meilisearch/milli/src/vector/ollama.rs

// Copied from "openai.rs" with the sections I actually understand changed for Ollama.
// The common components of the Ollama and OpenAI interfaces might need to be extracted.

use std::fmt::Display;

use reqwest::StatusCode;

use super::error::{EmbedError, NewEmbedderError};
use super::openai::Retry;
use super::{DistributionShift, Embedding, Embeddings};

#[derive(Debug)]
pub struct Embedder {
    headers: reqwest::header::HeaderMap,
    options: EmbedderOptions,
}

#[derive(Debug, Clone, Hash, PartialEq, Eq, serde::Deserialize, serde::Serialize)]
pub struct EmbedderOptions {
    pub embedding_model: EmbeddingModel,
}

#[derive(
    Debug, Clone, Hash, PartialEq, Eq, serde::Serialize, serde::Deserialize, deserr::Deserr,
)]
#[deserr(deny_unknown_fields)]
pub struct EmbeddingModel {
    name: String,
    dimensions: usize,
}

#[derive(Debug, serde::Serialize)]
struct OllamaRequest<'a> {
    model: &'a str,
    prompt: &'a str,
}

#[derive(Debug, serde::Deserialize)]
struct OllamaResponse {
    embedding: Embedding,
}

#[derive(Debug, serde::Deserialize)]
pub struct OllamaError {
    error: String,
}

impl EmbeddingModel {
    pub fn max_token(&self) -> usize {
        // this might not be the same for all models
        8192
    }

    pub fn default_dimensions(&self) -> usize {
        // Dimensions for nomic-embed-text
        768
    }

    pub fn name(&self) -> String {
        self.name.clone()
    }

    pub fn from_name(name: &str) -> Self {
        Self { name: name.to_string(), dimensions: 0 }
    }

    pub fn supports_overriding_dimensions(&self) -> bool {
        false
    }
}

impl Default for EmbeddingModel {
    fn default() -> Self {
        Self { name: "nomic-embed-text".to_string(), dimensions: 0 }
    }
}

impl EmbedderOptions {
    pub fn with_default_model() -> Self {
        Self { embedding_model: Default::default() }
    }

    pub fn with_embedding_model(embedding_model: EmbeddingModel) -> Self {
        Self { embedding_model }
    }
}

impl Embedder {
    pub fn new_client(&self) -> Result<reqwest::Client, EmbedError> {
        reqwest::ClientBuilder::new()
            .default_headers(self.headers.clone())
            .build()
            .map_err(EmbedError::openai_initialize_web_client)
    }

    pub fn new(options: EmbedderOptions) -> Result<Self, NewEmbedderError> {
        let mut headers = reqwest::header::HeaderMap::new();
        headers.insert(
            reqwest::header::CONTENT_TYPE,
            reqwest::header::HeaderValue::from_static("application/json"),
        );

        let mut embedder = Self { options, headers };

        let rt = tokio::runtime::Builder::new_current_thread()
            .enable_io()
            .enable_time()
            .build()
            .map_err(EmbedError::openai_runtime_init)
            .map_err(NewEmbedderError::ollama_could_not_determine_dimension)?;

        // Get dimensions from Ollama
        let request =
            OllamaRequest { model: &embedder.options.embedding_model.name(), prompt: "test" };
        // TODO: Refactor into shared error type
        let client = embedder
            .new_client()
            .map_err(NewEmbedderError::ollama_could_not_determine_dimension)?;

        rt.block_on(async move {
            let response = client
                .post(get_ollama_path())
                .json(&request)
                .send()
                .await
                .map_err(EmbedError::ollama_unexpected)
                .map_err(NewEmbedderError::ollama_could_not_determine_dimension)?;

            // Process error in case model not found
            let response = Self::check_response(response).await.map_err(|_err| {
                let e = EmbedError::ollama_model_not_found(OllamaError {
                    error: format!("model: {}", embedder.options.embedding_model.name()),
                });
                NewEmbedderError::ollama_could_not_determine_dimension(e)
            })?;

            let response: OllamaResponse = response
                .json()
                .await
                .map_err(EmbedError::ollama_unexpected)
                .map_err(NewEmbedderError::ollama_could_not_determine_dimension)?;

            let embedding = Embeddings::from_single_embedding(response.embedding);

            embedder.options.embedding_model.dimensions = embedding.dimension();

            tracing::info!(
                "ollama model {} with dimensionality {} added",
                embedder.options.embedding_model.name(),
                embedding.dimension()
            );

            Ok(embedder)
        })
    }

    async fn check_response(response: reqwest::Response) -> Result<reqwest::Response, Retry> {
        if !response.status().is_success() {
            // Not the same number of possible error cases covered as with OpenAI.
            match response.status() {
                StatusCode::TOO_MANY_REQUESTS => {
                    let error_response: OllamaError = response
                        .json()
                        .await
                        .map_err(EmbedError::ollama_unexpected)
                        .map_err(Retry::retry_later)?;

                    return Err(Retry::rate_limited(EmbedError::ollama_too_many_requests(
                        OllamaError { error: error_response.error },
                    )));
                }
                StatusCode::SERVICE_UNAVAILABLE => {
                    let error_response: OllamaError = response
                        .json()
                        .await
                        .map_err(EmbedError::ollama_unexpected)
                        .map_err(Retry::retry_later)?;
                    return Err(Retry::retry_later(EmbedError::ollama_internal_server_error(
                        OllamaError { error: error_response.error },
                    )));
                }
                StatusCode::NOT_FOUND => {
                    let error_response: OllamaError = response
                        .json()
                        .await
                        .map_err(EmbedError::ollama_unexpected)
                        .map_err(Retry::give_up)?;

                    return Err(Retry::give_up(EmbedError::ollama_model_not_found(OllamaError {
                        error: error_response.error,
                    })));
                }
                code => {
                    return Err(Retry::give_up(EmbedError::ollama_unhandled_status_code(
                        code.as_u16(),
                    )));
                }
            }
        }
        Ok(response)
    }

    pub async fn embed(
        &self,
        texts: Vec<String>,
        client: &reqwest::Client,
    ) -> Result<Vec<Embeddings<f32>>, EmbedError> {
        // Ollama only embedds one document at a time.
        let mut results = Vec::with_capacity(texts.len());

        // The retry loop is inside the texts loop, might have to switch that around
        for text in texts {
            // Retries copied from openai.rs
            for attempt in 0..7 {
                let retry_duration = match self.try_embed(&text, client).await {
                    Ok(result) => {
                        results.push(result);
                        break;
                    }
                    Err(retry) => {
                        tracing::warn!("Failed: {}", retry.error);
                        retry.into_duration(attempt)
                    }
                }?;
                tracing::warn!(
                    "Attempt #{}, retrying after {}ms.",
                    attempt,
                    retry_duration.as_millis()
                );
                tokio::time::sleep(retry_duration).await;
            }
        }

        Ok(results)
    }

    async fn try_embed(
        &self,
        text: &str,
        client: &reqwest::Client,
    ) -> Result<Embeddings<f32>, Retry> {
        let request = OllamaRequest { model: &self.options.embedding_model.name(), prompt: text };
        let response = client
            .post(get_ollama_path())
            .json(&request)
            .send()
            .await
            .map_err(EmbedError::openai_network)
            .map_err(Retry::retry_later)?;

        let response = Self::check_response(response).await?;

        let response: OllamaResponse = response
            .json()
            .await
            .map_err(EmbedError::openai_unexpected)
            .map_err(Retry::retry_later)?;

        tracing::trace!("response: {:?}", response.embedding);

        let embedding = Embeddings::from_single_embedding(response.embedding);
        Ok(embedding)
    }

    pub fn embed_chunks(
        &self,
        text_chunks: Vec<Vec<String>>,
    ) -> Result<Vec<Vec<Embeddings<f32>>>, EmbedError> {
        let rt = tokio::runtime::Builder::new_current_thread()
            .enable_io()
            .enable_time()
            .build()
            .map_err(EmbedError::openai_runtime_init)?;
        let client = self.new_client()?;
        rt.block_on(futures::future::try_join_all(
            text_chunks.into_iter().map(|prompts| self.embed(prompts, &client)),
        ))
    }

    // Defaults copied from openai.rs
    pub fn chunk_count_hint(&self) -> usize {
        10
    }

    pub fn prompt_count_in_chunk_hint(&self) -> usize {
        10
    }

    pub fn dimensions(&self) -> usize {
        self.options.embedding_model.dimensions
    }

    pub fn distribution(&self) -> Option<DistributionShift> {
        None
    }
}

impl Display for OllamaError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "{}", self.error)
    }
}

fn get_ollama_path() -> String {
    // Important: Hostname not enough, has to be entire path to embeddings endpoint
    std::env::var("MEILI_OLLAMA_URL").unwrap_or("http://localhost:11434/api/embeddings".to_string())
}