How your node searches by meaning

Imagine your node employs a librarian. Every time you give it a page of text, the librarian reads it and writes down a fingerprint: a list of 384 numbers that captures what the page means. Pages about "resetting passwords" and pages about "recovering logins" get fingerprints that are nearly identical, even though the words differ. When you ask a question, your question gets a fingerprint too, and your node returns the pages whose fingerprints sit closest to it.

The librarian is a small machine-learning model: Snowflake's arctic-embed-s, an open-source (Apache-2.0) embedding model from a US vendor. It runs as a compact, quantized artifact directly inside the platform process, on ordinary CPUs. There is no GPU requirement and no external service involved.

Questions and documents are treated differently on purpose. The model was trained to expect search queries in a particular framing, so your node applies that framing to queries automatically. You never see it, but it measurably improves what comes back.

The embedding model is pinned to one exact version, and its checksum is recorded in the platform source code. Every time a process starts the embedder, it re-verifies the model file against that checksum before serving a single request. A swapped, corrupted, or tampered model does not degrade quietly; it refuses to load, and every successful, verified load is written to the process logs for audit.

Every stored fingerprint is stamped with the model that produced it. Fingerprints from two different models look superficially alike, but comparing them is like comparing meters to miles, so search only ever compares fingerprints carrying the same stamp. When the platform adopts a newer model, re-ingesting your content re-embeds it under the new stamp automatically. Your documents never need to be uploaded again.

The same rule is why every node runs the same embedder. Nodes in a mesh compare fingerprints when they federate queries, and that only works when everyone speaks the same fingerprint language. Enterprise deployments with fully isolated nodes can override the model at their own boundary, because their fingerprints never cross it.

The model is small by design so it runs on the modest CPU nodes the platform provisions. It is strongest in English, and it reads at most 512 tokens of a passage (roughly 350 to 400 words); longer passages are truncated and the truncation is logged. Keeping ingested chunks inside that window gives the best results.