aux4/ai-rag

Standalone document ingestion and semantic retrieval for aux4. It owns the FAISS index and document lifecycle independently of the AI agent runtime, while using aux4/embedding for the default local EmbeddingGemma vectors.

Installation

aux4 aux4 pkger install aux4/ai-rag

Commands remain backwards compatible under aux4 ai agent:

• learn indexes a document or folder.
• search retrieves indexed content.
• forget removes indexed content.

Existing FAISS stores remain compatible. ids.json retains document chunk IDs, and embedding metadata prevents mixing incompatible vectors. Legacy stores without metadata continue to use the OpenAI embedding default.

Embeddings support multiple providers: local (default, fully offline), openai, bedrock, cohere, databricks, mistral, ollama, and vertex. Pass --embeddings '{"type":"<provider>","config":{...}}' to learn/search/forget to select one; a store remembers which provider/config it was built with and rejects a mismatched one with a clear "forget and re-learn" error.