---
title: Understanding Embeddings
description: Learn how Msty makes your documents searchable
navTruncate: false
---

::MstyStudioDocsNotice
::

Embeddings are what make Knowledge Stacks "smarter" than a regular search. They help Msty understand the meaning behind your words, not just match exact phrases.

## What Makes it Special?

Think about how you'd search for a recipe:
- Regular search looks for exact words like "chocolate cake"
- Msty understands that "dessert with cocoa" means the same thing
- It can even find recipes that never use the exact words you typed

::image-plain{src="/how-to/create-knowledge-stack/search-settings-interface.webp" alt="Search settings showing similarity controls"}
::

## How Embeddings Work

Imagine translating colors into numbers:
- Red = [255, 0, 0]
- Blue = [0, 0, 255]
- Purple = [128, 0, 128]

Embeddings do the same thing with words and ideas:
- "happy" and "joyful" get similar numbers
- "bank" (money) and "bank" (river) get different numbers
- "I'm freezing" and "it's cold" get similar numbers

## Embedding Models

::image-plain{src="/how-to/create-knowledge-stack/chunk-settings-interface.webp" alt="Embedding model selection"}
::

### 1. Local Embeddings
- ✅ Completely private - nothing leaves your computer
- ✅ Free to use forever
- ✅ Works offline
- ⚠️ Slightly less accurate
- ⚠️ Uses more CPU power

### 2. Remote Embeddings (via OpenAI)
- ✅ More accurate understanding
- ✅ Less CPU usage
- ⚠️ Requires API key
- ⚠️ Small cost per use
- ⚠️ Sends text to external service

::alert{type="tip" icon="tabler:bulb"}
Start with local embeddings - they work great for most uses! Switch to remote only if you need extra accuracy.
::

## Similarity Settings

Control how strictly Msty matches content:

1. **Similarity Threshold**
   - Low: Broader matches, more results
   - Medium: Balanced relevance
   - High: Stricter matching
   - Highest: Only very close matches

2. **Number of Chunks**
   - Default: 15 chunks
   - Adjust based on needs:
     - More chunks → broader context
     - Fewer chunks → focused answers

::alert{type="info" icon="tabler:info-circle"}
These settings can be adjusted per chat session in the Knowledge Stack selector!
::

## Getting Started

1. Create a new Knowledge Stack
2. Click the gear icon
3. Choose your embedding model
4. Adjust similarity settings as needed

::alert{type="tip" icon="tabler:bulb"}
Want to fine-tune your results? Check out our guide on [Advanced Knowledge Stack Settings](/features/knowledge-stack/advanced-features).
::