Evaluate & Benchmark
Semantic Embedding Space Auditor
UMAP + HDBSCAN clustering with five semantic health checks: outliers, boundary ambiguity, polarity mismatch, and global collapse.
The problem
Embedding quality silently degrades over time, and teams have no way to catch drift before it breaks RAG retrieval or dedup in production.
- Search results get worse for months before anyone connects it to the embeddings.
- A retrain quietly changes the vector space and nothing catches it before it ships.
- Dedup starts missing obvious duplicates and no one can say exactly when it broke.
What NEO built
NEO built an audit framework running five semantic health checks (outlier detection, boundary ambiguity, polarity mismatch, cluster consistency, and drift vs. baseline) with exportable, CI-ready reports.
The result
5 semantic checks, reproducible
Surfaces reproducible regression flags across five semantic checks before a bad embedding update reaches production.
From the blog · 8 min
EmbedAudit: A CLI Tool for Auditing Semantic Embedding Spaces
NEO built EmbedAudit: UMAP, HDBSCAN, and five semantic checks (outliers, boundary ambiguity, isolated tokens, polarity mismatch, global collapse) with JSON/Markdown reports and reproducible audits.
Try this in your workspace
Paste this into NEO chat to kick off the same workflow on your own data.
Audit my embedding space for outliers, boundary ambiguity, and drift against a baseline, and give me a CI-ready report I can run on every retrain.
Paste it in · review the plan · get the diff
Get NEO
