lakehouse/crates/vectord/src/store.rs

/// Vector storage as Parquet files.
/// Each embedding index is stored as: source, doc_id, chunk_idx, chunk_text, vector (binary blob).
/// Vectors are stored as raw f32 bytes for compact storage and fast loading.

use arrow::array::{ArrayRef, BinaryArray, Int32Array, RecordBatch, StringArray};
use arrow::datatypes::{DataType, Field, Schema};
use object_store::ObjectStore;
use std::sync::Arc;

use storaged::ops;

use crate::chunker::TextChunk;

/// A stored embedding — chunk text + its vector.
#[derive(Debug, Clone)]
pub struct StoredEmbedding {
    pub source: String,
    pub doc_id: String,
    pub chunk_idx: u32,
    pub chunk_text: String,
    pub vector: Vec<f32>,
}

/// Store embeddings as a Parquet file in object storage.
pub async fn store_embeddings(
    store: &Arc<dyn ObjectStore>,
    index_name: &str,
    chunks: &[TextChunk],
    vectors: &[Vec<f64>], // from embedding API (f64), we store as f32
) -> Result<String, String> {
    if chunks.len() != vectors.len() {
        return Err(format!("chunk count ({}) != vector count ({})", chunks.len(), vectors.len()));
    }

    let n = chunks.len();
    let sources: Vec<&str> = chunks.iter().map(|c| c.source.as_str()).collect();
    let doc_ids: Vec<&str> = chunks.iter().map(|c| c.doc_id.as_str()).collect();
    let chunk_idxs: Vec<i32> = chunks.iter().map(|c| c.chunk_idx as i32).collect();
    let texts: Vec<&str> = chunks.iter().map(|c| c.text.as_str()).collect();

    // Store vectors as raw f32 bytes (compact binary blob)
    let vector_bytes: Vec<Vec<u8>> = vectors.iter().map(|v| {
        v.iter().map(|&x| x as f32).flat_map(|f| f.to_le_bytes()).collect()
    }).collect();
    let vector_refs: Vec<&[u8]> = vector_bytes.iter().map(|v| v.as_slice()).collect();

    let schema = Arc::new(Schema::new(vec![
        Field::new("source", DataType::Utf8, false),
        Field::new("doc_id", DataType::Utf8, false),
        Field::new("chunk_idx", DataType::Int32, false),
        Field::new("chunk_text", DataType::Utf8, false),
        Field::new("vector", DataType::Binary, false),
    ]));

    let arrays: Vec<ArrayRef> = vec![
        Arc::new(StringArray::from(sources)),
        Arc::new(StringArray::from(doc_ids)),
        Arc::new(Int32Array::from(chunk_idxs)),
        Arc::new(StringArray::from(texts)),
        Arc::new(BinaryArray::from(vector_refs)),
    ];

    let batch = RecordBatch::try_new(schema, arrays)
        .map_err(|e| format!("RecordBatch error: {e}"))?;

    let parquet = shared::arrow_helpers::record_batch_to_parquet(&batch)?;
    let key = format!("vectors/{index_name}.parquet");
    ops::put(store, &key, parquet).await?;

    tracing::info!("stored {n} embeddings in {key}");
    Ok(key)
}

/// Load all embeddings from a vector index file.
pub async fn load_embeddings(
    store: &Arc<dyn ObjectStore>,
    index_name: &str,
) -> Result<Vec<StoredEmbedding>, String> {
    let key = format!("vectors/{index_name}.parquet");
    let data = ops::get(store, &key).await?;

    let (_, batches) = shared::arrow_helpers::parquet_to_record_batches(&data)?;

    let mut embeddings = Vec::new();
    for batch in &batches {
        let sources = batch.column(0).as_any().downcast_ref::<StringArray>()
            .ok_or("source column not string")?;
        let doc_ids = batch.column(1).as_any().downcast_ref::<StringArray>()
            .ok_or("doc_id column not string")?;
        let chunk_idxs = batch.column(2).as_any().downcast_ref::<Int32Array>()
            .ok_or("chunk_idx column not int")?;
        let texts = batch.column(3).as_any().downcast_ref::<StringArray>()
            .ok_or("chunk_text column not string")?;
        let vectors = batch.column(4).as_any().downcast_ref::<BinaryArray>()
            .ok_or("vector column not binary")?;

        for i in 0..batch.num_rows() {
            let vec_bytes = vectors.value(i);
            let vector: Vec<f32> = vec_bytes.chunks_exact(4)
                .map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]]))
                .collect();

            embeddings.push(StoredEmbedding {
                source: sources.value(i).to_string(),
                doc_id: doc_ids.value(i).to_string(),
                chunk_idx: chunk_idxs.value(i) as u32,
                chunk_text: texts.value(i).to_string(),
                vector,
            });
        }
    }

    tracing::info!("loaded {} embeddings from {key}", embeddings.len());
    Ok(embeddings)
}