Reviewing files that changed from the base of the PR and between 09068e2 and 81374af.

• lib/quantization/src/encoded_vectors.rs (1 hunks)
• lib/quantization/src/encoded_vectors_binary.rs (1 hunks)
• lib/quantization/src/encoded_vectors_pq.rs (1 hunks)
• lib/quantization/src/encoded_vectors_u8.rs (1 hunks)
• lib/segment/src/vector_storage/quantized/mod.rs (1 hunks)
• lib/segment/src/vector_storage/quantized/quantized_custom_query_scorer.rs (2 hunks)
• lib/segment/src/vector_storage/quantized/quantized_multi_custom_query_scorer.rs (1 hunks)
• lib/segment/src/vector_storage/quantized/quantized_multi_query_scorer.rs (1 hunks)
• lib/segment/src/vector_storage/quantized/quantized_multivector_storage.rs (2 hunks)
• lib/segment/src/vector_storage/quantized/quantized_query_scorer.rs (2 hunks)
• lib/segment/src/vector_storage/quantized/quantized_scorer_builder.rs (7 hunks)

• lib/segment/src/vector_storage/quantized/mod.rs

• lib/segment/src/vector_storage/quantized/quantized_multivector_storage.rs

• GitHub Check: storage-compat-test
• GitHub Check: test-snapshot-operations-s3-minio
• GitHub Check: test-low-resources
• GitHub Check: integration-tests-consensus
• GitHub Check: Basic TLS/HTTPS tests
• GitHub Check: test-consistency
• GitHub Check: test-consensus-compose
• GitHub Check: test-shard-snapshot-api-s3-minio
• GitHub Check: integration-tests
• GitHub Check: lint
• GitHub Check: rust-tests (windows-latest)
• GitHub Check: rust-tests (macos-latest)
• GitHub Check: rust-tests (ubuntu-latest)

lib/quantization/src/encoded_vectors.rs (1)

40-40: LGTM! Clean trait method addition.

The new quantized_vector_size method provides a standardized interface for accessing quantized vector sizes across all encoded vector implementations, supporting the hardware counter refactoring.

lib/quantization/src/encoded_vectors_pq.rs (1)

572-574: LGTM! Correct PQ quantized vector size implementation.

The implementation correctly returns the length of vector division, which represents the number of quantized components in Product Quantization. This aligns with the PQ algorithm where vectors are divided into chunks and each chunk is quantized to a centroid index.

lib/quantization/src/encoded_vectors_u8.rs (1)

505-507: LGTM! Correct U8 quantized vector size implementation.

The implementation correctly returns the vector dimension since in U8 scalar quantization, each dimension is quantized to a single byte, making the quantized vector size equal to the original dimension.

lib/quantization/src/encoded_vectors_binary.rs (1)

359-361: LGTM! Good delegation to existing logic.

The implementation correctly delegates to the existing get_quantized_vector_size() method, avoiding code duplication and reusing the established binary quantization size calculation logic.

lib/segment/src/vector_storage/quantized/quantized_query_scorer.rs (2)

67-69: LGTM! Hardware counter increment properly tracks IO reads.

The addition of the hardware counter increment before scoring ensures accurate tracking of vector IO reads. This aligns with the PR objective to fix the measurement issue where cached reads weren't being properly accounted for.

---

8-8: Good simplification to single-vector focus.

The import changes appropriately reflect the removal of multi-vector support, making this scorer dedicated to single-vector operations only.

lib/segment/src/vector_storage/quantized/quantized_scorer_builder.rs (3)

16-17: Correct imports for new multi-vector scorer types.

The addition of imports for QuantizedMultiCustomQueryScorer and QuantizedMultiQueryScorer properly supports the refactoring to dedicated multi-vector scorers.

---

206-206: Good parameter rename for clarity.

Renaming the parameter from quantized_storage to quantized_multivector_storage better reflects its specific purpose in multi-vector scoring contexts.

---

224-230:

✅ Verification successful

Verify the new multi-vector scorer constructor signatures.

The update to use QuantizedMultiQueryScorer with new_multi constructor looks correct. Ensure the new scorer types are properly implemented with the expected constructor signatures.

---

🏁 Script executed:

#!/bin/bash
# Description: Verify the new multi-vector scorer types have the expected constructors

# Check for QuantizedMultiQueryScorer implementation
ast-grep --pattern 'impl $_ QuantizedMultiQueryScorer $_ {
  $$$
  pub fn new_multi($_) $_ {
    $$$
  }
  $$$
}'

# Check for QuantizedMultiCustomQueryScorer implementation  
ast-grep --pattern 'impl $_ QuantizedMultiCustomQueryScorer $_ {
  $$$
  pub fn new_multi($_) $_ {
    $$$
  }
  $$$
}'

Length of output: 237

---

🏁 Script executed:

#!/bin/bash
# Search for the scorer types and their new_multi constructors
set +e

echo "Searching for QuantizedMultiQueryScorer definitions and usages..."
rg -n "QuantizedMultiQueryScorer" -A5 -B5

echo "---"
echo "Searching for QuantizedMultiCustomQueryScorer definitions and usages..."
rg -n "QuantizedMultiCustomQueryScorer" -A5 -B5

echo "---"
echo "Searching for all `pub fn new_multi` implementations..."
rg -n "pub fn new_multi" -A5 -B5

Length of output: 21301

---

Constructors Verified: new_multi Signatures Match Usage

Both QuantizedMultiQueryScorer::new_multi and QuantizedMultiCustomQueryScorer::new_multi have the correct parameter lists and return types as used in quantized_scorer_builder.rs. No changes are needed.

– lib/segment/src/vector_storage/quantized/quantized_multi_query_scorer.rs defines
pub fn new_multi(&MultiDenseVectorInternal, &TEncodedVectors, &QuantizationConfig, HardwareCounterCell) -> Self

– lib/segment/src/vector_storage/quantized/quantized_multi_custom_query_scorer.rs defines
pub fn new_multi<TOriginalQuery, TInputQuery>(TInputQuery, &TEncodedVectors, &QuantizationConfig, HardwareCounterCell) -> Self

Both match their invocations in quantized_scorer_builder.rs.

lib/segment/src/vector_storage/quantized/quantized_custom_query_scorer.rs (2)

95-97: Consistent hardware counter tracking implementation.

The hardware counter increment before scoring matches the pattern in quantized_query_scorer.rs, ensuring consistent IO measurement across all quantized scorers.

---

8-8: Clean removal of multi-vector dependencies.

The import simplification appropriately reflects the focus on single-vector operations, consistent with the overall refactoring strategy.

lib/segment/src/vector_storage/quantized/quantized_multi_custom_query_scorer.rs (1)

112-118: Good implementation of hardware counter delegation.

The delegation of scoring to the quantized multivector storage, along with passing the hardware counter, aligns well with the PR objective of moving IO measurement to the multivector storage level for accurate tracking.

lib/segment/src/vector_storage/quantized/quantized_multi_query_scorer.rs (2)

73-77: Well-structured delegation pattern.

The implementation correctly delegates scoring operations to the quantized multivector storage while passing the hardware counter. This ensures accurate IO measurement at the storage level as intended by the PR.

---

83-86: Appropriate implementation of internal scoring.

Unlike the custom query scorer, this implementation correctly delegates internal scoring to the storage layer, which is the expected behavior for standard multi-query scenarios.