hermes-core 1.8.77

//! Dense vector merge strategies
//!
//! Every field always gets a Flat entry (raw vectors for reranking/merge).
//! Optionally, an ANN entry is also written alongside:
//! 1. O(1) cluster merge for compatible IVF-PQ payloads
//! 2. ANN rebuild with trained structures
//!
//! Raw vectors are read from source segments' lazy flat data (mmap-backed),
//! never from the document store.
//!
//! **Streaming**: ANN blobs are written immediately after serialization and
//! freed, so peak memory is bounded by one ANN blob at a time (not all blobs
//! simultaneously). Doc_id maps are streamed in chunks (384 KB per chunk).

use std::io::Write;
use std::sync::Arc;

use super::OffsetWriter;
use super::SegmentMerger;
use super::TrainedVectorStructures;
use super::doc_offsets;
use crate::Result;
use crate::directories::{Directory, DirectoryWriter};
use crate::dsl::{DenseVectorQuantization, FieldType, VectorIndexType};
use crate::segment::format::{DenseVectorTocEntry, write_dense_toc_and_footer};
use crate::segment::reader::SegmentReader;
use crate::segment::types::SegmentFiles;
use crate::segment::vector_data::{FlatVectorData, dequantize_raw};

/// Batch size for streaming vector reads (1024 vectors at a time)
const VECTOR_BATCH_SIZE: usize = 1024;

/// Chunk size for streaming flat vector bytes during merge (8 MB)
const FLAT_VECTOR_CHUNK: u64 = 8 * 1024 * 1024;

/// Chunk size for streaming doc_id+ordinal entries during merge.
/// 64K entries × 6 bytes = 384 KB per chunk (vs 60 MB unbounded).
const DOC_ID_CHUNK: usize = 64 * 1024;

/// Streams vectors from a segment's lazy flat data into an add_fn callback.
///
/// Reads vectors in batches of VECTOR_BATCH_SIZE to bound memory usage.
/// Each batch is a single range read via LazyFileSlice. Vectors are
/// dequantized to f32 regardless of storage quantization (f16, u8, f32).
///
/// Returns number of vectors added.
async fn feed_segment(
    segment: &SegmentReader,
    field: crate::dsl::Field,
    doc_id_offset: u32,
    mut add_batch: impl FnMut(&[(u32, u16)], &[f32]) -> Result<()>,
) -> crate::Result<usize> {
    let lazy_flat = match segment.flat_vectors().get(&field.0) {
        Some(f) => f,
        None => return Ok(0),
    };
    let n = lazy_flat.num_vectors;
    if n == 0 {
        return Ok(0);
    }
    let dim = lazy_flat.dim;
    let quant = lazy_flat.quantization;
    let mut count = 0;
    // Only allocate dequantize buffer for non-f32 quantizations
    let needs_dequant = quant != DenseVectorQuantization::F32;
    let mut f32_buf: Vec<f32> = Vec::new();
    let mut labels = Vec::with_capacity(VECTOR_BATCH_SIZE);

    for batch_start in (0..n).step_by(VECTOR_BATCH_SIZE) {
        let batch_count = VECTOR_BATCH_SIZE.min(n - batch_start);
        let batch_bytes = lazy_flat
            .read_vectors_batch(batch_start, batch_count)
            .await
            .map_err(crate::Error::Io)?;
        let raw = batch_bytes.as_slice();
        let batch_floats = batch_count.checked_mul(dim).ok_or_else(|| {
            crate::Error::Corruption("dense merge batch size overflows usize".into())
        })?;

        // For f32: reinterpret mmap bytes directly (zero-copy).
        // For f16/u8: dequantize into buffer.
        let vectors: &[f32] = if needs_dequant {
            f32_buf.resize(batch_floats, 0.0);
            dequantize_raw(raw, quant, batch_floats, &mut f32_buf).map_err(crate::Error::Io)?;
            &f32_buf
        } else {
            if !(raw.as_ptr() as usize).is_multiple_of(std::mem::align_of::<f32>()) {
                return Err(crate::Error::Corruption(
                    "f32 flat vector data is not 4-byte aligned".into(),
                ));
            }
            // Safety: mmap-backed vector data is page-aligned. Assertion above
            // guards against unexpected misalignment.
            unsafe { std::slice::from_raw_parts(raw.as_ptr() as *const f32, batch_floats) }
        };

        labels.clear();
        for i in 0..batch_count {
            let (doc_id, ordinal) = lazy_flat.get_doc_id(batch_start + i);
            labels.push((doc_id_offset + doc_id, ordinal));
        }
        add_batch(&labels, vectors)?;
        count += batch_count;
    }
    Ok(count)
}

/// Write a Flat entry: header + raw vectors (chunked) + doc_ids (chunked).
///
/// Doc_id map is streamed in DOC_ID_CHUNK entries (384 KB) instead of
/// buffering all at once (was 60 MB for 10M vectors).
async fn write_flat_entry(
    field_id: u32,
    dim: usize,
    total_vectors: usize,
    quantization: DenseVectorQuantization,
    segments: &[SegmentReader],
    doc_offs: &[u32],
    writer: &mut OffsetWriter,
) -> Result<()> {
    FlatVectorData::write_binary_header(dim, total_vectors, quantization, writer)?;

    // Pass 1: stream raw vector bytes in chunks
    for segment in segments {
        if let Some(lazy_flat) = segment.flat_vectors().get(&field_id) {
            let total_bytes = lazy_flat.vector_bytes_len();
            let base_offset = lazy_flat.vectors_byte_offset();
            let handle = lazy_flat.handle();
            for chunk_start in (0..total_bytes).step_by(FLAT_VECTOR_CHUNK as usize) {
                let chunk_end = chunk_start
                    .saturating_add(FLAT_VECTOR_CHUNK)
                    .min(total_bytes);
                let range_start = base_offset.checked_add(chunk_start).ok_or_else(|| {
                    crate::Error::Corruption("flat vector source offset exceeds u64".into())
                })?;
                let range_end = base_offset.checked_add(chunk_end).ok_or_else(|| {
                    crate::Error::Corruption("flat vector source range exceeds u64".into())
                })?;
                let bytes = handle
                    .read_bytes_range(range_start..range_end)
                    .await
                    .map_err(crate::Error::Io)?;
                let expected_len = usize::try_from(chunk_end - chunk_start).map_err(|_| {
                    crate::Error::Corruption("flat vector merge chunk exceeds usize".into())
                })?;
                if bytes.len() != expected_len {
                    return Err(crate::Error::Corruption(format!(
                        "flat vector merge read returned {} bytes, expected {expected_len}",
                        bytes.len()
                    )));
                }
                super::block_in_place_if_multithread(|| writer.write_all(bytes.as_slice()))?;
            }
        }
    }

    // Pass 2: stream doc_ids with offset adjustment (chunked, 384 KB per chunk)
    let mut buf = Vec::with_capacity(DOC_ID_CHUNK * 6);
    for (seg_idx, segment) in segments.iter().enumerate() {
        if let Some(lazy_flat) = segment.flat_vectors().get(&field_id) {
            let offset = doc_offs[seg_idx];
            let count = lazy_flat.num_vectors;
            for chunk_start in (0..count).step_by(DOC_ID_CHUNK) {
                buf.clear();
                let chunk_end = (chunk_start + DOC_ID_CHUNK).min(count);
                for i in chunk_start..chunk_end {
                    let (doc_id, ordinal) = lazy_flat.get_doc_id(i);
                    buf.extend_from_slice(&(offset + doc_id).to_le_bytes());
                    buf.extend_from_slice(&ordinal.to_le_bytes());
                }
                super::block_in_place_if_multithread(|| writer.write_all(&buf))?;
            }
        }
    }

    Ok(())
}

impl SegmentMerger {
    /// Merge dense vector indexes - returns output size in bytes
    ///
    /// Single-pass streaming: for each field, the ANN blob (if any) is built,
    /// serialized, written to disk, and freed immediately. Then the Flat entry
    /// is streamed. Peak memory = one ANN blob at a time, not all simultaneously.
    ///
    /// No document store reads — all vector data comes from .vectors files.
    pub(crate) async fn merge_dense_vectors<D: Directory + DirectoryWriter>(
        &self,
        dir: &D,
        segments: &[SegmentReader],
        files: &SegmentFiles,
        trained: Option<&TrainedVectorStructures>,
    ) -> Result<usize> {
        let doc_offs = doc_offsets(segments)?;

        // Collect fields that need writing (in schema field_id order)
        struct FieldInfo {
            field: crate::dsl::Field,
            dim: usize,
            total_vectors: usize,
            quantization: DenseVectorQuantization,
        }
        let mut fields_to_write: Vec<FieldInfo> = Vec::new();

        for (field, entry) in self.schema.fields() {
            if !matches!(
                entry.field_type,
                FieldType::DenseVector | FieldType::BinaryDenseVector
            ) || !(entry.indexed || entry.stored)
            {
                continue;
            }

            let dim: usize = segments
                .iter()
                .filter_map(|s| s.flat_vectors().get(&field.0).map(|f| f.dim))
                .find(|&d| d > 0)
                .unwrap_or(0);
            if dim == 0 {
                continue;
            }

            let total_vectors = segments
                .iter()
                .filter_map(|s| s.flat_vectors().get(&field.0).map(|f| f.num_vectors))
                .try_fold(0usize, |total, count| total.checked_add(count))
                .ok_or_else(|| {
                    crate::Error::Corruption(format!(
                        "flat vector count overflows usize for field {}",
                        field.0
                    ))
                })?;
            if total_vectors == 0 {
                continue;
            }

            let quantization = if entry.field_type == FieldType::BinaryDenseVector {
                DenseVectorQuantization::Binary
            } else {
                entry
                    .dense_vector_config
                    .as_ref()
                    .map(|c| c.quantization)
                    .or_else(|| {
                        segments
                            .iter()
                            .find_map(|s| s.flat_vectors().get(&field.0).map(|f| f.quantization))
                    })
                    .unwrap_or(DenseVectorQuantization::F32)
            };

            fields_to_write.push(FieldInfo {
                field,
                dim,
                total_vectors,
                quantization,
            });
        }

        if fields_to_write.is_empty() {
            return Ok(0);
        }

        let write_start = std::time::Instant::now();
        let mut writer = OffsetWriter::new(dir.streaming_writer_cold(&files.vectors).await?);
        let mut toc: Vec<DenseVectorTocEntry> = Vec::new();

        for fi in &fields_to_write {
            let field = fi.field;
            let entry = self.schema.get_field_entry(field).unwrap();
            let config = entry.dense_vector_config.as_ref();

            // ── ANN entry (written first, index_type != FLAT_TYPE) ───────
            if entry.field_type == FieldType::BinaryDenseVector {
                if let Some(index) = self
                    .try_build_binary_ivf(
                        field,
                        entry,
                        segments,
                        &doc_offs,
                        fi.total_vectors,
                        trained,
                    )
                    .await?
                {
                    let data_offset = writer.offset();
                    super::block_in_place_if_multithread(|| index.write_to(&mut writer))
                        .map_err(crate::Error::Io)?;
                    let data_size = writer.offset() - data_offset;
                    toc.push(DenseVectorTocEntry {
                        field_id: field.0,
                        index_type: crate::segment::ann_build::BINARY_IVF_TYPE,
                        offset: data_offset,
                        size: data_size,
                    });
                    // Drop the in-memory index before streaming the Flat entry.
                    drop(index);
                    let pad = (8 - (writer.offset() % 8)) % 8;
                    if pad > 0 {
                        super::block_in_place_if_multithread(|| {
                            writer.write_all(&[0u8; 8][..pad as usize])
                        })?;
                    }
                }
            } else if let Some((index_type, bytes)) = self
                .try_build_ann(field, config, segments, &doc_offs, trained)
                .await?
            {
                let data_offset = writer.offset();
                super::block_in_place_if_multithread(|| writer.write_all(&bytes))?;
                let data_size = writer.offset() - data_offset;
                toc.push(DenseVectorTocEntry {
                    field_id: field.0,
                    index_type,
                    offset: data_offset,
                    size: data_size,
                });
                // Pad to 8-byte boundary
                let pad = (8 - (writer.offset() % 8)) % 8;
                if pad > 0 {
                    super::block_in_place_if_multithread(|| {
                        writer.write_all(&[0u8; 8][..pad as usize])
                    })?;
                }
                // `bytes` dropped here — frees the ANN blob immediately
            }

            // ── Flat entry (always written, index_type = FLAT_TYPE) ──────
            let data_offset = writer.offset();
            write_flat_entry(
                field.0,
                fi.dim,
                fi.total_vectors,
                fi.quantization,
                segments,
                &doc_offs,
                &mut writer,
            )
            .await?;
            let data_size = writer.offset() - data_offset;
            toc.push(DenseVectorTocEntry {
                field_id: field.0,
                index_type: crate::segment::ann_build::FLAT_TYPE,
                offset: data_offset,
                size: data_size,
            });
            // Pad to 8-byte boundary
            let pad = (8 - (writer.offset() % 8)) % 8;
            if pad > 0 {
                super::block_in_place_if_multithread(|| {
                    writer.write_all(&[0u8; 8][..pad as usize])
                })?;
            }
        }

        // Write TOC + footer
        let toc_offset = writer.offset();
        write_dense_toc_and_footer(&mut writer, toc_offset, &toc)?;

        let output_size = writer.offset() as usize;
        super::block_in_place_if_multithread(move || writer.finish())?;
        log::info!(
            "[merge_vectors] file written: {} ({} entries) in {:.1}s",
            super::format_bytes(output_size),
            toc.len(),
            write_start.elapsed().as_secs_f64()
        );
        Ok(output_size)
    }

    /// Try to build an ANN index for a field. Returns (index_type, serialized_bytes)
    /// or None if no ANN path is available.
    ///
    /// Tries an O(1) compatible cluster merge, then rebuilds from global artifacts.
    /// Rebuild the binary IVF index for a merged binary dense vector field.
    ///
    /// Reads all packed codes from the source segments' flat storage and
    /// rebuilds with fresh k-majority centroids (training is sample-bounded
    /// via `max_train_samples`; assignment covers all vectors). Codes are the
    /// exact vectors, so the rebuild is lossless.
    async fn try_build_binary_ivf(
        &self,
        field: crate::dsl::Field,
        entry: &crate::dsl::FieldEntry,
        segments: &[SegmentReader],
        doc_offs: &[u32],
        total_vectors: usize,
        trained: Option<&TrainedVectorStructures>,
    ) -> Result<Option<crate::structures::BinaryIvfIndex>> {
        let Some(cfg) = entry.binary_dense_vector_config.as_ref() else {
            return Ok(None);
        };
        if cfg.index_type != crate::dsl::BinaryIndexType::Ivf {
            return Ok(None);
        }
        let Some(quantizer) = trained.and_then(|trained| trained.binary_quantizers.get(&field.0))
        else {
            return Ok(None);
        };

        // Fast path: every source already uses this quantizer generation.
        // Merge cluster columns directly; no centroid scoring or code rewrite.
        let mut merged: Option<crate::structures::BinaryIvfIndex> = None;
        let mut compatible = true;
        for (segment, &doc_offset) in segments.iter().zip(doc_offs) {
            let Some(crate::segment::VectorIndex::BinaryIvf(lazy)) =
                segment.vector_indexes().get(&field.0)
            else {
                compatible = false;
                break;
            };
            let Some(source) = lazy.get() else {
                compatible = false;
                break;
            };
            if source.quantizer_version != quantizer.version {
                compatible = false;
                break;
            }
            if let Some(target) = &mut merged {
                target.merge_into(source, doc_offset).map_err(|error| {
                    crate::Error::Corruption(format!(
                        "global binary IVF merge compatibility failure: {error}"
                    ))
                })?;
            } else {
                let mut first = (**source).clone();
                if doc_offset != 0 {
                    let empty = crate::structures::BinaryIvfIndex::build(
                        quantizer,
                        cfg.ivf_routing,
                        &[],
                        &[],
                    )
                    .map_err(crate::Error::Io)?;
                    let mut shifted = empty;
                    shifted
                        .merge_into(&first, doc_offset)
                        .map_err(|error| crate::Error::Corruption(error.to_string()))?;
                    first = shifted;
                }
                merged = Some(first);
            }
        }
        if compatible && let Some(index) = merged {
            log::debug!(
                "[merge_vectors] field {}: global binary IVF block-merged ({} vectors)",
                field.0,
                index.len(),
            );
            return Ok(Some(index));
        }

        let byte_len = cfg.dim.div_ceil(8);
        let code_capacity = total_vectors.checked_mul(byte_len).ok_or_else(|| {
            crate::Error::Internal("binary IVF code capacity exceeds usize".into())
        })?;
        let mut codes: Vec<u8> = Vec::with_capacity(code_capacity);
        let mut labels: Vec<(u32, u16)> = Vec::with_capacity(total_vectors);

        const CODE_BATCH: usize = 65536;
        for (seg_idx, segment) in segments.iter().enumerate() {
            let Some(lazy_flat) = segment.flat_vectors().get(&field.0) else {
                continue;
            };
            let offset = doc_offs[seg_idx];
            let n = lazy_flat.num_vectors;
            for batch_start in (0..n).step_by(CODE_BATCH) {
                let batch_count = CODE_BATCH.min(n - batch_start);
                let bytes = lazy_flat
                    .read_vectors_batch(batch_start, batch_count)
                    .await
                    .map_err(crate::Error::Io)?;
                codes.extend_from_slice(bytes.as_slice());
                for i in 0..batch_count {
                    let (doc_id, ordinal) = lazy_flat.get_doc_id(batch_start + i);
                    let merged_doc_id = doc_id.checked_add(offset).ok_or_else(|| {
                        crate::Error::Corruption(format!(
                            "binary IVF doc-id offset overflow: {doc_id} + {offset}"
                        ))
                    })?;
                    labels.push((merged_doc_id, ordinal));
                }
            }
        }

        if labels.is_empty() {
            return Ok(None);
        }

        if codes.len() != labels.len().saturating_mul(byte_len) {
            return Err(crate::Error::Corruption(format!(
                "binary IVF code/label mismatch: {} bytes for {} labels × {} bytes",
                codes.len(),
                labels.len(),
                byte_len,
            )));
        }

        let vector_count = labels.len();
        let num_clusters = quantizer.num_clusters;
        let quantizer = Arc::clone(quantizer);
        let routing = cfg.ivf_routing;
        let pool = self.background_pool.clone();
        let index = tokio::task::spawn_blocking(move || {
            let build =
                || crate::structures::BinaryIvfIndex::build(&quantizer, routing, &codes, &labels);
            let index = if let Some(pool) = pool {
                pool.install(build)
            } else {
                build()
            }?;
            // The built clusters now own the only code copy needed by the
            // serializer. Release the raw merge inputs before returning.
            drop(codes);
            drop(labels);
            Ok::<_, std::io::Error>(index)
        })
        .await
        .map_err(|error| crate::Error::Internal(format!("binary IVF build task failed: {error}")))?
        .map_err(crate::Error::Io)?;
        log::debug!(
            "[merge_vectors] field {}: binary IVF rebuilt ({} vectors, {} clusters, estimated {} bytes)",
            field.0,
            vector_count,
            num_clusters,
            index.estimated_memory_bytes(),
        );
        Ok(Some(index))
    }

    async fn try_build_ann(
        &self,
        field: crate::dsl::Field,
        config: Option<&crate::dsl::DenseVectorConfig>,
        segments: &[SegmentReader],
        doc_offs: &[u32],
        trained: Option<&TrainedVectorStructures>,
    ) -> Result<Option<(u8, Vec<u8>)>> {
        let segments_with_flat = segments
            .iter()
            .filter(|s| s.flat_vectors().contains_key(&field.0))
            .count();

        // Try an O(1) cluster merge for compatible IVF-PQ payloads.
        // Pair each ANN index with its own segment's doc offset BEFORE
        // filtering: segments without the field are dropped from the list, so
        // zipping the filtered list against the full `doc_offs` array would
        // remap every later segment's vectors with the wrong offset.
        let ivf_pq_indexes: Vec<_> = segments
            .iter()
            .enumerate()
            .filter_map(|(seg_idx, s)| {
                s.get_ivf_pq_vector_index(field)
                    .map(|index| (doc_offs[seg_idx], index))
            })
            .collect();

        let matches_published_generation = trained
            .and_then(|trained| {
                trained
                    .centroids
                    .get(&field.0)
                    .zip(trained.codebooks.get(&field.0))
            })
            .is_some_and(|(centroids, codebook)| {
                ivf_pq_indexes.iter().all(|(_, index)| {
                    index.centroids_version == centroids.version
                        && index.codebook_version == codebook.version
                })
            });
        if matches_published_generation
            && ivf_pq_indexes.len() == segments_with_flat
            && !ivf_pq_indexes.is_empty()
        {
            let refs: Vec<&crate::structures::IVFPQIndex> = ivf_pq_indexes
                .iter()
                .map(|(_, index)| index.as_ref())
                .collect();
            let offsets: Vec<u32> = ivf_pq_indexes.iter().map(|(off, _)| *off).collect();
            match crate::structures::IVFPQIndex::merge(&refs, &offsets) {
                Ok(merged) => {
                    let bytes = merged
                        .to_bytes()
                        .map_err(|e| crate::Error::Serialization(e.to_string()))?;
                    return Ok(Some((crate::segment::ann_build::IVF_PQ_TYPE, bytes)));
                }
                Err(e) => {
                    log::warn!("IVF-PQ merge failed: {}, falling back to rebuild", e);
                }
            }
        }

        // --- Try ANN rebuild from lazy flat vectors using trained structures ---
        let ann_type = trained
            .zip(config)
            .and_then(|(trained, config)| match config.index_type {
                VectorIndexType::IvfPq
                    if trained.centroids.contains_key(&field.0)
                        && trained.codebooks.contains_key(&field.0) =>
                {
                    Some(VectorIndexType::IvfPq)
                }
                _ => None,
            });

        let ann = match ann_type {
            Some(ann) => ann,
            None => {
                log::debug!(
                    "[merge_vectors] field {}: no IVF-PQ path available (trained={}, config={}, existing={}/{})",
                    field.0,
                    trained.is_some(),
                    config
                        .map(|c| format!("{:?}", c.index_type))
                        .unwrap_or_else(|| "None".into()),
                    ivf_pq_indexes.len(),
                    segments_with_flat,
                );
                return Ok(None);
            }
        };

        let trained = trained.unwrap();
        let dim = segments
            .iter()
            .filter_map(|s| s.flat_vectors().get(&field.0).map(|f| f.dim))
            .find(|&d| d > 0)
            .unwrap();
        let mut total_fed = 0usize;
        let ann_start = std::time::Instant::now();

        let (index_type, bytes) = match ann {
            VectorIndexType::Flat => unreachable!(),
            VectorIndexType::IvfPq => {
                let centroids = &trained.centroids[&field.0];
                let codebook = &trained.codebooks[&field.0];
                let routing = config.map_or(crate::dsl::IvfRoutingMode::Auto, |config| {
                    config.ivf_routing
                });
                let mut index =
                    crate::segment::ann_build::new_ivf_pq(dim, routing, centroids, codebook);

                for (seg_idx, segment) in segments.iter().enumerate() {
                    let offset = doc_offs[seg_idx];
                    let fed = feed_segment(segment, field, offset, |labels, vectors| {
                        super::block_in_place_if_multithread(|| {
                            let mut add =
                                || index.add_vectors_parallel(centroids, codebook, labels, vectors);
                            if let Some(pool) = &self.background_pool {
                                pool.install(add)
                            } else {
                                add()
                            }
                        })
                        .map_err(|error| {
                            crate::Error::Internal(format!(
                                "parallel IVF-PQ rebuild failed for field {}: {error}",
                                field.0,
                            ))
                        })
                    })
                    .await?;
                    total_fed += fed;
                    if fed > 0 {
                        log::debug!(
                            "[merge_vectors] field {} IVF-PQ: fed {} vectors from segment {} ({} total, {:.1}s)",
                            field.0,
                            fed,
                            seg_idx,
                            total_fed,
                            ann_start.elapsed().as_secs_f64()
                        );
                    }
                }

                log::info!(
                    "[merge_vectors] field {} IVF-PQ: serializing index ({} vectors, {:.1}s elapsed)",
                    field.0,
                    total_fed,
                    ann_start.elapsed().as_secs_f64()
                );
                let bytes = super::block_in_place_if_multithread(|| {
                    crate::segment::ann_build::serialize_ivf_pq(index)
                })?;
                (crate::segment::ann_build::IVF_PQ_TYPE, bytes)
            }
        };

        log::info!(
            "[merge_vectors] field {} ANN(type={}) rebuilt: {} vectors, blob={}, {:.1}s",
            field.0,
            index_type,
            total_fed,
            super::format_bytes(bytes.len()),
            ann_start.elapsed().as_secs_f64()
        );
        Ok(Some((index_type, bytes)))
    }
}

#[cfg(test)]
mod tests {
    use std::sync::Arc;

    use super::SegmentMerger;
    use crate::directories::RamDirectory;
    use crate::dsl::{DenseVectorConfig, Document, SchemaBuilder};
    use crate::index::{IndexConfig, IndexWriter};
    use crate::segment::reader::SegmentReader;
    use crate::segment::types::SegmentId;

    async fn committed_segment_ids(dir: &RamDirectory) -> Vec<String> {
        crate::index::IndexMetadata::load(dir)
            .await
            .unwrap()
            .segment_ids()
    }

    async fn newly_committed_segment(dir: &RamDirectory, known: &mut Vec<String>) -> String {
        let ids = committed_segment_ids(dir).await;
        let new: Vec<String> = ids
            .iter()
            .filter(|id| !known.contains(id))
            .cloned()
            .collect();
        assert_eq!(
            new.len(),
            1,
            "expected exactly one new segment, got {new:?}"
        );
        *known = ids;
        new.into_iter().next().unwrap()
    }

    /// Regression: the O(1) IVF-PQ cluster merge used to zip the filtered
    /// list of per-field ANN indexes against the FULL per-segment doc-offset
    /// array. With merge sources [A(has field), B(no field), C(has field)],
    /// C's vectors were remapped with B's offset instead of its own, silently
    /// persisting wrong doc ids in the merged ANN index.
    #[tokio::test]
    async fn o1_ann_cluster_merge_skips_offsets_of_segments_without_the_field() {
        let dim = 8;
        let mut sb = SchemaBuilder::default();
        let title = sb.add_text_field("title", true, true);
        let embedding = sb.add_dense_vector_field_with_config(
            "embedding",
            true,
            true,
            DenseVectorConfig::with_ivf_pq(dim, Some(1), 1),
        );
        let schema = sb.build();

        let dir = RamDirectory::new();
        let config = IndexConfig {
            merge_policy: Box::new(crate::merge::NoMergePolicy),
            num_indexing_threads: 1,
            ..Default::default()
        };
        let mut writer = IndexWriter::create(dir.clone(), schema.clone(), config)
            .await
            .unwrap();

        // Training segment: provides the sample, stays out of the merge.
        for i in 0..4 {
            let mut doc = Document::new();
            doc.add_text(title, format!("train {i}"));
            doc.add_dense_vector(embedding, vec![i as f32 + 1.0; dim]);
            writer.add_document(doc).unwrap();
        }
        writer.commit().await.unwrap();
        writer.build_vector_index().await.unwrap();
        let mut known = committed_segment_ids(&dir).await;

        // Segment A: 2 docs WITH the field (gets a per-segment IVF at flush).
        for i in 0..2 {
            let mut doc = Document::new();
            doc.add_text(title, format!("a {i}"));
            doc.add_dense_vector(embedding, vec![0.5; dim]);
            writer.add_document(doc).unwrap();
        }
        writer.commit().await.unwrap();
        let seg_a = newly_committed_segment(&dir, &mut known).await;

        // Segment B: 3 docs WITHOUT the field (no flat entry, no ANN entry).
        for i in 0..3 {
            let mut doc = Document::new();
            doc.add_text(title, format!("b {i}"));
            writer.add_document(doc).unwrap();
        }
        writer.commit().await.unwrap();
        let seg_b = newly_committed_segment(&dir, &mut known).await;

        // Segment C: 2 docs WITH the field.
        for i in 0..2 {
            let mut doc = Document::new();
            doc.add_text(title, format!("c {i}"));
            doc.add_dense_vector(embedding, vec![0.25; dim]);
            writer.add_document(doc).unwrap();
        }
        writer.commit().await.unwrap();
        let seg_c = newly_committed_segment(&dir, &mut known).await;

        // Open the merge sources in a fixed order: [A, B, C].
        let schema = Arc::new(schema);
        let mut readers = Vec::new();
        for id in [&seg_a, &seg_b, &seg_c] {
            readers.push(
                SegmentReader::open(
                    &dir,
                    SegmentId::from_hex(id).unwrap(),
                    Arc::clone(&schema),
                    16,
                )
                .await
                .unwrap(),
            );
        }
        assert!(readers[0].get_ivf_pq_vector_index(embedding).is_some());
        assert!(
            readers[1].get_ivf_pq_vector_index(embedding).is_none(),
            "segment B must not carry the dense field"
        );
        assert!(readers[2].get_ivf_pq_vector_index(embedding).is_some());

        let merged_id = SegmentId::new();
        let trained = writer.segment_manager().trained().unwrap();
        SegmentMerger::new(Arc::clone(&schema))
            .merge(&dir, &readers, merged_id, Some(trained.as_ref()))
            .await
            .unwrap();

        let merged = SegmentReader::open(&dir, merged_id, Arc::clone(&schema), 16)
            .await
            .unwrap();
        let ivf_pq = merged
            .get_ivf_pq_vector_index(embedding)
            .expect("homogeneous IVF-PQ sources must take the O(1) cluster-merge path");

        let mut doc_ids: Vec<u32> = ivf_pq
            .clusters
            .values()
            .flat_map(|cluster| cluster.doc_ids.iter().copied())
            .collect();
        doc_ids.sort_unstable();
        // A occupies merged docs 0..2, B (no vectors) 2..5, C 5..7. C's
        // vectors must be remapped with C's own offset (5), not B's (2).
        assert_eq!(
            doc_ids,
            vec![0, 1, 5, 6],
            "merged ANN doc ids must use each field-bearing segment's own offset"
        );
    }
}