input_tokens: usize,
425 output_tokens: usize,
426 inner_stream: S,
427}
428
429impl<S> Stream for TokenCountingStream<S>
430where
431 S: Stream<Item = Result<(Vec<u8>, usize, usize), anyhow::Error>> + Unpin,
432{
433 type Item = Result<Vec<u8>, anyhow::Error>;
434
435 fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
436 match Pin::new(&mut self.inner_stream).poll_next(cx) {
437 Poll::Ready(Some(Ok((mut bytes, input_tokens, output_tokens)))) => {
438 bytes.push(b'\n');
439 self.input_tokens += input_tokens;
440 self.output_tokens += output_tokens;
441 Poll::Ready(Some(Ok(bytes)))
442 }
443 Poll::Ready(Some(Err(e))) => Poll::Ready(Some(Err(e))),
444 Poll::Ready(None) => Poll::Ready(None),
445 Poll::Pending => Poll::Pending,
446 }
447 }
448}
449
450impl<S> Drop for TokenCountingStream<S> {
451 fn drop(&mut self) {
452 let state = self.state.clone();
453 let claims = self.claims.clone();
454 let provider = self.provider;
455 let model = std::mem::take(&mut self.model);
456 let input_token_count = self.input_tokens;
457 let output_token_count = self.output_tokens;
458 self.state.executor.spawn_detached(async move {
459 let usage = state
460 .db
461 .record_usage(
462 claims.user_id as i32,
463 provider,
464 &model,
465 input_token_count,
466 output_token_count,
467 Utc::now(),
468 )
469 .await
470 .log_err();
471
472 if let Some((clickhouse_client, usage)) = state.clickhouse_client.as_ref().zip(usage) {
473 report_llm_usage(
474 clickhouse_client,
475 LlmUsageEventRow {
476 time: Utc::now().timestamp_millis(),
477 user_id: claims.user_id as i32,
478 is_staff: claims.is_staff,
479 plan: match claims.plan {
480 Plan::Free => "free".to_string(),
481 Plan::ZedPro => "zed_pro".to_string(),
482 },
483 model,
484 provider: provider.to_string(),
485 input_token_count: input_token_count as u64,
486 output_token_count: output_token_count as u64,
487 requests_this_minute: usage.requests_this_minute as u64,
488 tokens_this_minute: usage.tokens_this_minute as u64,
489 tokens_this_day: usage.tokens_this_day as u64,
490 input_tokens_this_month: usage.input_tokens_this_month as u64,
491 output_tokens_this_month: usage.output_tokens_this_month as u64,
492 spending_this_month: usage.spending_this_month as u64,
493 },
494 )
495 .await
496 .log_err();
497 }
498 })
499 }
500}