Home/ Products/ Bud Cache/ Product Brief
Bud Cache overview
Product Brief · Resource Aware Attention · Semantic Caching

Bud Cache

The enterprise-grade, accuracy-first response cache for large language models — reuse answers across genuinely equivalent requests, on commodity CPUs, cutting inference cost and latency without serving a wrong answer. Built on the Resource Aware Attention family behind Bud Sentinel.

Product reference v1.0 July 2026 ~11 min read
01At a glance

Reuse only what is genuinely equivalent.

When a request arrives, Bud Cache decides — in under a millisecond, on a CPU — whether it has already produced a trustworthy answer to a genuinely equivalent request. If so, that answer returns instantly and the expensive model call is avoided. If not, the request passes through untouched and the new answer is learned for next time.

Decision accuracy92%
Wrong answers1.5%
Useful hit rate79.4%
Median hit latency0.2ms
sealed CacheBench held-out set · single CPU core · methodology in §06
What it is
An accuracy-first semantic cache in front of your model — reuses answers across genuinely equivalent requests, paraphrases and re-phrasings, not just identical strings
Inference cost falls roughly one-for-one with the hit rate, with no change to answer quality
The cache tier runs entirely on commodity CPUs — it cuts GPU demand without adding any of its own
A served hit returns before a model call has finished parsing the prompt
What it is not
An exact-match key-value cache — byte-identical requests almost never recur in natural language, so those caches barely hit
A naive semantic cache — it judges equivalence, not surface resemblance, and holds the wrong-answer rate near zero with a statistical bound
A stack layer — it sits in the serving path in front of Bud AI Foundry, in the Resource Aware Attention model family alongside Bud Sentinel
02Where it fits

In the serving path, in front of Layer 04.

Bud Cache sits in front of Bud AI Foundry's serving plane. A hit returns from cache before the model is touched; a miss flows through Foundry unchanged. Either way, the response is governed.

On a hit — the cached answer returns in ~0.2ms median from a single CPU core. The model is never called, and the GPU capacity it would have used serves new work instead.

On a miss — the request passes through to Bud AI Foundry serving unchanged. The model's new answer is cached, so the next genuinely equivalent request is a hit.

Bud Cache is not a stack layer of its own. It sits in the serving path in front of Bud AI Foundry (Layer 04), and is part of the Resource Aware Attention model family alongside Bud Sentinel, the guardrail engine inside Bud SENTRY. Hits return from cache, misses pass through to the model — and SENTRY governance applies either way.

03Capabilities, in full

Safety and savings, at once.

Six capability groups — the equivalence decision, the accuracy it delivers, the guarantee behind it, and the freshness and operability that make it deployable, expanded to the specifics an evaluator needs.

01Equivalence, not similarityEvery request decomposed into the parts that decide the answer — entities, quantities, constraints, negations, and intent · a naive semantic cache collapses all of that into one similarity score, so a flipped negation or a changed quantity reads as "nearly identical" and gets served wrongly · an answer is reused only when the signals agree two requests are genuinely equivalent, not merely alike5 signals · 1 decision
02An accuracy-first decision92% decision accuracy and a 1.5% wrong-answer rate on the sealed CacheBench held-out set · 2.4% wrong under adversarial paraphrases — where GPTCache serves 96.3% wrong · deliberately trades some coverage (79.4% vs. GPTCache's 93.3%) for a wrong-answer rate a customer-facing product can stand behind92% · 1.5% wrong
03A statistical safety guaranteeWrong-answer rate held under a target you set — with a mathematical (PAC) bound, not a hopeful threshold · one Lenient / Balanced / Strict knob trades coverage for caution to match the risk appetite of the use case, with sensible defaults out of the box · biased toward caution — when the signals disagree, the request passes through to the modelPAC-bounded
04Sub-millisecond, CPU-only~0.2ms median served hit, ≈1ms p99, on a single commodity CPU core · ~285× faster per lookup than a transformer-embedder semantic cache (~57ms), and ~7,500× faster than a model call · no GPU in the cache tier — it cuts GPU demand without adding any of its own, with a smaller energy footprint to match0.2ms · no GPU
05Freshness & correctnessTime-sensitive questions — prices, "today", "latest" — detected automatically, given short lifetimes, never reused across a day boundary · TTL configurable per request, per domain, and globally; highly non-deterministic generations are not cached at all · honours standard HTTP cache directives — no-store, no-cache, max-age, min-fresh — per request, RFC 9111 compliantRFC 9111 native
06Self-tuning, zero-configCalibrates itself to each workload and keeps the safety target on track automatically · no manual threshold tuning, and no ML team required to operate it · misses are learned — the model's new answer is cached, so coverage grows with the workloadno ML team required
04How it works

How the decision is made.

Sitting in front of your model, Bud Cache reads the parts of a request that actually decide the answer — and reuses one only when two requests are genuinely equivalent. The whole check stays under a millisecond, on CPU.

The decision flow — six steps, every request

01RequestArrives at the cache tier, before the model
02DecomposeEntities, quantities, constraints, negation, intent
03DecideEquivalence check — <1ms, CPU-only
04HitCached answer returns · ~0.2ms median
05MissPasses through to the model, untouched
06LearnThe new answer is cached for next time

Hits and misses alike stay inside the governed serving path — SENTRY policy and audit apply to every response.

Why one similarity score fails

The signals that decide whether an answer transfers are exactly the ones a single embedding distance blurs together. Two requests that differ by one negation or one quantity read as nearly identical to an embedder — and mean different things.

Naive semantic cache — one collapsed score

  • Reduces two requests to a single similarity number; above the threshold, reuse — below, don't.
  • Flipped negations, changed quantities, and swapped entities all hide inside a high score.
  • Every available threshold trades wrong answers for coverage — on the benchmark, no GPTCache setting keeps useful coverage below roughly 35% wrong answers.
  • Each lookup runs a transformer embedder: ~57ms per request.

Bud Cache — five signals, one verdict

entitiesquantitiesconstraintsnegationintent
  • Each signal is checked in its own right; an answer is reused only when they agree the requests are equivalent.
  • Operates below the whole threshold curve: 79.4% coverage at 1.5% wrong, under a 5% risk ceiling.
  • The full decision completes in under a millisecond on a commodity CPU — no transformer in the lookup path.

A missed reuse costs one model call. A wrong reuse costs trust. Bud Cache is biased toward caution by design — the asymmetry is built into the decision, not left to a threshold.

Part of the Resource Aware Attention family

Bud Cache is built on Resource Aware Attention — the CPU-native architecture behind Bud Sentinel, designed against its resource envelope from the start rather than compressed down to it. The same property that lets Sentinel classify at single-digit milliseconds lets Bud Cache judge equivalence at serving speed, on the CPUs the application already runs on.

05Deployment & compatibility

Run it like infrastructure, not like a model.

The cache tier runs entirely on commodity CPUs — cheap to run, easy to place in front of the serving path, and shipped with the security, multi-tenancy, data-governance, and operability controls an enterprise needs, on by default.

Where it runs

Commodity CPUsNo GPUSingle-core lookups

A served hit returns in ~0.2ms on a single CPU core; the equivalence check itself stays under a millisecond. Cheap to run, and easy to place anywhere in the stack.

How it integrates

Serving pathIn front of Bud AI FoundryRFC 9111

Hits skip the model; misses pass through unchanged. Standard HTTP cache directives are honoured per request, so existing caching policy carries straight over — and SENTRY governs both paths.

Network defaults

Loopback-bind defaultTenant-scoped keysRate limits

Binds to loopback until access control is configured. Tenant-scoped API keys with a separate admin key compared in constant time; per-caller rate and request-size limits.

Enterprise controls, on by default

01Security & access controlTenant-scoped API keys; a separate admin key, compared in constant time · network-safe defaults — binds to loopback until access control is configured · per-caller rate limiting and request-size limitssecure by default
02Multi-tenancy & isolationStrict per-tenant isolation — one tenant's answers can never be reused for another · verified with zero isolation violations in testing · keys scoped per-tenant or organisation-widezero violations
03GDPR-ready governanceRight to erasure (Art. 17) — tombstoned, so erasure survives backup restoration · automatic age-out of stored data and logs (Art. 5(1)(e)) · per-request never-store for sensitive traffic (Art. 5(1)(c))Art. 17 · 5(1)(c) · 5(1)(e)
04Freshness & correctnessConfigurable TTL per request, per domain, and globally · time-sensitive questions get short lifetimes automatically and are never reused across a day boundary; highly non-deterministic generations are not cached · honours no-store / no-cache / max-age / min-fresh per request — RFC 9111 compliantnever stale
05Data protection at restLeast-privilege, owner-only file permissions (0600) · runs on an encrypted volume or cloud KMS-backed disk · application-level encryption seam for envelope encryption0600 + KMS seam
06Operability & reliabilityPrometheus metrics, a health endpoint, and structured tracing · zero-downtime restarts with snapshot and warm-start · proven under shadow-replay, soak, and chaos testingproduction-proven
06Proof & methodology

Every headline number, with its basis.

All accuracy figures come from the sealed CacheBench held-out set — 400 pairs, held out from any tuning — with GPTCache 0.1.44 measured on the same set. Latency is reported on a single commodity CPU core.

92%
Decision accuracy
How it's measuredShare of reuse/pass-through decisions the cache gets right on the sealed CacheBench held-out set (400 pairs). On the same benchmark, GPTCache 0.1.44 scores 52.6% — barely better than a coin flip.
1.5%
Wrong-answer rate
How it's measuredShare of reuses that served an answer the request didn't warrant, on the same sealed set — versus 68.2% for GPTCache. Under adversarial paraphrases built to fool similarity scoring: 2.4% versus 96.3%. The rate is held under a configured ceiling with a PAC bound.
0.2ms
Median hit latency
How it's measuredServed-hit latency on a single commodity CPU core, ≈1ms at p99. Per lookup that is ~285× faster than GPTCache's transformer-embedder path (~57ms), and ~7,500× faster than a fresh model call (~1,500ms).
79.4%
Useful hit rate (coverage)
How it's measuredShare of true matches actually reused, same sealed set. GPTCache reaches 93.3% coverage only by accepting 68.2% wrong answers — Bud Cache trades that coverage away deliberately and holds the result under a 5% risk ceiling.

Head-to-head — same benchmark, both systems

Every threshold setting available to a single-similarity cache trades wrong answers for coverage; no GPTCache setting keeps useful coverage below roughly 35% wrong answers. Bud Cache operates below that curve entirely.

MetricGPTCache 0.1.44Bud CacheDirection
Decision accuracy52.6%92.0%higher is better
Wrong-answer ratethe safety metric68.2%1.5%lower is better
Adversarial wrong-answer rate96.3%2.4%lower is better
Coverage (recall)93.3%79.4%higher is better
Lookup latency, per request~57ms0.2mslower is better · ~285×

Sealed CacheBench held-out set (400 pairs). Latency on a single commodity CPU core; GPTCache's lookup cost is dominated by its transformer embedder. GPTCache's higher coverage is the direct consequence of the reuse behaviour that also produces its 68.2% wrong-answer rate — the two numbers are one trade, not two results.

07Why Bud Optional

Against the alternatives an enterprise actually weighs.

Reuse is only the first requirement. Here is how Bud Cache compares with the real options — no cache at all, a generic exact-match key-value store, and a naive semantic cache.

CapabilityNo cacheExact-match KVNaive semantic (GPTCache)Bud Cache
Reuses paraphrasesNoNoYes✓ Yes
Useful hit rateNoneVery lowHigh only if unsafe✓ High (79%)
Wrong-answer riskNoneNoneHigh (68%)✓ Near zero (1.5%)
Statistical safety guaranteeNone✓ Bounded false-hit rate
Multi-tenant isolationManualNot built in✓ Strict, built in
Auth & rate limitingYes (mature)No (library)✓ Secure by default
GDPR erasure that survives restoresDelete onlyNo✓ Yes (tombstoned)
Data / log retention age-outTTL onlyNo✓ Configurable
Encryption at restYes (enterprise)No✓ 0600 + volume/KMS seam
Freshness / time-sensitivityNoNo✓ Built in
Standard cache-control (RFC 9111)NoNo✓ Yes
Observability (metrics / health)YesLimited✓ Prometheus + health
Self-tuningn/aManual threshold✓ Automatic
CPU-only, sub-millisecond✓ YesVaries✓ Yes (0.2ms)
Equivalence, not similarity PAC-bounded safety Sub-ms, CPU-only Self-tuning, zero-config GDPR-ready by default Never stale by design
08Who it's for Optional

Where repeat questions pay for themselves.

Any workload where the same questions recur in different words is a caching workload — provided the cache never serves a wrong answer. Six places that holds.

Platforms

LLM API platforms & gateways

Absorb redundant traffic across many customers, cutting GPU spend per request — while strict tenant isolation guarantees answers are never mismatched across tenants.

Support

Customer support & FAQ assistants

The same questions recur endlessly in countless phrasings — the ideal high-redundancy, high-value workload for safe reuse.

Internal

Enterprise copilots & knowledge bots

Employees ask overlapping questions about policies, code, and docs. Reuse is high, and correctness matters.

Retrieval

Retrieval-augmented (RAG) Q&A

Repeated questions over a stable corpus reuse cleanly — and freshness rules keep answers current when the corpus changes.

Agents

Agentic & tool-using systems

Multi-step agents repeat sub-queries and planning steps. Caching them shortens chains and cuts cost per task.

Consumer

High-traffic consumer chat

Popular prompts and trending questions serve instantly from cache, freeing model capacity for the long tail.

09Go deeper & next steps

The platform behind the cache.

This brief is the reference for Bud Cache. For the platform-level argument — why serving, governance, and cost control belong on one plane — read the whitepaper, or return to the product overview.

Get started with Bud

Put your data on it.

The fastest way to see what an integrated AI operating system does for your enterprise is a proof-of-concept on your infrastructure, with your data.

01 Identify a use case where complexity, cost, or governance is a known pain point.
02 Joint discovery — Bud maps your AI pain points to platform capabilities.
03 POC in days, on your hardware, with your data.