The Qur’an Semantic Atlas: what 0.80 milliseconds enables

Over the past several months, we have been building something I care deeply about at Sedrah AI: the Qur’an Semantic Atlas. The idea is simple, even if the infrastructure behind it is not.

The Atlas allows users to explore all 6,236 ayahs through semantic search, thematic maps, contextual exploration, recitation audio, and multimodal analysis. But for this experience to feel truly exploratory, search cannot become the bottleneck. That led us to NVIDIA cuVS and CAGRA.

From Qur’anic text to a GPU-native semantic index

Using an NVIDIA H100 80GB GPU, we indexed 6,236 normalized 1,024-dimensional Qur’an embeddings and evaluated retrieval under both unrestricted and highly selective filtered searches. Our final cuVS CAGRA configuration achieved:

In other words, on our evaluation set, CAGRA reproduced 100% of the top-10 neighbors returned by exhaustive search, while keeping index-level retrieval below a millisecond at the median.

And we did not evaluate only easy, unfiltered queries. We tested filtering across surah selections, juz ranges, thematic thresholds, and highly selective intersections of those constraints. The measured Recall@10 remained 1.000.

Moving closer to the infrastructure

Getting there taught us something important. We initially evaluated GPU CAGRA through a higher-level vector-database path. In our setup, the CAGRA search parameters were not being applied as reliably as we needed, and recall plateaued below our target.

So we moved closer to the infrastructure. Direct integration with NVIDIA cuVS CAGRA gave us predictable control over search behavior, exact filtering behavior for our implementation, and the retrieval quality we were looking for.

That change produced something more important than a benchmark number. It changed what we could build on top of the search engine.

What the milliseconds are for

When nearest-neighbor retrieval takes around one millisecond, semantic search effectively disappears as the performance bottleneck. That gives us room to concentrate on the parts the user actually experiences:

query understanding → semantic retrieval → contextual passages → visualization → thematic exploration → recitation → audio analysis

That is where GPU infrastructure becomes interesting to me. Not because 0.80 ms is an impressive number by itself, but because of what that 0.80 ms enables.

A student can search for an idea in Arabic or English without knowing the exact Qur’anic vocabulary. A reader can discover semantically related ayahs located in distant surahs. A researcher can constrain exploration by surah, juz, or thematic relevance without sacrificing interactive responsiveness. A user can move from a semantic result into its surrounding Qur’anic context.

And because the Atlas also connects the text with ayah-level recitations, users can compare reciters and explore differences in timing, pitch, pace, and other acoustic characteristics.

A discovery system, not an interpreter

That distinction is important to everything we are building at Sedrah. We want AI to help people explore more deeply, not replace the act of learning.

This is also why evaluation matters so much to us. We are building the Atlas so that retrieval metrics, confidence, model choices, latency, filtering behavior, and benchmark results can be inspected rather than hidden behind an AI interface.

Sometimes the most valuable result of an experiment is learning what not to claim. Our multimodal experiments showed that text remains substantially stronger for semantic theme prediction, while audio is valuable for studying reciter identity, performance characteristics, and verse-level acoustic variation. That boundary is a result too.

Bigger than a benchmark

This is where the combination of NVIDIA GPU infrastructure, responsible evaluation, and a meaningful educational problem becomes particularly exciting. The H100 gave us the compute to experiment. cuVS and CAGRA gave us the retrieval infrastructure to make semantic exploration genuinely interactive.

Now our challenge is to turn that infrastructure into something much bigger than a benchmark: an experience that helps the next generation explore the Qur’an through meaning, context, relationships, visualization, and recitation, while keeping scholarship and interpretation firmly in human hands.

A special thank you to NVIDIA for the H100 platform and the cuVS ecosystem that made this level of experimentation possible.

6,236 ayahs. 1,024 dimensions. 100% measured Recall@10. 0.80 ms median retrieval.

But ultimately, the number I care about most isn’t the latency. It’s how many new connections those milliseconds can help someone discover.