RFC: Multi-Vector Distance Function Architecture

[suri-kumkaran]

RFC: Multi-Vector Distance Function Architecture

Description

This RFC proposes the architectural design for introducing multi-vector support (e.g., ColBERT/MaxSim, Chamfer distance) into the DiskANN ecosystem. This design must balance high-performance, specialized distance evaluations with strict API compatibility and standalone usability.

Goals

• Define the core mathematical operations required for multi-vector distances, starting with Chamfer and MaxSim.
• Maintain strict compatibility with DiskANN's DistanceFunctionMut trait.
• Provide a clean API that enables standalone distance function usage without requiring full index integration.
• Establish the strategy for supporting various datatypes (f32 initially, with future f16 and u8 support) and subsequent quantizations.
• Outline proposed memory layouts specifically optimized to achieve a 2x+ speedup over baseline SIMD.

Action Item: Solicit feedback from maintainers on the DistanceFunctionMut implementation and standalone API boundaries before committing to implementation.