Skip to main content

Vector Database

LLM Client includes a powerful built-in vector database system that enables you to store text embeddings and perform semantic search. This is the foundation for Retrieval-Augmented Generation (RAG).

Overview​

The vector database allows you to:

  • Store text content with embeddings
  • Perform similarity searches
  • Build knowledge bases
  • Enable RAG-enhanced conversations

Quick Start​

# 1. Generate and store embeddings
lc embed -m text-embedding-3-small -v knowledge "Machine learning is a subset of AI"

# 2. Search for similar content
lc similar -v knowledge "What is neural network programming?"

# 3. Use in conversations
lc -v knowledge "Explain the relationship between AI and ML"

Creating a Vector Database​

Vector databases are created automatically when you store your first embedding:

# This creates a database named "docs" and stores the embedding
lc embed -m text-embedding-3-small -v docs "Your first document"

Managing Databases​

List All Databases​

lc vectors list
lc v l

Output:

Vector databases:
  • knowledge (1,234 entries, model: text-embedding-3-small, provider: openai)
  • docs (567 entries, model: text-embedding-3-large, provider: openai)
  • codebase (890 entries, model: embed-english-v3.0, provider: cohere)

View Database Information​

lc vectors info <database>
lc v i <database>

Output shows:

  • Total number of entries
  • Embedding model used
  • Provider used
  • Recent entries with timestamps

Example:

lc vectors info knowledge
# Output:
# Vector database: knowledge
# Entries: 1,234
# Model: text-embedding-3-small
# Provider: openai
# 
# 📝 Recent entries:
#   1. Machine learning is a subset of AI... (2025-01-28 10:15)
#   2. Deep learning uses neural networks... (2025-01-28 10:16)
#   3. Python is popular for data science... (2025-01-28 10:17)

Delete a Database​

lc vectors delete <database>
lc v d <database>

File Embedding​

One of the most powerful features is the ability to embed entire files or directories:

Single File​

lc embed -m text-embedding-3-small -v docs -f README.md

Multiple Files​

# Using glob patterns
lc embed -m text-embedding-3-small -v docs -f "*.md"
lc embed -m text-embedding-3-small -v docs -f "src/**/*.rs"

# Multiple specific files
lc embed -m text-embedding-3-small -v docs -f "file1.txt,file2.md"

How File Embedding Works​

  1. Intelligent Chunking: Files are split into ~1200 character chunks with 200 character overlap
  2. Smart Boundaries: Chunks break at sentence, paragraph, or line boundaries
  3. Metadata Storage: Each chunk stores file path, chunk index, and total chunks
  4. Binary Detection: Automatically filters out binary files

Supported File Types​

The system automatically detects and processes text files:

Documentation: .txt, .md, .markdown, .rst, .org

Code: .rs, .py, .js, .ts, .java, .cpp, .go, .rb, .php

Web: .html, .css, .xml, .json, .yaml, .yml

Config: .toml, .ini, .cfg, .conf

Storage Details​

Location​

Vector databases are stored in platform-specific directories:

PlatformDirectory
Linux~/.config/lc/embeddings/
macOS~/Library/Application Support/lc/embeddings/
Windows%APPDATA%\lc\embeddings\

Database Structure​

Each database is a SQLite file containing:

  • vectors table: Text content, embeddings, and metadata
  • model_info table: Embedding model and provider information
  • Optimized indexes for fast similarity search

Model Consistency​

The vector database ensures consistency by:

  • Storing the model and provider used for embeddings
  • Automatically using the same model for similarity searches
  • Validating vector dimensions match the stored model

Best Practices​

1. Choose the Right Model​

  • Smaller models (e.g., text-embedding-3-small) are faster and cheaper
  • Larger models (e.g., text-embedding-3-large) provide better accuracy
  • Use the same model consistently within a database

2. Organize by Purpose​

Create separate databases for different types of content:

lc embed -v technical-docs -f "docs/technical/*.md"
lc embed -v user-guides -f "docs/guides/*.md"
lc embed -v api-reference -f "docs/api/*.md"

3. Regular Updates​

Keep your databases current:

# Re-embed updated files
lc embed -v docs -f "updated-file.md"

4. Chunk Size Considerations​

The default chunk size (1200 chars) works well for most content. For specialized needs:

  • Smaller chunks: Better for precise retrieval
  • Larger chunks: Better for maintaining context

Performance​

The vector database is optimized for:

  • Fast Retrieval: Indexed similarity searches
  • Efficient Storage: Compressed embeddings
  • Scalability: Handles thousands of documents
  • Low Memory: Streaming processing for large files

Integration with RAG​

Vector databases seamlessly integrate with chat:

# Direct prompt with context
lc -v knowledge "What do you know about Python?"

# Interactive chat with context
lc chat -v knowledge -m gpt-4

The system automatically:

  1. Generates embeddings for your query
  2. Finds similar content (similarity > 0.3)
  3. Includes relevant context in the prompt
  4. Returns an enhanced response

Troubleshooting​

"Database not found"​

  • Check spelling: lc vectors list
  • Ensure you've stored at least one embedding

"Dimension mismatch"​

  • The database uses a specific embedding model
  • Check model info: lc vectors info <database>
  • Use the same model for all operations

"No similar content found"​

  • Verify database has content: lc vectors info <database>
  • Try different search terms
  • Check if content is relevant to your query

Next Steps​