Search & RAG Settings

These settings control Atlas’s local search and retrieval system — how it breaks your notes into searchable pieces, how it finds relevant content, and how much memory it uses to do so. These are free features that run entirely on your device.

How Indexing Works

When you point Atlas at a vault, it reads every note and builds two things: a vector index (for meaning-based search) and a keyword index (for exact-match search). These two indexes are combined at search time for the best of both worlds.

The settings below let you tune how this indexing works.

Chunk Size

Range: 100 – 2000 characters | Default: 500

When indexing a note, Atlas splits it into overlapping segments called chunks. Each chunk is what gets searched and returned as a result.

Smaller chunks (100 – 300): More precise retrieval — you get the exact paragraph that’s relevant. Good for dense notes with lots of distinct facts.
Larger chunks (700 – 2000): More context per result — each hit includes more surrounding text. Good for narrative notes, meeting summaries, or long-form writing.

The default of 500 characters works well for most vaults.

Chunk Overlap

Range: 0 – 500 characters | Default: 50

The number of characters shared between adjacent chunks. Overlap helps avoid splitting a sentence or idea right at a chunk boundary.

If you find that search results sometimes feel like they’re missing the beginning or end of a thought, try increasing overlap to 100–150.

Embedding Model

Two models are available:

Model	Speed	Memory	Accuracy
BGE-small	Faster	Less (~100MB)	Good
BGE-base	Slower	More (~400MB)	Better

BGE-small is the right choice for most users, especially on older hardware or laptops without dedicated GPU memory. BGE-base produces slightly more accurate semantic matches and is worth trying if you have a modern machine.

Relevance Threshold

Range: 0.0 – 0.9 | Default: 0.3

The minimum score a result must have to be included. Results below this threshold are discarded.

Lower values (0.1 – 0.2): More results, but some may not be very relevant
Higher values (0.5 – 0.9): Only very strong matches get through

If search feels too noisy, raise the threshold. If useful notes aren’t showing up, lower it.

Hybrid Search

Default: On

Hybrid search combines semantic (meaning-based) and keyword (exact-match) search into a single ranked list. This almost always outperforms either approach alone.

Turn this off only if you’re debugging or have a specific reason to use pure vector search.

Vector Weight

Default: 0.7

The weight given to the semantic search score when combining results. Higher values favor results that match the meaning of your query. The keyword weight is automatically the remainder (1 minus the vector weight).

Text Weight

Default: 0.3

The weight given to the keyword search score. Higher values favor results that contain the exact words you searched for.

Embedding Cache Size

Default: 50,000 entries

Atlas caches computed embeddings in a local SQLite database so it doesn’t need to recompute them every time. This setting controls how many embeddings are kept in the cache before old ones are evicted.

Unless you have an unusually large vault (50,000+ notes), you don’t need to change this.

Embedding Idle Timeout

Default: 10 minutes

The ONNX embedding model is loaded into memory when needed and unloaded after this many minutes of inactivity. Unloading frees up RAM and is safe — the model reloads automatically the next time you search or chat.

Lower values (5 min): Frees memory faster, good for low-RAM machines
Higher values (30+ min): Keeps the model loaded if you search frequently, avoids reload delay

For more on how Atlas uses these search results in AI chat, see How Context Works.