memvid_cli/
openai_embeddings.rs

1//! OpenAI Embeddings Provider
2//!
3//! This module provides an `EmbeddingProvider` implementation that uses
4//! OpenAI's text-embedding API for generating high-quality embeddings.
5//!
6//! ## Environment Variables
7//! - `OPENAI_API_KEY`: Required API key for OpenAI
8//! - `OPENAI_EMBEDDING_MODEL`: Optional model override (default: text-embedding-3-large)
9//!
10//! ## Features
11//! - Supports all OpenAI embedding models
12//! - Efficient batch processing (up to 100 texts per request)
13//! - Automatic rate limiting with exponential backoff
14//! - Thread-safe for concurrent use
15
16use anyhow::{anyhow, bail, Result};
17use memvid_core::{EmbeddingConfig, EmbeddingProvider, VecEmbedder};
18use reqwest::blocking::Client;
19use serde::{Deserialize, Serialize};
20use std::sync::atomic::{AtomicBool, Ordering};
21use std::time::Duration;
22use tracing::{debug, info, warn};
23
24/// OpenAI embeddings API endpoint
25const OPENAI_EMBEDDINGS_URL: &str = "https://api.openai.com/v1/embeddings";
26
27/// Maximum texts per batch (OpenAI limit)
28const MAX_BATCH_SIZE: usize = 100;
29
30/// Request timeout
31const REQUEST_TIMEOUT: Duration = Duration::from_secs(60);
32
33/// Maximum characters for embedding text to avoid exceeding OpenAI's 8192 token limit.
34/// Using ~3 chars/token estimate (conservative for dense content), 20K chars ≈ 6.6K tokens.
35const MAX_EMBEDDING_TEXT_LEN: usize = 20_000;
36
37/// Truncate text to MAX_EMBEDDING_TEXT_LEN to avoid token limit errors.
38fn truncate_for_embedding(text: &str) -> std::borrow::Cow<'_, str> {
39    if text.len() <= MAX_EMBEDDING_TEXT_LEN {
40        std::borrow::Cow::Borrowed(text)
41    } else {
42        // Find a safe char boundary
43        let end = text[..MAX_EMBEDDING_TEXT_LEN]
44            .char_indices()
45            .rev()
46            .next()
47            .map(|(i, c)| i + c.len_utf8())
48            .unwrap_or(MAX_EMBEDDING_TEXT_LEN);
49        warn!(
50            "Truncating embedding text from {} to {} chars to avoid token limit",
51            text.len(),
52            end
53        );
54        std::borrow::Cow::Owned(text[..end].to_string())
55    }
56}
57
58/// OpenAI embedding request payload
59#[derive(Debug, Serialize)]
60struct OpenAIEmbeddingRequest<'a> {
61    model: &'a str,
62    input: Vec<&'a str>,
63    #[serde(skip_serializing_if = "Option::is_none")]
64    dimensions: Option<usize>,
65}
66
67/// OpenAI embedding response
68#[derive(Debug, Deserialize)]
69struct OpenAIEmbeddingResponse {
70    data: Vec<OpenAIEmbeddingData>,
71    model: String,
72    usage: OpenAIUsage,
73}
74
75#[derive(Debug, Deserialize)]
76struct OpenAIEmbeddingData {
77    embedding: Vec<f32>,
78    index: usize,
79}
80
81#[derive(Debug, Deserialize)]
82struct OpenAIUsage {
83    #[allow(dead_code)]
84    prompt_tokens: usize,
85    total_tokens: usize,
86}
87
88/// OpenAI error response
89#[derive(Debug, Deserialize)]
90struct OpenAIErrorResponse {
91    error: OpenAIError,
92}
93
94#[derive(Debug, Deserialize)]
95struct OpenAIError {
96    message: String,
97    #[serde(rename = "type")]
98    error_type: String,
99}
100
101/// OpenAI Embedding Provider
102///
103/// Implements `EmbeddingProvider` trait for generating embeddings via OpenAI API.
104#[derive(Clone)]
105pub struct OpenAIEmbeddingProvider {
106    api_key: String,
107    config: EmbeddingConfig,
108    client: Client,
109    ready: std::sync::Arc<AtomicBool>,
110}
111
112impl std::fmt::Debug for OpenAIEmbeddingProvider {
113    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
114        f.debug_struct("OpenAIEmbeddingProvider")
115            .field("model", &self.config.model)
116            .field("dimension", &self.config.dimension)
117            .field("ready", &self.ready.load(Ordering::Relaxed))
118            .finish()
119    }
120}
121
122impl OpenAIEmbeddingProvider {
123    /// Create a new OpenAI embedding provider
124    ///
125    /// # Arguments
126    /// * `api_key` - OpenAI API key
127    /// * `config` - Embedding configuration (model, dimension, etc.)
128    ///
129    /// # Example
130    /// ```ignore
131    /// let provider = OpenAIEmbeddingProvider::new(
132    ///     "sk-...".to_string(),
133    ///     EmbeddingConfig::openai_large(),
134    /// )?;
135    /// ```
136    pub fn new(api_key: String, config: EmbeddingConfig) -> Result<Self> {
137        if api_key.is_empty() {
138            bail!("OpenAI API key cannot be empty");
139        }
140
141        let client = crate::http::blocking_client(REQUEST_TIMEOUT)
142            .map_err(|e| anyhow!("Failed to create HTTP client: {}", e))?;
143
144        Ok(Self {
145            api_key,
146            config,
147            client,
148            ready: std::sync::Arc::new(AtomicBool::new(false)),
149        })
150    }
151
152    /// Create provider from environment variables
153    ///
154    /// Uses `OPENAI_API_KEY` for authentication and optionally
155    /// `OPENAI_EMBEDDING_MODEL` to override the default model.
156    pub fn from_env() -> Result<Self> {
157        let api_key = std::env::var("OPENAI_API_KEY")
158            .map_err(|_| anyhow!("OPENAI_API_KEY environment variable not set"))?;
159
160        let config = match std::env::var("OPENAI_EMBEDDING_MODEL") {
161            Ok(model) => match model.as_str() {
162                "text-embedding-3-small" => EmbeddingConfig::openai_small(),
163                "text-embedding-ada-002" => EmbeddingConfig::openai_ada(),
164                "text-embedding-3-large" | _ => EmbeddingConfig::openai_large(),
165            },
166            Err(_) => EmbeddingConfig::openai_large(),
167        };
168
169        Self::new(api_key, config)
170    }
171
172    /// Create provider with text-embedding-3-large (default, highest quality)
173    pub fn large(api_key: String) -> Result<Self> {
174        Self::new(api_key, EmbeddingConfig::openai_large())
175    }
176
177    /// Create provider with text-embedding-3-small (faster, lower cost)
178    pub fn small(api_key: String) -> Result<Self> {
179        Self::new(api_key, EmbeddingConfig::openai_small())
180    }
181
182    /// Create provider with text-embedding-ada-002 (legacy)
183    pub fn ada(api_key: String) -> Result<Self> {
184        Self::new(api_key, EmbeddingConfig::openai_ada())
185    }
186
187    /// Internal method to call OpenAI API
188    fn call_openai(&self, texts: &[&str]) -> Result<Vec<Vec<f32>>> {
189        if texts.is_empty() {
190            return Ok(Vec::new());
191        }
192
193        let request = OpenAIEmbeddingRequest {
194            model: &self.config.model,
195            input: texts.to_vec(),
196            dimensions: None, // Use model's native dimension
197        };
198
199        let response = self
200            .client
201            .post(OPENAI_EMBEDDINGS_URL)
202            .header("Authorization", format!("Bearer {}", self.api_key))
203            .header("Content-Type", "application/json")
204            .json(&request)
205            .send()
206            .map_err(|e| anyhow!("OpenAI API request failed: {}", e))?;
207
208        let status = response.status();
209        let body = response
210            .text()
211            .map_err(|e| anyhow!("Failed to read response body: {}", e))?;
212
213        if !status.is_success() {
214            // Try to parse error response
215            if let Ok(error_response) = serde_json::from_str::<OpenAIErrorResponse>(&body) {
216                bail!(
217                    "OpenAI API error ({}): {}",
218                    error_response.error.error_type,
219                    error_response.error.message
220                );
221            }
222            bail!("OpenAI API request failed with status {}: {}", status, body);
223        }
224
225        let embedding_response: OpenAIEmbeddingResponse = serde_json::from_str(&body)
226            .map_err(|e| anyhow!("Failed to parse OpenAI response: {}", e))?;
227
228        debug!(
229            "OpenAI embeddings: {} texts, {} tokens, model={}",
230            texts.len(),
231            embedding_response.usage.total_tokens,
232            embedding_response.model
233        );
234
235        // Sort by index and extract embeddings
236        let mut data = embedding_response.data;
237        data.sort_by_key(|d| d.index);
238
239        let embeddings: Vec<Vec<f32>> = data.into_iter().map(|d| d.embedding).collect();
240
241        // Validate dimensions
242        if let Some(first) = embeddings.first() {
243            if first.len() != self.config.dimension {
244                warn!(
245                    "OpenAI returned dimension {} but expected {}",
246                    first.len(),
247                    self.config.dimension
248                );
249            }
250        }
251
252        Ok(embeddings)
253    }
254
255    /// Embed texts with retry logic
256    fn embed_with_retry(&self, texts: &[&str], max_retries: usize) -> Result<Vec<Vec<f32>>> {
257        let mut last_error = None;
258
259        for attempt in 0..max_retries {
260            match self.call_openai(texts) {
261                Ok(embeddings) => return Ok(embeddings),
262                Err(e) => {
263                    let error_str = e.to_string();
264                    if error_str.contains("rate_limit") || error_str.contains("429") {
265                        let backoff = Duration::from_millis(500 * (1 << attempt));
266                        warn!(
267                            "Rate limited by OpenAI, retrying in {:?} (attempt {}/{})",
268                            backoff,
269                            attempt + 1,
270                            max_retries
271                        );
272                        std::thread::sleep(backoff);
273                        last_error = Some(e);
274                        continue;
275                    }
276                    return Err(e);
277                }
278            }
279        }
280
281        Err(last_error.unwrap_or_else(|| anyhow!("Failed to embed after {} retries", max_retries)))
282    }
283}
284
285impl EmbeddingProvider for OpenAIEmbeddingProvider {
286    fn kind(&self) -> &str {
287        "openai"
288    }
289
290    fn model(&self) -> &str {
291        &self.config.model
292    }
293
294    fn dimension(&self) -> usize {
295        self.config.dimension
296    }
297
298    fn embed_text(&self, text: &str) -> memvid_core::Result<Vec<f32>> {
299        let text = truncate_for_embedding(text);
300        self.embed_with_retry(&[&text], 3)
301            .map(|mut v| v.pop().unwrap_or_default())
302            .map_err(|e| memvid_core::MemvidError::EmbeddingFailed {
303                reason: e.to_string().into_boxed_str(),
304            })
305    }
306
307    fn embed_batch(&self, texts: &[&str]) -> memvid_core::Result<Vec<Vec<f32>>> {
308        if texts.is_empty() {
309            return Ok(Vec::new());
310        }
311
312        // Truncate all texts first to avoid token limit errors
313        let truncated: Vec<std::borrow::Cow<'_, str>> =
314            texts.iter().map(|t| truncate_for_embedding(t)).collect();
315        let truncated_refs: Vec<&str> = truncated.iter().map(|c| c.as_ref()).collect();
316
317        // Process in batches of MAX_BATCH_SIZE
318        let batch_size = self
319            .config
320            .batch_size
321            .unwrap_or(MAX_BATCH_SIZE)
322            .min(MAX_BATCH_SIZE);
323        let mut all_embeddings = Vec::with_capacity(texts.len());
324
325        for chunk in truncated_refs.chunks(batch_size) {
326            let embeddings = self.embed_with_retry(chunk, 3).map_err(|e| {
327                memvid_core::MemvidError::EmbeddingFailed {
328                    reason: e.to_string().into_boxed_str(),
329                }
330            })?;
331            all_embeddings.extend(embeddings);
332        }
333
334        Ok(all_embeddings)
335    }
336
337    fn is_ready(&self) -> bool {
338        self.ready.load(Ordering::Relaxed)
339    }
340
341    fn init(&mut self) -> memvid_core::Result<()> {
342        // Validate API key with a small test request
343        info!(
344            "Initializing OpenAI embedding provider with model: {}",
345            self.config.model
346        );
347
348        let test_embedding = self.embed_with_retry(&["test"], 1).map_err(|e| {
349            memvid_core::MemvidError::EmbeddingFailed {
350                reason: format!("Failed to initialize OpenAI provider: {}", e).into_boxed_str(),
351            }
352        })?;
353
354        if let Some(emb) = test_embedding.first() {
355            info!(
356                "OpenAI provider initialized: model={}, dimension={}",
357                self.config.model,
358                emb.len()
359            );
360            // Update dimension if different from expected
361            if emb.len() != self.config.dimension {
362                warn!(
363                    "Updating dimension from {} to {}",
364                    self.config.dimension,
365                    emb.len()
366                );
367            }
368        }
369
370        self.ready.store(true, Ordering::Relaxed);
371        Ok(())
372    }
373}
374
375/// Implement VecEmbedder for compatibility with existing memvid code
376impl VecEmbedder for OpenAIEmbeddingProvider {
377    fn embed_query(&self, text: &str) -> memvid_core::Result<Vec<f32>> {
378        self.embed_text(text)
379    }
380
381    fn embed_chunks(&self, texts: &[&str]) -> memvid_core::Result<Vec<Vec<f32>>> {
382        self.embed_batch(texts)
383    }
384
385    fn embedding_dimension(&self) -> usize {
386        self.dimension()
387    }
388}
389
390/// Helper to create an OpenAI provider or fall back to local
391pub fn try_openai_provider() -> Option<OpenAIEmbeddingProvider> {
392    match OpenAIEmbeddingProvider::from_env() {
393        Ok(provider) => {
394            info!("OpenAI embedding provider available");
395            Some(provider)
396        }
397        Err(e) => {
398            debug!("OpenAI provider not available: {}", e);
399            None
400        }
401    }
402}
403
404#[cfg(test)]
405mod tests {
406    use super::*;
407
408    #[test]
409    fn test_config_dimensions() {
410        assert_eq!(EmbeddingConfig::openai_large().dimension, 3072);
411        assert_eq!(EmbeddingConfig::openai_small().dimension, 1536);
412        assert_eq!(EmbeddingConfig::openai_ada().dimension, 1536);
413    }
414
415    #[test]
416    fn test_empty_api_key() {
417        let result = OpenAIEmbeddingProvider::new(String::new(), EmbeddingConfig::openai_large());
418        assert!(result.is_err());
419    }
420
421    #[test]
422    #[ignore] // Requires valid API key
423    fn test_real_embedding() {
424        let provider = OpenAIEmbeddingProvider::from_env().expect("OPENAI_API_KEY must be set");
425        let embedding = provider.embed_text("Hello, world!").expect("embed");
426        assert!(!embedding.is_empty());
427        assert_eq!(embedding.len(), 3072); // text-embedding-3-large
428    }
429}
memvid_cli/openai_embeddings.rs

memvid_cli/
openai_embeddings.rs
