Repository navigation
Veynrel 2.1.0
Discover: optional search refinement
Discover's existing Semantic Search can now refine results with Rerank. New installations leave it off and select manual mode when enabled: search first, then choose Refine results if the original results need another look. Automatic refinement after an explicit search is also available; previously enabled installations keep that behavior.
Both modes reuse one semantic candidate pool of up to 30 notes and display up to 10 results. Manual refinement does not request another query embedding. Compare Original order and After refinement locally, without another API call. A failed refinement keeps eligible original results; retrying requires another explicit action and may incur another charge.
New settings start with voyageai/rerank-3-lite. Existing explicit model IDs and independent API keys are preserved. Previously enabled settings without a model retain their effective Cohere configuration. You can enter another supported OpenRouter rerank model ID. Choosing a model does not connect automatically or supply credentials.
Refinement sends the search query and selected note fragments to OpenRouter and the selected provider, billed to your account. Prices depend on the model, provider and text volume. This is optional external processing, not a fully local or anonymous search mode.
Rerank does not guarantee improvement or retrieve notes missing from the original pool. In our small synthetic hard benchmark, it improved nDCG@10 on 11 queries, left 11 unchanged and regressed on 2. Semantic search already placed a relevant note first on 23 of 24 positive queries. The judgments were authored by an agent; these results do not establish universal model superiority.
Semantic duplicate assessment / Decisions
An optional, explicit assessment compares two previewed fragments and offers an experimental hint for manual review. It does not merge notes, delete duplicates or automatically modify Findings. A relationship between fragments does not establish that entire notes are duplicates.
Historical live validation in PR #71 recorded 16 primary matches, 2 acceptable alternatives and 2 disagreements across 20 ordered cases. Insufficient-context examples were misclassified. Confidence is a model signal, not the probability that a verdict is correct; no automatic acceptance threshold is provided. No new paid Decisions evaluation is part of this release candidate.
Reliability, privacy and compatibility
Outgoing Rerank and Decisions fragments are reconstructed from current source and checked against indexed chunk identity and content hash. Changes to sources, settings or the index invalidate pending work; late answers cannot replace a newer search session.
Automatically generated filename breadcrumbs are excluded from these outgoing fragments. Authored headings and names or paths written inside notes remain intact: this is not anonymization, and it does not change embedding, RAG or Companion payloads.
The plugin keeps its ai-knowledge-hub identity, existing user settings and semantic index format. Upgrading does not require rebuilding the index. New Rerank settings remain off/manual; old enabled configurations without an explicit mode remain automatic. See Rerank setup and privacy, Decisions and the release-candidate verification report.