5.0.0

[cybaea]

5.0.0 — The Dual-Loop Update

This release rewrites the search architecture to mimic human cognition: fast reflexes followed by deep thought. We call it Dual-Loop Search.

Alongside this, we've solved the "I can't find my note" frustration with asymmetric embeddings and fuzzy matching, ensuring you find what you're looking for even if you make a typo or use different phrasing.

Warning

Action Required: Re-Index Your Vault
Because we have changed the fundamental way vectors are calculated (switching to Asymmetric Embeddings), your old index is incompatible with this version.

1. Update the plugin.
2. Go to Settings > Explorer.
3. Click Re-index vault.

Dual-Loop Search Architecture

We have split the search experience into two distinct cognitive loops:

• Loop 1: Reflex Search (The "Spotlight")

  • Instant (<100ms): Results appear as you type.
  • Local Hybrid: Powered by a new on-device engine that blends permissive keyword matching with lightweight vector scanning.
  • Typo Tolerance: The engine now handles "fuzzy" queries. Searching for storis will correctly find stories, making the UI feel robust and forgiving.

• Loop 2: Analyst Search (The Agent)

  • Deep Reasoning: When you ask the Agent a question, it engages the "Deep" loop.
  • Deep Recall: It scans semantic "hidden threads" (bridging notes via topics, tags, author) to gather a broad set of candidates.
  • Re-Ranking: It then uses Gemini to re-rank these candidates based on true semantic relevance, ensuring it reads the right notes before answering.

Intelligence Upgrades (The "Fixes")

Three major changes ensure that "Search" actually means "Find".

• Asymmetric Embeddings: Previously, notes and queries were treated the same. Now, we explicitly distinguish them: queries are embedded as Questions and notes as Documents. This aligns with how modern models (like Gemini) are trained, drastically improving retrieval accuracy.
• Permissive Natural Language: We replaced strict keyword matching with a logical OR strategy. A query like "Where are my stories about cats?" will now find a note containing just "Cat", even if "stories" is missing.
• Expanded Vision: We increased the indexing context window (from 300 to 1000+ chars). The AI now sees your full frontmatter, tags, and header metadata, ensuring it never misses a connection because it was "cut off."
• Zero-Noise Excalidraw: We now strips away internal JSON from drawing files before indexing. Your search results are no longer polluted by metadata, matching only the text labels you actually wrote.

Architecture & Stability

• Worker Authority: The background worker is now the "Single Source of Truth," eliminating split-brain bugs where search results didn't match file contents.
• Self-Healing Vault: Changing your embedding model or search dimensions now automatically triggers the necessary re-indexing.
• Persistence Manager: A dedicated system now handles index serialization, ensuring data integrity and faster startup times.

Quality of Life

• Keyword Match Calibration: Fine-tune the balance between exact keyword matches and conceptual matches in Explorer settings.
• Tuning Resets: granular reset buttons for Advanced Settings.
• Crash loop fix: Resolved a race condition that could cause re-indexing loops on startup.

---

This discussion was created from the release 5.0.0.