Skip to content

Concepts

Release binding: v0.1.0-enterprise-ready · AMI ami-010806d4d3445660e · 2026-05-19

Workload representations

HX-SDP chooses a representation according to the data and the required operation. Its measured embedding-retrieval path uses low-rank latent factors plus an int8 scalar-quantized (SQ8) sidecar. Quantized Tensor Train (QTT) is an alternate ingest path for supported tensor data and callers that already hold TT-cores.

Latent factors and SQ8

For an embedding corpus with N rows and D dimensions, the latent representation approximates the corpus as Z @ V_T, where Z has shape N × r and V_T has shape r × D. The Rank Governor selects rank r from the measured structure of the dataset. A latent scan produces candidates; the SQ8 sidecar can rescore those candidates in the original dimension space.

The Performance Evidence page separates synthetic latent-only scale measurements from two-tier retrieval on named embedding corpora. Those results must not be attributed to QTT or combined into an operating point that was not measured. Compression depends on rank, precision, sidecar residency, and the stated dense-data baseline.

QTT and TT-cores

QTT represents tensor data as a chain of small linked tensors called TT-cores. The alternate ingest path accepts pre-built cores for supported operations. Its applicability and fidelity must be evaluated for the workload; the embedding measurements above do not establish a QTT performance envelope.

A TT-core is a small 3D tensor slice; the full chain of cores for one entry represents the complete data within the stated tolerance.

  • Bond dimension (rank): controls the fidelity and storage trade-off. Higher rank means higher fidelity and more storage.
  • Chain length: for a fixed physical dimension and bounded ranks, storage grows with the number of cores and the square of the rank. Rank growth can remove that advantage.

The Oracle

Every PUT passes through the Oracle, a fidelity evaluator that compares the compressed representation against the original:

Verdict Meaning
exact Effectively lossless at the configured tolerance
safe Suitable for most applications
weak Lossy but structure-preserving
passthrough Compression does not pay for this entry; stored uncompressed

The verdict is returned with every PUT response and stored in metadata. Loss is an explicit, tolerance-governed choice; it never happens silently.

When compression does not pay for an entry, passthrough preserves the uncompressed data. The Oracle's ingest verdict and the retrieval class below answer different questions: representation fidelity and query-serving behavior, respectively.

Two-tier query architecture

Approximate queries run in two tiers: a latent scan over the compressed representation produces candidates, and an int8 scalar-quantized (SQ8) sidecar rescores them. Exact-recall requirements bypass approximation entirely through the EXACT path (POST /v1/query/exact), which serves from the registered raw corpus and returns an ML-DSA-65 signed receipt.

The Rank Governor

The Rank Governor adapts the retrieval configuration automatically from measured per-query gate metrics; there is no manual tuning surface. It assigns each namespace-workload combination a retrieval class that fixes the compression aggressiveness, rerank behavior, and monitoring contract. The operator view of the class ladder is on the Retrieval Classes page.

Namespaces

Namespaces provide logical isolation within a single appliance:

  • Each tenant key is bound to a namespace ACL.
  • Keys are unique within a namespace.
  • Cross-namespace queries are not supported, by design.

Versions

Every PUT to the same key creates a new immutable version; the engine never overwrites data. GET and SERVE return the latest version by default; pass version=N for a specific one.

Tenants

Each tenant has an API key (SHA-256 hashed at rest), a namespace ACL, operator-set rate limits and CU quota, and an audit trail. Tenants are operator-defined profiles; see Admin and Tenants.

Compute Units

Every operation reports a metering weight in Compute Units so an operator can attribute usage across tenants. Weights are accounting units, not prices. See Metering.

Architecture

Client ──HTTPS 8443──> hx-gate ──loopback──> hx-engine ──loopback──> redis
                        auth, ACL,            decomposition, Oracle,
                        rate limits,          two-tier query, GPU serving
                        metering, audit

Only hx-gate binds a routable interface. The engine and Redis are loopback-only appliance internals; their addresses are listed in the Operations Guide.