Overview
GodFocus is Bible Search and RAG as a service. It answers questions against Scripture with citations, over a retrieval layer built on pgvector. The distinguishing feature is not the interface but what sits behind it: the models serving it were trained here rather than called over an API.
Objective
Build a retrieval system accurate enough to be trusted on a text where a wrong citation is a real failure, not a cosmetic one. That ruled out relying on a general-purpose model and generic embeddings, and made training and evaluation the centre of the work rather than an optimisation done at the end.
Process
Two training tracks run alongside the product. The assistant is a QLoRA fine-tune, then preference-optimised with DPO, with adapters merged and evaluated through a dedicated harness before anything is promoted. Separately, a custom cross-encoder reranker is trained with sentence-transformers and PyTorch on a purpose-built pairs dataset, then exported to ONNX so it runs inline in production instead of adding a network hop. Models are packaged with Ollama Modelfiles, and training is monitored through TensorBoard.
Technical Implementation
Leveraged advanced machine learning algorithms and AI models to enhance user experience and system performance. Implemented intelligent recommendation systems using collaborative filtering and deep learning techniques. Utilized modern AI frameworks including TensorFlow.js for real-time inference and Python-based ML pipelines for model training and optimization.
Impact
The retrieval layer returns passages ranked by a model trained on this corpus rather than by generic embedding similarity alone, and the reranker runs in-process rather than as a remote call. The platform spans six repositories covering the API, web, mobile, marketing sites and documentation.