Discovery engine & verifiable serving
Reduce a large corpus to a small, verified candidate set — then serve an open LLM behind provenance-carrying, fail-closed contracts.
What it is
Discovery reduces a large corpus or signal store to a small, verified candidate set with sparse-spectral discovery designed to avoid a full scan, packs cache-stable RAG context, and serves an open-source LLM (Qwen3.5-4B) on an NVIDIA RTX 5070 Ti — every surfaced item carrying a SHA-256 trace behind a fail-closed contract. The pipeline is discover → verify → pack → serve → trace, and each stage is measured independently with a reproducible gate. Serving results are single-GPU and scope-bounded; the integrated structured-data → LLM fusion is the roadmap objective, not a current claim.
Single RTX 5070 Ti, WSL2, eager mode. Serving numbers are single-GPU and scope-bounded. Discovery-quality evidence and serving evidence are separate evidence classes and are never cross-quoted.
By the numbers
Measured · 5070 Ti serving laneThese are serving-lane feasibility results on the vLLM / RTX 5070 Ti stack, single GPU, eager mode. Throughput and the fp8 KV-cache trade are vLLM-native platform behaviours measured in this environment, not standalone algorithmic product claims; 1.72× is a static-prefix upper bound.
How it works
Discover → verify → pack → serve → trace
Each stage has a reproducible gate; the serving stages are measured on a single NVIDIA RTX 5070 Ti with retained artifacts. The prefix-cache result is a static-prefix upper bound, and the integrated structured-data → LLM fusion is the roadmap objective, not a current claim.
Discover
Sparse-spectral discovery reduces a large corpus / signal store to a small candidate set, designed to avoid a full linear scan. On six structured-numeric domains it recovers sparse structure at recall 1.0 with full determinism.
Verify
Exact verification confirms candidates deterministically. A fail-closed output contract passes only on validated engine metadata and rejects unverifiable runs — non-engine-backed evidence cannot enter the surfaced set.
Pack
The runtime assembles cache-stable, verified RAG context for the model. Whether packed dynamic context achieves the static-prefix cache benefit is the open dynamic-prefix measurement — not yet claimed.
Serve
An OpenAI-compatible payload is served to an open LLM (Qwen3.5-4B) on the RTX 5070 Ti via vLLM. Measured: prefix caching 1.72× TTFT (static-prefix upper bound); concurrency throughput 7.3× (38.9 → 284.6 tok/s), saturating 8–16 streams.
Trace
Every turn / manifest carries the previous manifest’s SHA-256 — a tamper-evident chain over the whole session — with evidence tiers enforced. Machine closure of a review queue is benchmark-tier reclassification, not human review.
Where it works
What is measured today, and what is scope-bounded
Each capability is claimed only where it is measured, with its evidence class and scope stated. Open rows are not claimed and mark where NVIDIA validation would extend the envelope.
| Capability | Status | Path | Claim |
|---|---|---|---|
| Prefix-cache TTFT benefit | supported (measured) | serving | 1.72× on a static synthetic prefix (2.11× prefill-only); APC-off control 1.03× |
| Concurrency throughput | supported (measured) | serving | 7.3×, 38.9 → 284.6 tok/s, single GPU eager, saturating 8–16 |
| GPU encode economics | supported (measured) | embedding | 23–41× encode-only same-environment (end-to-end 8–20×) |
| fp8 KV-cache trade | supported (measured) | serving | +25% KV capacity for −6% short-request throughput |
| Regulated-document discovery | supported (measured) | discovery | recall 1.0 on the 17-query evaluation subset (in-sample); certified 5,745/6,101 |
| Structured-numeric + real sensor | supported (measured) | discovery | recall 1.0 on 6 planted-GT domains; bearing-fault family within ±2% (fixed band) |
| Dynamic-prefix / production RAG cache benefit | open — unmeasured | fallback | no benefit claimed; the roadmap gate for future validation |
| Multi-GPU / production serving scale | open — not claimed | single GPU only | single consumer GPU, eager mode; no production or multi-GPU claim |
Application fit
Why this matters beyond a single benchmark
The broader pattern — reduce a large mixed structured / unstructured store to a small verified candidate set, then contextualize it with an LLM behind auditable traces — fits many NVIDIA-accelerated workloads where the useful output is a small candidate set and the answer must carry provenance. These are application-fit examples, not measured product claims.
Regulated discovery
Legal / compliance document discovery where every surfaced item needs an auditable, reproducible trace back to source evidence.
Enterprise structured data
Sparse discovery over sensor, tabular, telemetry, and log streams feeding an LLM analysis layer.
Defense mission data
Structured sensor + report fusion with verifiable analytic traces for defense analysts.
Condition monitoring
Fault-frequency and anomaly discovery on rotating machinery and power-grid signals (real bearing and rotating-machinery datasets).
Guardrails
Not claimed
- Not a production-ready SDK or hosted service.
- Not production-scale or multi-GPU serving numbers (single consumer GPU, eager mode).
- Not a dynamic-prefix / production RAG cache-benefit claim — 1.72× is a static-prefix upper bound.
- Not throughput headroom beyond 16 concurrent streams (saturation observed at 8–16).
- Not a custom-codec KV acceleration claim — the fp8 KV result is a vLLM-native trade, separate.
- Not buyer-proof or human-reviewed evidence — machine closure is benchmark-tier reclassification, not human review.
- Not SOTA superiority of any kind, retrieval or serving.
- Not a claim that the structured-data → LLM fusion is demonstrated; it is the roadmap objective.
- CUDA-graph-mode numbers are unmeasured; all quoted serving numbers are eager mode.
Verified context for LLMs, discovered without a full scan.
See the certification and fallback discipline shared across the SparseDSP portfolio — or request an evaluation packet.
NVIDIA and RTX are trademarks and/or registered trademarks of NVIDIA Corporation. Qwen is a trademark of its respective owner; vLLM is an open-source project. References to third-party software and hardware describe the measured environment and do not imply endorsement.