Update README.md

README.md

@@ -18,7 +18,7 @@ SAGE is a retrieval-augmented generation (RAG) pipeline that offers **non-prescr
 
 ## Introduction
 
-People routinely bring moral and existential quandaries — forgiveness, duty, grief, honesty, vocation — to the wisdom of their religious traditions, but a general-purpose LLM is a poor guide here: asked to cite scripture, it frequently **fabricates plausible-sounding verses**, misattributes them, or defaults to a [documented Western-Christian and anti-Muslim skew](https://arxiv.org/abs/2101.05783) because that content dominates pretraining. Fabricated or misattributed scripture is not a harmless error in this domain — it can distort a person's relationship with their own faith. SAGE addresses this by keeping the scripture **outside the model's weights**: it retrieves real passages from a curated, public-domain corpus, filters them to the seeker's declared tradition(s), and asks the generator to reflect only on what it was handed, citing each passage by name. On a hand-built evaluation set of 100 profile–quandary test cases, this moved the generator from citing plausible-but-off-topic verses recalled from memory to citing verses actually present in the retrieved context **62% of the time (versus 0% without retrieval)**, while a benchmark of recent-events questions ([RGB](https://arxiv.org/abs/2309.01431)) confirmed the base model answers only ~10% correctly closed-book. The results also cleanly localize the system's ceiling to **retrieval quality** rather than generation.
+People routinely bring moral and existential quandaries — forgiveness, duty, grief, honesty, vocation — to the wisdom of their religious traditions, but a general-purpose LLM is a poor guide here: asked to cite scripture, it frequently **fabricates plausible-sounding verses**, misattributes them, or defaults to a [documented Western-Christian and anti-Muslim skew](https://arxiv.org/abs/2101.05783) because that content dominates pretraining. Fabricated or misattributed scripture is not a harmless error in this domain — it can distort a person's relationship with their own faith. SAGE addresses this by keeping the scripture **outside the model's weights**: it retrieves real passages from a curated, public-domain corpus, filters them to the seeker's declared tradition(s), and asks the generator to reflect only on what it was handed, citing each passage by name. On a hand-built evaluation set of 100 profile–quandary test cases, this moved the generator from citing plausible-but-off-topic verses recalled from memory to citing verses actually present in the retrieved context **62% of the time (versus 0% without retrieval)**, while a benchmark of recent-events questions ([RGB](https://arxiv.org/abs/2309.01431)) confirmed the base model answers only ~19% correctly closed-book. The results also cleanly localize the system's ceiling to **retrieval quality** rather than generation.
 
 ## Data