bash-rag - Simple Terminal SQLite RAG Implementations

TLDR:

• cd vector-search && ./setup-and-prompt.sh to see RAG using only SQLite extensions for semantic search
• cd keyword-search && ./setup-and-prompt.sh to see RAG using only plain SQLite and its text search

Two lightweight Retrieval-Augmented Generation (RAG) examples using only terminal and SQLite with extensions, demonstrating how RAG can be implemented simply and effectively without complex infrastructure.

Vector search with SQLite extensions

Features

• Matches similar meanings better than keyword search
• No hosted database: uses embedding and vector search with only a SQLite extension
• Built on well-supported open source projects for long-term viability

Limitations

• Requires installation of extensions and downloading the embedding model, as these aren't built into SQLite.
• Keyword search may still perform better for certain queries - especially those with few keywords and limited context.
• A hosted database solution would offer easier maintenance and updates.

Usage

• Run ./setup-and-prompt.sh to run the full, slow example.
• After that, run prompt.sh to search, build the prompt and run the model.
• Note that the SQLite extensions for Mac M-chip is installed, but you will have to install it manually for other architectures.

Keyword search with plain SQLite

Features

• Just plain SQLite, no extensions or dependencies.
• Classic text search: Uses SQLite's built-in full-text search with BM25 ranking algorithm.
• Handles typos and partial matches: Uses SQLite's built-in trigram tokenizer.
• Source documents aren't stored in the database, only the index, making it space efficient.

Limitations

• Doesn't use embeddings or semantic search, so searching for "motor vehicle" wouldn't match "car". Relies on keyword search (BM25), though this is sufficient for many use cases.
• Also doesn't work well with long questions, and not just keywords.
• It would be useful to add performance comparisons with other RAG methods.

Usage

• Run ./setup-and-prompt.sh to run the full, slow example.
• After that, run prompt.sh to search, build the prompt and run the model.

Attribution

With appreciation to Jacques de Vos for the original bash-rag code and inspiration that this is forked from.