Semantic Search (RAG)

Overview

Open Note uses RAG (Retrieval-Augmented Generation) technology to implement intelligent semantic search.

Technical Architecture

Components

• ChromaDB: Local vector database
• ONNX Runtime: Local Embedding inference
• embeddinggemma-300m: Embedding model
• RRF algorithm: Result fusion ranking

Workflow

Note saved → Text chunking → Vectorization → Store in ChromaDB
                                     ↓
Search request → Query vectorization → Similarity matching → RRF fusion → Return results

Hybrid Search

Keyword Search

Traditional full-text search with exact keyword matching.

Advantages:

• Exact matching for proper nouns
• Fast
• Explainable results

Disadvantages:

• Does not understand semantics
• Cannot match synonyms

Semantic Search

Search based on vector similarity that understands search intent.

Advantages:

• Understands semantics
• Supports synonyms
• Fuzzy matching

Disadvantages:

• Computationally intensive
• Proper nouns may be less precise

RRF Fusion

Uses the Reciprocal Rank Fusion algorithm to merge two types of search results:

RRF(score) = Σ 1 / (k + rank)

where k = 60 (default)

Usage

Basic Search

Enter keywords in the search box.

Semantic Search

Use Cici Agent.

Index Management

Automatic Indexing

When a note is saved, Open Note automatically:

1. Extracts plain text (from Quill Delta JSON)
2. Chunks text (500 tokens per chunk)
3. Vectorizes chunks
4. Stores them in ChromaDB

Manually Rebuild Index

Settings → Search → Rebuild Index

Applicable when:

• Changing the Embedding model
• Index corruption
• After large imports

Index Status

View index statistics:

• Number of indexed notes
• Total number of vectors
• Index size

Embedding Configuration

Local Model (Default)

Model: embeddinggemma-300m
Location: embedding_service/models/
Memory usage: ~600MB

Chunking Strategy

Default Strategy

• Chunk size: 500 tokens
• Overlap: 50 tokens

Adjust Strategy

Advanced users can adjust it in the configuration file.

Performance Optimization

Cache

• Query result cache
• Embedding cache

Improve Accuracy

Tips

1. Use natural language: Semantic search understands natural language
2. Describe intent: Do not use only keywords
3. Combine methods: Keywords + semantic description

Examples

Good queries:

• "Notes about React component optimization"
• "Ideas I wrote about product design"

Poor queries:

• "React"
• "Product"

Troubleshooting

Q: No search results

1. Check whether the index is complete
2. Try keyword search
3. Rebuild the index

Q: Search results are irrelevant

1. Adjust the query description
2. Check the Embedding model
3. Check note index status

Q: Search is slow

1. Reduce the search scope (add category/tag filters)
2. Check hardware resources

Advanced Usage

Similarity Threshold

Adjust the minimum similarity (0-1):

• High threshold (0.8+): Precise matches
• Low threshold (0.5-): More results

Custom Embedding

Use a custom Embedding model (advanced).

Data Privacy

• All vector data is stored locally
• It is not uploaded to the cloud (when using a local model)
• Vector data can be cleared anytime