embedding.rs

  1use anyhow::{anyhow, Result};
  2use async_trait::async_trait;
  3use futures::AsyncReadExt;
  4use gpui::executor::Background;
  5use gpui::serde_json;
  6use isahc::http::StatusCode;
  7use isahc::prelude::Configurable;
  8use isahc::{AsyncBody, Response};
  9use lazy_static::lazy_static;
 10use ordered_float::OrderedFloat;
 11use parking_lot::Mutex;
 12use parse_duration::parse;
 13use postage::watch;
 14use rusqlite::types::{FromSql, FromSqlResult, ToSqlOutput, ValueRef};
 15use rusqlite::ToSql;
 16use serde::{Deserialize, Serialize};
 17use std::env;
 18use std::ops::Add;
 19use std::sync::Arc;
 20use std::time::{Duration, Instant};
 21use tiktoken_rs::{cl100k_base, CoreBPE};
 22use util::http::{HttpClient, Request};
 23
 24lazy_static! {
 25    static ref OPENAI_API_KEY: Option<String> = env::var("OPENAI_API_KEY").ok();
 26    static ref OPENAI_BPE_TOKENIZER: CoreBPE = cl100k_base().unwrap();
 27}
 28
 29#[derive(Debug, PartialEq, Clone)]
 30pub struct Embedding(Vec<f32>);
 31
 32impl From<Vec<f32>> for Embedding {
 33    fn from(value: Vec<f32>) -> Self {
 34        Embedding(value)
 35    }
 36}
 37
 38impl Embedding {
 39    pub fn similarity(&self, other: &Self) -> OrderedFloat<f32> {
 40        let len = self.0.len();
 41        assert_eq!(len, other.0.len());
 42
 43        let mut result = 0.0;
 44        unsafe {
 45            matrixmultiply::sgemm(
 46                1,
 47                len,
 48                1,
 49                1.0,
 50                self.0.as_ptr(),
 51                len as isize,
 52                1,
 53                other.0.as_ptr(),
 54                1,
 55                len as isize,
 56                0.0,
 57                &mut result as *mut f32,
 58                1,
 59                1,
 60            );
 61        }
 62        OrderedFloat(result)
 63    }
 64}
 65
 66impl FromSql for Embedding {
 67    fn column_result(value: ValueRef) -> FromSqlResult<Self> {
 68        let bytes = value.as_blob()?;
 69        let embedding: Result<Vec<f32>, Box<bincode::ErrorKind>> = bincode::deserialize(bytes);
 70        if embedding.is_err() {
 71            return Err(rusqlite::types::FromSqlError::Other(embedding.unwrap_err()));
 72        }
 73        Ok(Embedding(embedding.unwrap()))
 74    }
 75}
 76
 77impl ToSql for Embedding {
 78    fn to_sql(&self) -> rusqlite::Result<ToSqlOutput> {
 79        let bytes = bincode::serialize(&self.0)
 80            .map_err(|err| rusqlite::Error::ToSqlConversionFailure(Box::new(err)))?;
 81        Ok(ToSqlOutput::Owned(rusqlite::types::Value::Blob(bytes)))
 82    }
 83}
 84
 85#[derive(Clone)]
 86pub struct OpenAIEmbeddings {
 87    pub client: Arc<dyn HttpClient>,
 88    pub executor: Arc<Background>,
 89    rate_limit_count_rx: watch::Receiver<Option<Instant>>,
 90    rate_limit_count_tx: Arc<Mutex<watch::Sender<Option<Instant>>>>,
 91}
 92
 93#[derive(Serialize)]
 94struct OpenAIEmbeddingRequest<'a> {
 95    model: &'static str,
 96    input: Vec<&'a str>,
 97}
 98
 99#[derive(Deserialize)]
100struct OpenAIEmbeddingResponse {
101    data: Vec<OpenAIEmbedding>,
102    usage: OpenAIEmbeddingUsage,
103}
104
105#[derive(Debug, Deserialize)]
106struct OpenAIEmbedding {
107    embedding: Vec<f32>,
108    index: usize,
109    object: String,
110}
111
112#[derive(Deserialize)]
113struct OpenAIEmbeddingUsage {
114    prompt_tokens: usize,
115    total_tokens: usize,
116}
117
118#[async_trait]
119pub trait EmbeddingProvider: Sync + Send {
120    async fn embed_batch(&self, spans: Vec<String>) -> Result<Vec<Embedding>>;
121    fn max_tokens_per_batch(&self) -> usize;
122    fn truncate(&self, span: &str) -> (String, usize);
123    fn rate_limit_expiration(&self) -> Option<Instant>;
124}
125
126pub struct DummyEmbeddings {}
127
128#[async_trait]
129impl EmbeddingProvider for DummyEmbeddings {
130    fn rate_limit_expiration(&self) -> Option<Instant> {
131        None
132    }
133    async fn embed_batch(&self, spans: Vec<String>) -> Result<Vec<Embedding>> {
134        // 1024 is the OpenAI Embeddings size for ada models.
135        // the model we will likely be starting with.
136        let dummy_vec = Embedding::from(vec![0.32 as f32; 1536]);
137        return Ok(vec![dummy_vec; spans.len()]);
138    }
139
140    fn max_tokens_per_batch(&self) -> usize {
141        OPENAI_INPUT_LIMIT
142    }
143
144    fn truncate(&self, span: &str) -> (String, usize) {
145        let mut tokens = OPENAI_BPE_TOKENIZER.encode_with_special_tokens(span);
146        let token_count = tokens.len();
147        let output = if token_count > OPENAI_INPUT_LIMIT {
148            tokens.truncate(OPENAI_INPUT_LIMIT);
149            let new_input = OPENAI_BPE_TOKENIZER.decode(tokens.clone());
150            new_input.ok().unwrap_or_else(|| span.to_string())
151        } else {
152            span.to_string()
153        };
154
155        (output, tokens.len())
156    }
157}
158
159const OPENAI_INPUT_LIMIT: usize = 8190;
160
161impl OpenAIEmbeddings {
162    pub fn new(client: Arc<dyn HttpClient>, executor: Arc<Background>) -> Self {
163        let (rate_limit_count_tx, rate_limit_count_rx) = watch::channel_with(None);
164        let rate_limit_count_tx = Arc::new(Mutex::new(rate_limit_count_tx));
165
166        OpenAIEmbeddings {
167            client,
168            executor,
169            rate_limit_count_rx,
170            rate_limit_count_tx,
171        }
172    }
173
174    fn resolve_rate_limit(&self) {
175        let reset_time = *self.rate_limit_count_tx.lock().borrow();
176
177        if let Some(reset_time) = reset_time {
178            if Instant::now() >= reset_time {
179                *self.rate_limit_count_tx.lock().borrow_mut() = None
180            }
181        }
182
183        log::trace!(
184            "resolving reset time: {:?}",
185            *self.rate_limit_count_tx.lock().borrow()
186        );
187    }
188
189    fn update_reset_time(&self, reset_time: Instant) {
190        let original_time = *self.rate_limit_count_tx.lock().borrow();
191
192        let updated_time = if let Some(original_time) = original_time {
193            if reset_time < original_time {
194                Some(reset_time)
195            } else {
196                Some(original_time)
197            }
198        } else {
199            Some(reset_time)
200        };
201
202        log::trace!("updating rate limit time: {:?}", updated_time);
203
204        *self.rate_limit_count_tx.lock().borrow_mut() = updated_time;
205    }
206    async fn send_request(
207        &self,
208        api_key: &str,
209        spans: Vec<&str>,
210        request_timeout: u64,
211    ) -> Result<Response<AsyncBody>> {
212        let request = Request::post("https://api.openai.com/v1/embeddings")
213            .redirect_policy(isahc::config::RedirectPolicy::Follow)
214            .timeout(Duration::from_secs(request_timeout))
215            .header("Content-Type", "application/json")
216            .header("Authorization", format!("Bearer {}", api_key))
217            .body(
218                serde_json::to_string(&OpenAIEmbeddingRequest {
219                    input: spans.clone(),
220                    model: "text-embedding-ada-002",
221                })
222                .unwrap()
223                .into(),
224            )?;
225
226        Ok(self.client.send(request).await?)
227    }
228}
229
230#[async_trait]
231impl EmbeddingProvider for OpenAIEmbeddings {
232    fn max_tokens_per_batch(&self) -> usize {
233        50000
234    }
235
236    fn rate_limit_expiration(&self) -> Option<Instant> {
237        *self.rate_limit_count_rx.borrow()
238    }
239    fn truncate(&self, span: &str) -> (String, usize) {
240        let mut tokens = OPENAI_BPE_TOKENIZER.encode_with_special_tokens(span);
241        let output = if tokens.len() > OPENAI_INPUT_LIMIT {
242            tokens.truncate(OPENAI_INPUT_LIMIT);
243            OPENAI_BPE_TOKENIZER
244                .decode(tokens.clone())
245                .ok()
246                .unwrap_or_else(|| span.to_string())
247        } else {
248            span.to_string()
249        };
250
251        (output, tokens.len())
252    }
253
254    async fn embed_batch(&self, spans: Vec<String>) -> Result<Vec<Embedding>> {
255        const BACKOFF_SECONDS: [usize; 4] = [3, 5, 15, 45];
256        const MAX_RETRIES: usize = 4;
257
258        let api_key = OPENAI_API_KEY
259            .as_ref()
260            .ok_or_else(|| anyhow!("no api key"))?;
261
262        let mut request_number = 0;
263        let mut rate_limiting = false;
264        let mut request_timeout: u64 = 15;
265        let mut response: Response<AsyncBody>;
266        while request_number < MAX_RETRIES {
267            response = self
268                .send_request(
269                    api_key,
270                    spans.iter().map(|x| &**x).collect(),
271                    request_timeout,
272                )
273                .await?;
274            request_number += 1;
275
276            match response.status() {
277                StatusCode::REQUEST_TIMEOUT => {
278                    request_timeout += 5;
279                }
280                StatusCode::OK => {
281                    let mut body = String::new();
282                    response.body_mut().read_to_string(&mut body).await?;
283                    let response: OpenAIEmbeddingResponse = serde_json::from_str(&body)?;
284
285                    log::trace!(
286                        "openai embedding completed. tokens: {:?}",
287                        response.usage.total_tokens
288                    );
289
290                    // If we complete a request successfully that was previously rate_limited
291                    // resolve the rate limit
292                    if rate_limiting {
293                        self.resolve_rate_limit()
294                    }
295
296                    return Ok(response
297                        .data
298                        .into_iter()
299                        .map(|embedding| Embedding::from(embedding.embedding))
300                        .collect());
301                }
302                StatusCode::TOO_MANY_REQUESTS => {
303                    rate_limiting = true;
304                    let mut body = String::new();
305                    response.body_mut().read_to_string(&mut body).await?;
306
307                    let delay_duration = {
308                        let delay = Duration::from_secs(BACKOFF_SECONDS[request_number - 1] as u64);
309                        if let Some(time_to_reset) =
310                            response.headers().get("x-ratelimit-reset-tokens")
311                        {
312                            if let Ok(time_str) = time_to_reset.to_str() {
313                                parse(time_str).unwrap_or(delay)
314                            } else {
315                                delay
316                            }
317                        } else {
318                            delay
319                        }
320                    };
321
322                    // If we've previously rate limited, increment the duration but not the count
323                    let reset_time = Instant::now().add(delay_duration);
324                    self.update_reset_time(reset_time);
325
326                    log::trace!(
327                        "openai rate limiting: waiting {:?} until lifted",
328                        &delay_duration
329                    );
330
331                    self.executor.timer(delay_duration).await;
332                }
333                _ => {
334                    let mut body = String::new();
335                    response.body_mut().read_to_string(&mut body).await?;
336                    return Err(anyhow!(
337                        "open ai bad request: {:?} {:?}",
338                        &response.status(),
339                        body
340                    ));
341                }
342            }
343        }
344        Err(anyhow!("openai max retries"))
345    }
346}
347
348#[cfg(test)]
349mod tests {
350    use super::*;
351    use rand::prelude::*;
352
353    #[gpui::test]
354    fn test_similarity(mut rng: StdRng) {
355        assert_eq!(
356            Embedding::from(vec![1., 0., 0., 0., 0.])
357                .similarity(&Embedding::from(vec![0., 1., 0., 0., 0.])),
358            0.
359        );
360        assert_eq!(
361            Embedding::from(vec![2., 0., 0., 0., 0.])
362                .similarity(&Embedding::from(vec![3., 1., 0., 0., 0.])),
363            6.
364        );
365
366        for _ in 0..100 {
367            let size = 1536;
368            let mut a = vec![0.; size];
369            let mut b = vec![0.; size];
370            for (a, b) in a.iter_mut().zip(b.iter_mut()) {
371                *a = rng.gen();
372                *b = rng.gen();
373            }
374            let a = Embedding::from(a);
375            let b = Embedding::from(b);
376
377            assert_eq!(
378                round_to_decimals(a.similarity(&b), 1),
379                round_to_decimals(reference_dot(&a.0, &b.0), 1)
380            );
381        }
382
383        fn round_to_decimals(n: OrderedFloat<f32>, decimal_places: i32) -> f32 {
384            let factor = (10.0 as f32).powi(decimal_places);
385            (n * factor).round() / factor
386        }
387
388        fn reference_dot(a: &[f32], b: &[f32]) -> OrderedFloat<f32> {
389            OrderedFloat(a.iter().zip(b.iter()).map(|(a, b)| a * b).sum())
390        }
391    }
392}