Hillock v0.4: Low-Resource RAG Engine for Ollama

Hillock v0.4 is a local neuro-symbolic memory engine that can be used in conjunction with Ollama. It is optimized for building an efficient RAG (Retrieval-Augmented Generation) environment while minimizing VRAM occupancy.

Key features include:

• Fast document parsing: Instead of LLM calls, it uses a CUDA tensor classification pipeline (GLiREL + MiniLM) to convert documents into SQLite SPO triples in about 5 seconds.
• Efficient query gating: It performs query gating on the CPU in under 1ms using 10,000-dimensional VSA (Vector Symbolic Architecture) vectors, with actual Ollama calls occurring only when the gate is passed.
• Low-spec optimization: It can run using less than 1.2GB of VRAM even in a GTX 1070 environment.
• v0.4 updates: Added schema type constraints and fixed inverted relation errors.