Supported by
What Is SEMQ?
SEMQ is a semantic state layer for AI systems. It encodes embeddings as compact, deterministic angular codes instead of raw floating-point vectors, so the same semantic state always maps to the same symbolic representation across any hardware or provider.
These codes are stable and diffable across model versions, so you can run semq diff between snapshots the same way you'd diff a git commit and catch semantic drift before it reaches production.
SEMQ sits underneath your existing embedding pipeline, compatible with Pinecone, Weaviate, pgvector, Qdrant, or your own store. Your models and workflows don't need to change; SEMQ just encodes, stores, and diffs the resulting state.
New to SEMQ? Start with the overview
Built with SEMQ
ARI: the AI Reproducibility Index
Does an AI system give the same answer twice, even when it's configured to be "deterministic"? ARI is a public benchmark that evaluates this for embedding models; it sends the same inputs with a fresh process, under concurrent load, and a day later, encodes every output as a SEMQ code, and counts how often the codes match exactly.
Representation reproducibility ARI
Code agreement across process, concurrency and time for hosted embedding APIs, with 95% confidence intervals.
- 1gemini-embedding-001Google1.000ARI-R 1.000, plus or minus 0.000. same 1.000, proc 1.000, conc 1.000, time 1.000
- 2text-embedding-3-largeOpenAI0.845ARI-R 0.845, plus or minus 0.022. same 0.859, proc 0.862, conc 0.851, time 0.822
- 3mistral-embedMistral0.699ARI-R 0.699, plus or minus 0.028. same 0.734, proc 0.753, conc 0.791, time 0.554
- 4voyage-4-largeVoyage AI0.500ARI-R 0.500, plus or minus 0.028. same 0.571, proc 0.615, conc 0.209, time 0.676
- 5embed-v4.0Cohere0.169ARI-R 0.169, plus or minus 0.024. same 0.133, proc 0.146, conc 0.130, time 0.232
ARI v0.1-preview · 1,000 frozen inputs · refreshed 2026-09-01
What SEMQ Enables
Reproducibility
Deterministic encodings make AI pipeline outputs stable across hardware and time.
Portability
Hardware-agnostic encoding. Move workloads freely without state migration pain.
Versioning
Diff model outputs across versions like source code. Catch semantic drift early.
Observability
Structured angle sequences make AI state auditable and loggable for the first time.
Caching
Stable angle keys enable exact-match and approximate semantic caches at scale.
Compression
16× smaller state footprint. Cache more, store longer, move faster.
In The Spotlight
Changing AI math could reduce the hardware burden, researchers show
The Register on SEMQ: separating semantic meaning from data representation to cut storage and memory requirements.
Read MoreEl teorema del cuaderno: el emprendedor argentino que guardó una idea durante quince años y hoy impresiona a Silicon Valley
Andrés Mac Allister desarrolló en 2010 una observación matemática sobre vectores que hoy transformó en The SEMQ Group.
Read MoreThe SEMQ Group: cómo una startup argentina quiere hacer más eficiente y estable la inteligencia artificial
La startup fundada por Andrés Mac Allister comprime embeddings hasta 5.33 veces mediante una transformación matemática novedosa.
Read MoreNació en Pergamino, guardó una idea inútil en un cuaderno y 15 años después conquistó Silicon Valley
La historia de Andrés Mac Allister, el emprendedor argentino que incubó una idea matemática durante 15 años.
Read More