Skip to content

feedback (validated): brain vector signal confirmed in fresh processes on v0.15.0; conceptual-query noise + embed-tier ceiling remain #552

Description

@ajianaz

Summary

External user feedback on brain search quality has been validated. On binaries \u22640.14.x, brain search silently degraded to FTS+graph in fresh processes (root cause #545). Fix shipped via #546 in v0.15.0. This issue records the validation evidence and the remaining (non-bug) quality feedback.

Validation evidence (v0.15.0)

Repo: cora-code self-index, 1,220 symbols / 134 files, embedding tier hashing-trick (256d, static).

Test Binary Result
Fresh-process conceptual query (\u0022why does search degrade to keyword only fallback in a new process\u0022) v0.15.0 20/20 results carry vector signal
Same query, same index v0.13.0 0 results (vector never fired, FTS+graph returned nothing useful)
Exact-term query (ensure_vector_cache) v0.15.0 Fusion works: top-1 = [fts, vector], score ~2.3\u00d7 vector-only hits
cora explore 'vector_search' v0.15.0 1 precise hit, instant

vecq backend note: the documented fresh-process degradation (vecq#32) is closed by #548 (vecq-core 0.3.0 keyed persistence) \u2014 so on v0.15.0 both backends should fire vector in fresh processes. vecq path not re-verified at runtime locally.

Remaining feedback (quality, not bug)

  1. Conceptual queries still surface noise in top-10 (e.g. regex constants from extract.rs rank high for a prose query about process-boundary degradation).
  2. Current static hashing 256d embed tier caps semantic depth \u2014 the vector signal is now wired end-to-end, but embedding quality is the next ceiling.

Proposed follow-ups

/cc validation session 2026-09-01

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions