Skip to content

Opt-in embedding layer for deep semantic paraphrase (feature-flagged) #12

@tomas-samek

Description

@tomas-samek

Summary

Honest residual limit from Scenario 4: deep semantic paraphrase ("how do threads coordinate with counters?" ↔ semaphore memory) produces Unknown. There is no synonym table and no embeddings.

What to do

  • Add an optional embedding layer that produces a second index alongside the trie path index.
  • Retrieval path: path-key retrieval first; on Unknown/Partial, fall back to embedding nearest-neighbor.
  • Keep embedding layer opt-in and gated behind a feature flag — the core "no embeddings" claim should still hold for the default build.

Acceptance

  • Feature-flagged build recovers the semaphore-via-counters paraphrase query.
  • Default build unchanged; all current tests pass.

Links

  • docs/design/honest_agent/progress.md Phase B residual.
  • M2 baseline comparison issue (compare fallback quality to pure-vector baseline).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions