Veynrel ranked the right notes first in 3 out of 3 queries

Conventional note search breaks down in an all-too-familiar situation: you remember the idea, but not the exact wording or filename. A query such as “how to test yourself without notes” does not necessarily contain any words from a note titled “Practicing an Oral Answer.” A string-matching search may therefore return nothing even though the relevant material is close at hand.

The Veynrel plugin for Obsidian addresses this with semantic search. It converts note fragments into numerical representations—embeddings—in advance, stores the index, and then compares it with a question written in ordinary language. In a small experiment by the author, BGE-M3 ranked the correct note first for all three queries. That is a useful result, but not yet a full comparative benchmark.

What this looks like in a simple example

For the test, the author assembled a study vault containing 16 fictional notes: exam questions, exercises, and preparation plans. The author then entered three questions worded differently from the relevant files. For example, a query about a graph that “goes up and down” was expected to lead to an exam question about a function increasing and decreasing.

Obsidian's built-in search did not find that file using the two selected words. Local MiniLM ranked the correct note fourth, while BGE-M3 ranked it first. An even more revealing query asked why rearranging characters changes a password: the source note did not contain the word “rearranging” at all. Keyword search was powerless, while semantic search connected the description of changing character order with the question.

Three configurations were tested. MiniLM ran locally through ONNX and placed the correct files at positions 1, 4, and 1. Cloud-based BGE-M3 through OpenRouter and local BGE-M3 through Ollama both produced positions 1, 1, and 1. A separate Ollama index was built from 19 vectors with 1,024 components each. With this configuration, the notes never left the computer.

More than search

Veynrel includes a Health mode that checks the vault's structure locally. In the demonstration folder, it detected a deliberately planted link to a nonexistent note and an exact copy of another file. This check requires neither a language model nor an API key.

Recall mode creates review cards from questions and answers written in Markdown. The user reveals the answer and rates how well they remembered it, after which the plugin schedules the next review. Studio stores text-processing workflows, while the Companion integration allows an AI agent to read synchronized copies of notes and suggest changes. Applying edits remains under the user's control.

What to check before adopting it

Three queries and 16 prepared notes do not prove that BGE-M3 will win on a production vault containing thousands of heterogeneous files. The author explicitly notes that rankings may change in another folder. Numerical similarity scores from different models also cannot be compared directly. Rerank did not improve the top result in the example shown because the correct note was already at the top.

Privacy is another consideration. When using OpenRouter, note fragments are sent to an external provider. For confidential data, it is more sensible to start with a local configuration: install Ollama, download bge-m3, enter http://127.0.0.1:11434, and build a new index. The model occupies about 1.2 GB, so disk space and memory should be checked beforehand.

The practical conclusion is straightforward: semantic search is useful when people remember meaning better than filenames. But it is too early to choose a model based on three impressive hits. First, collect 20–50 real queries from your own history, mark the expected notes, and compare their result positions. That will show whether the new index justifies its memory use, latency, and the potential transfer of data to the cloud.