Accepted at GroundLM 2026, an international workshop co-located with EMNLP 2026.
This paper introduces TripletRAG, a method that transfers useful patterns from a large language model into an SLM-based RAG system without fine-tuning. An LLM first generates triplets consisting of a synthetic question, retrieved passages, and a reference answer. Relevant triplets are then used as structured in-context examples at inference time.
Experiments on SQuAD and Natural Questions show the strongest gains for models with roughly 1B–3B parameters, improving groundedness and judge-based accuracy. Ablations suggest the benefits come not merely from few-shot formatting, but from semantic relevance and retrieved evidence.