Embedding Models

Choose the right embedding model for your semantic search needs.

Available Models

BGE-Small (Default)

ck --index --model bge-small .

Specifications:

• Chunk size: 400 tokens
• Model capacity: 512 tokens
• Dimensions: 384
• Size: ~80MB

Best for:

• General code search
• Fast indexing
• Smaller codebases
• Quick iteration

Pros:

• Fastest indexing
• Smallest model download
• Good general understanding
• Low memory usage

Cons:

• Smaller chunks may split large functions
• Lower context window

Nomic V1.5

ck --index --model nomic-v1.5 .

Specifications:

• Chunk size: 1024 tokens
• Model capacity: 8192 tokens
• Dimensions: 768
• Size: ~500MB

Best for:

• Large functions
• Documentation-heavy code
• Complex code structures
• Long-context understanding

Pros:

• Large context window (8K tokens)
• Better for big functions
• Handles documentation well
• Strong semantic understanding

Cons:

• Slower indexing
• Larger model download
• Higher memory usage

Jina Code

ck --index --model jina-code .

Specifications:

• Chunk size: 1024 tokens
• Model capacity: 8192 tokens
• Dimensions: 768
• Size: ~500MB

Best for:

• Code-specific searches
• Programming language understanding
• API/function signatures
• Code structure awareness

Pros:

• Specialized for code
• Understands programming concepts
• Large context window
• Strong for refactoring

Cons:

• Slower indexing
• Larger model download
• May be overkill for simple searches

Comparison Table

Feature	BGE-Small	Nomic V1.5	Jina Code
Chunk Size	400 tokens	1024 tokens	1024 tokens
Context Window	512 tokens	8K tokens	8K tokens
Dimensions	384	768	768
Download Size	~80MB	~500MB	~500MB
Index Speed	⚡⚡⚡	⚡⚡	⚡⚡
Memory Usage	Low	Medium	Medium
Code Understanding	Good	Good	Excellent
Large Functions	Fair	Excellent	Excellent

Model Selection Guide

By Project Size

Small projects (<10K LOC):

ck --index --model bge-small .
# Fast, sufficient for small codebases

Medium projects (10K-100K LOC):

ck --index --model bge-small .      # Fast iteration
# or
ck --index --model jina-code .      # Better understanding

Large projects (>100K LOC):

ck --index --model nomic-v1.5 .     # Large contexts
# or
ck --index --model jina-code .      # Code-specialized

By Code Characteristics

Many small functions:

ck --index --model bge-small .
# 400-token chunks handle small functions well

Large functions/classes:

ck --index --model nomic-v1.5 .
# 1024-token chunks avoid splitting

Documentation-heavy:

ck --index --model nomic-v1.5 .
# Better for docs and comments

Pure code focus:

ck --index --model jina-code .
# Code-specialized understanding

Switching Models

Check Current Model

ck --status .
# Shows current model and dimensions

Switch to Different Model

# Smart switch (rebuilds if needed)
ck --switch-model nomic-v1.5 .

# Force rebuild
ck --switch-model jina-code --force .

Manual Rebuild

# Remove old index
ck --clean .

# Build with new model
ck --index --model jina-code .

Model Cache Location

Models are downloaded once and cached:

• Linux/macOS – ~/.cache/ck/models/
• Windows – %LOCALAPPDATA%\ck\cache\models\
• Fallback – .ck_models/models/ in current directory

# Check cache
ls ~/.cache/ck/models/

# Clear cache (will re-download)
rm -rf ~/.cache/ck/models/

Performance Impact

Indexing Time

For 1M LOC codebase:

Model	Time	Notes
bge-small	~2 min	Fastest
nomic-v1.5	~4 min	Larger chunks
jina-code	~4 min	Code-specific processing

Search Speed

All models have similar search speed (~400-600ms). Differences are in indexing, not search.

Disk Usage

Index size (typical 1M LOC):

Model	Size	Notes
bge-small	~200MB	384 dimensions
nomic-v1.5	~400MB	768 dimensions
jina-code	~400MB	768 dimensions

Best Practices

Start Simple

# Begin with default
ck --index .
ck --sem "pattern" src/

# If results aren't great, try specialized model
ck --switch-model jina-code .

Test Different Models

# Inspect chunking without rebuilding
ck --inspect --model bge-small src/large_file.py
ck --inspect --model nomic-v1.5 src/large_file.py

# Compare results

Consider Trade-offs

• Fast iteration – Use bge-small
• Best quality – Use jina-code
• Balanced – Use nomic-v1.5

Troubleshooting

Model Download Fails

# Check network connection
ping huggingface.co

# Check disk space
df -h ~/.cache/ck/

# Manual retry
rm -rf ~/.cache/ck/models/
ck --index --model bge-small .

Index Size Too Large

# Use smaller model
ck --switch-model bge-small .

# Exclude unnecessary files
echo "*.md" >> .ckignore
echo "docs/" >> .ckignore
ck --clean .
ck --index .

Results Not Good

# Try code-specialized model
ck --switch-model jina-code .

# Adjust threshold
ck --sem --threshold 0.5 "pattern" src/

# Use hybrid search
ck --hybrid "pattern" src/

Next Steps

• Learn about semantic search
• Check configuration options
• See CLI reference
• Read basic usage