d1-3B decision model tops RAG gate test on images

The tutorial positions decision models as a solution to two problems in agentic RAG: knowing when retrieved evidence is insufficient and acting on that decision. Unlike generative models, decision models return structured judgments such as relevance scores or probabilities, which the application uses to generate a response or escalate to a human.

The build uses InsForge for backend, database, auth, and human-review workflow, EmbeddingGemma 2 for multimodal embeddings, and Liquid AI's d1-3B for scoring passages and assessing answerability. A FastAPI server in Lightning AI Studio hosts the embedding and decision models, which fit on a single NVIDIA L4 and use about 7.2 GB of GPU memory.

The evaluation compared three RAG pipelines on 75 labelled questions across four cases. The decision gate had the highest accuracy and was the only pipeline that held full performance on image-only questions, while the BGE reranker avoided wrong answers but struggled with image-based questions, escalating most of them. This suggests decision models can do more than rank passages; they can also judge whether visual evidence supports an answer, though the results are from a single tutorial test.