fathomdb-embedder-api 0.6.1

pub type Vector = Vec<f32>;

#[derive(Clone, Debug, Eq, PartialEq)]
pub struct EmbedderIdentity {
    pub name: String,
    pub revision: String,
    pub dimension: u32,
}

impl EmbedderIdentity {
    #[must_use]
    pub fn new(name: impl Into<String>, revision: impl Into<String>, dimension: u32) -> Self {
        Self { name: name.into(), revision: revision.into(), dimension }
    }
}

#[derive(Clone, Debug, Eq, PartialEq)]
pub enum EmbedderError {
    Failed { message: String },
    Timeout,
}

pub trait Embedder: Send + Sync {
    fn identity(&self) -> EmbedderIdentity;

    fn embed(&self, input: &str) -> Result<Vector, EmbedderError>;

    /// Embed many inputs in one call. The default implementation loops [`embed`],
    /// so every backend works unchanged; backends with a true batched forward
    /// (e.g. the candle GPU path) override this to amortize per-call overhead and
    /// saturate the device (minutes -> seconds on a full-corpus embed).
    ///
    /// Contract: `embed_batch` MUST be numerically equivalent (within float
    /// tolerance) to calling [`embed`] on each input, so a caller can switch to
    /// batching WITHOUT changing the vectors written to an index. Locked by a
    /// parity test in the default-embedder crate.
    ///
    /// [`embed`]: Embedder::embed
    fn embed_batch(&self, inputs: &[&str]) -> Result<Vec<Vector>, EmbedderError> {
        inputs.iter().map(|input| self.embed(input)).collect()
    }
}