Skip to content

Add chunked embeddings with overlap and chunk stitching #13

@aurexav

Description

@aurexav

Context
Single-vector notes dilute important content in long memories. QMD uses fixed-size chunks with overlap to improve recall at boundaries.

Goal
Improve recall for long memories by indexing overlapping chunks and merging results back to the note level.

Scope

  • Add configurable chunk size and overlap.
  • Store chunk embeddings tied to note_id, sequence, and offsets.
  • Retrieve by chunk similarity and stitch adjacent chunks into a single memory result.

Acceptance Criteria

  • Long-memory recall improves without degrading short-note precision.

References

Metadata

Metadata

Assignees

No one assigned

    Labels

    kind:archArchitecture and design changes that affect system shape, boundaries, or major flows.kind:specSpecification or contract definition (APIs, schemas, invariants, query semantics).

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions