Skip to main content

rig_core/providers/openai/
embedding.rs

1use serde::{Deserialize, Serialize};
2
3/// `text-embedding-3-large` embedding model
4pub const TEXT_EMBEDDING_3_LARGE: &str = "text-embedding-3-large";
5/// `text-embedding-3-small` embedding model
6pub const TEXT_EMBEDDING_3_SMALL: &str = "text-embedding-3-small";
7/// `text-embedding-ada-002` embedding model
8pub const TEXT_EMBEDDING_ADA_002: &str = "text-embedding-ada-002";
9
10/// Typed raw response from [`Embeddings`](super::wire::Embeddings).
11/// Missing object and model fields default to empty strings; usage is optional.
12#[derive(Debug, Clone, Serialize, Deserialize)]
13pub struct CompatibleEmbeddingResponse {
14    #[serde(default)]
15    pub object: String,
16    pub data: Vec<EmbeddingData>,
17    #[serde(default)]
18    pub model: String,
19    #[serde(default)]
20    pub usage: Option<EmbeddingUsage>,
21}
22
23/// One embedded input. Missing object and index fields default to empty and zero.
24#[derive(Debug, Clone, Serialize, Deserialize)]
25pub struct EmbeddingData {
26    #[serde(default)]
27    pub object: String,
28    pub embedding: Vec<f64>,
29    #[serde(default)]
30    pub index: usize,
31}
32
33/// Return default dimensions for a known model identifier, or `None`.
34pub(crate) fn model_dimensions_from_identifier(identifier: &str) -> Option<usize> {
35    match identifier {
36        TEXT_EMBEDDING_3_LARGE => Some(3_072),
37        TEXT_EMBEDDING_3_SMALL | TEXT_EMBEDDING_ADA_002 => Some(1_536),
38        _ => None,
39    }
40}
41
42/// The token counts an embeddings reply reports.
43#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
44pub struct EmbeddingUsage {
45    pub prompt_tokens: u64,
46    pub total_tokens: u64,
47    // Not aliased to Mistral's singular `prompt_token_details`: Mistral's
48    // embeddings reply carries *both* keys (the singular always `null`), and
49    // an alias makes serde reject the document as a duplicate field.
50    #[serde(default, skip_serializing_if = "Option::is_none")]
51    pub prompt_tokens_details: Option<EmbeddingTokensDetails>,
52    /// Mistral's top-level cached-prompt count, read when the details are absent.
53    #[serde(default, skip_serializing_if = "Option::is_none")]
54    pub num_cached_tokens: Option<u64>,
55}
56
57/// Where an embeddings reply's prompt tokens went.
58#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
59pub struct EmbeddingTokensDetails {
60    /// Tokens read from the prompt cache.
61    #[serde(default)]
62    pub cached_tokens: u64,
63}
64
65/// The provider's counts are passed through unchanged, so a reply whose
66/// `prompt_tokens` exceeds its `total_tokens` keeps both as sent.
67impl From<&EmbeddingUsage> for crate::completion::Usage {
68    fn from(usage: &EmbeddingUsage) -> Self {
69        Self {
70            input_tokens: Some(usage.prompt_tokens),
71            // What the total counts beyond the prompt, normally zero.
72            output_tokens: Some(usage.total_tokens.saturating_sub(usage.prompt_tokens)),
73            total_tokens: Some(usage.total_tokens),
74            cached_input_tokens: usage
75                .prompt_tokens_details
76                .map(|details| details.cached_tokens)
77                .or(usage.num_cached_tokens),
78            ..Self::default()
79        }
80    }
81}
82
83#[cfg(test)]
84mod tests;