Glossary¶
Terms used across the Private AI Platform Kit docs, runbooks, charts, and source. Each definition describes how this repo uses the term, not the concept in the abstract, and links the authoritative document, runbook, or file where one exists. Entries are alphabetized.
A¶
Admission control: Gateway-side request validation applied before a request reaches Ollama or
vLLM. The admission block in
deploy/charts/inference-gateway/values.yaml
caps message count (maxMessages), prompt size (maxPromptChars), completion tokens
(maxCompletionTokens), temperature range, and streaming (allowStreaming), plus the prompt
secret-detection policy. Rejections return HTTP 400 and are counted by
inference_gateway_admission_rejections_total. See the
Guardrails
and Traceability sandbox
runbooks.
Agent sandbox (workspace runtime): The hardened kubernetes-sigs/agent-sandbox Sandbox that
every coding-agent workspace runs on (ADR 0010):
non-root, read-only root filesystem, no ambient ServiceAccount token, and an optional
kernel-isolation runtime class (sandbox.runtimeClassName). Each workspace is bound 1:1 to the
sandbox id the gateway sees as X-Sandbox-ID. The controller is a platform prerequisite (synced
as the agent-sandbox-controller Application, or installed with make agent-sandbox-install),
and make agent-sandbox-smoke verifies the runtime contract. Distinct from the traceable sandbox
namespace (see below). See
Agent workspaces.
B¶
Budget window / estimated-token budget: The rolling time window (windowSeconds, default 86400)
over which a sandbox's usage is accumulated, and the estimated-token ceiling enforced within it.
Estimated tokens are computed as ceil(prompt chars / estimatedCharsPerToken) + requested max_tokens
(falling back to the admission completion-token ceiling when max_tokens is omitted). An overage
returns HTTP 400 with sandbox_token_budget_exceeded. See
Budget controls.
Burn-rate alert: A Prometheus alert that fires when the inference gateway consumes its
availability error budget too fast. InferenceGatewayErrorBudgetFastBurn and
...SlowBurn in
deploy/observability/alerts/ai-platform-alerts.yaml
watch the 99.5 percent objective on short and long horizons. See
SLO and error budget.
C¶
Canary vs shadow (progressive delivery): Two gateway routing behaviors used to roll out a model
or backend change safely. Canary sends a weighted fraction of live traffic to a candidate route and
counts it with inference_gateway_canary_routed_total. Shadow mirrors requests fire-and-forget to a
candidate route without returning its response to the caller
(inference_gateway_shadow_requests_total). Both are first-class in the gateway; see
ADR 0005 and Architecture.
catalogRef: The field that ties a requested egress destination back to an approved entry in the
egress catalog. A tenant or workspace values file references an approved
platform/network/egress-catalog.yaml entry by catalogRef (alongside cidr and ports); the
validator rejects any external destination without a matching, non-expired catalog entry. See
Egress governance.
Context precision (RAGAS-style): See faithfulness / context precision.
Cost-center: One of the required attribution labels (platform.ai/cost-center) stamped on
namespaces, pods, and tenant specs so Prometheus, OpenCost-style reporting, logs, and evidence packs
attribute usage to the same owner. Kept stable across upgrades so cost history stays comparable. See
Quota and chargeback.
Cross-encoder reranker: An optional second retrieval stage that reorders over-fetched candidates
by joint query-document relevance, raising precision where dense retrieval is weakest (paraphrase,
synonymy, multi-hop). Configured via retrieval.reranker in the RAG chart; the default provider is
none, and an openai-compatible provider calls a Cohere/Jina/TEI-style /rerank endpoint. A
reranker outage is non-fatal; the first-stage hybrid ranking is kept. See
src/rag-service/app/reranker.py.
D¶
dataClassification: A model catalog field recording the sensitivity of the data a model is
approved to handle. Validated against {public, internal, confidential, restricted} in
scripts/model-catalog.py
and must match the model's provenance record. Distinct from riskTier (see below): dataClassification
answers "how sensitive is the data?", riskTier answers "how much harm if the model misbehaves?".
Digest: See model provenance / immutableRef / digest.
E¶
Egress catalog: The single reviewed allowlist of external network destinations, at
platform/network/egress-catalog.yaml. Agent and tenant namespaces are default-deny; any external
CIDR must be an approved, non-expired catalog entry (owner, environments, expiry, use cases, data
classification, CIDRs, ports) referenced from values by catalogRef. See
Egress governance.
Error budget / SLO: An SLO (service-level objective) is a target defined in
platform/slo/objectives.yaml: gateway error rate, p95/p99 latency, eval pass rate, restore-drill
pass rate, and coding-agent readiness. The error budget is the allowed amount of SLO violation before
burn-rate alerts fire; a failed objective is treated as a customer-readiness blocker. See
SLO and error budget.
Evidence pack: The bundle of static readiness controls plus pointers to the latest generated
operational artifacts, produced by make evidence (add LIVE=1 for live Kubernetes checks). Reports
are written under results/evidence/; the Markdown report is the customer-facing summary and the JSON
report feeds automation and audit. See
Evidence pack.
F¶
Faithfulness / context precision (RAGAS-style): Deterministic, offline generation-grounding
proxies scored by the RAG eval only on cases that ship a ground-truth answer. Context precision
measures how much of what retrieval returned was on-topic; faithfulness measures how much of the
expected answer is supported by the retrieved context. They are floored by minContextPrecision and
minFaithfulness in
platform/evals/rag-retrieval-suite.yaml
and computed in
scripts/rag-eval.py.
H¶
Hybrid retrieval: A recall-oriented first-stage ranking that combines dense (embedding) similarity
with lexical term overlap. In the Qdrant profile, candidates are over-fetched and reordered by a hybrid
score; retrieval.lexicalWeight controls the blend (0 reproduces pure dense ranking). Feeds the
optional cross-encoder reranker. See
src/rag-service/app/retriever.py
and the Vector RAG
runbook.
I¶
immutableRef: See model provenance / immutableRef / digest.
M¶
MIG (Multi-Instance GPU): NVIDIA's hardware partitioning of one A100/H100 into isolated GPU
instances, configured through the NVIDIA GPU Operator. To use it, request the MIG resource name (e.g.
nvidia.com/mig-1g.10gb) in accelerator.resourceName with accelerator.count: 1. It is the
memory-isolated way to share a GPU across tenants (unlike time-slicing). The device-plugin/operator
config is operator-owned. See
GPU capacity.
Model card: The human-readable companion to a model's machine-checked catalog entry and
provenance record. Every status: approved model must have a card under
platform/model-catalog/model-cards/;
cards introduce no new facts; every field is copied from the governed YAML, and make model-check
fails if an approved model is missing one.
Model provenance / immutableRef / digest: The attested origin of a served model artifact, recorded
in platform/governance/model-provenance.yaml. Each approved model carries a sourceUri, an
immutableRef (an immutable reference pinned by a SHA-256 digest), a digest scope and verification
command, and license/risk/classification metadata matching the catalog. The bundled lab uses
source-reference digests; customer production replaces them with real registry/object-store artifact
digests, and the vLLM model.revision should be pinned to the attested immutableRef so the runtime
artifact cannot drift. See
Model provenance.
O¶
Output guardrail: A response-path control that inspects the model's completion before it is
returned or cached, contributing controls for OWASP LLM02:2025 (sensitive information disclosure)
and LLM05:2025 (improper output handling); the input-side prompt secret detection cannot catch a
secret the model emits. Modes are
flag, redact (default), and block; streaming responses are detected/flagged only. Configured
under guardrails.outputGuardrail. See
Guardrails.
P¶
Per-tenant RAG isolation: A retrieval scope filter that limits a caller to documents stamped with
their tenant on the configured retrieval.tenantIsolation.field (default owner, stamped per point at
ingest and matched against the caller's X-Sandbox-ID). Both backends enforce it (the Qdrant query
filter appends the match; the lexical corpus is stamped with the default sandbox id) and it fails closed:
a missing or unasserted tenant returns no documents, never the whole corpus. Enabled by default;
the single-tenant local lexical lab is the only shipped profile that turns it off. See
src/rag-service/app/retriever.py.
Pod Security Admission (restricted): The Kubernetes built-in that enforces the restricted Pod
Security Standard on platform namespaces via the pod-security.kubernetes.io/enforce|audit|warn:
restricted labels (see
deploy/sandbox/base/namespace.yaml).
Sandbox and tenant-lab namespaces get these labels so workloads run non-root, drop capabilities, and
use a restricted seccomp profile.
Promotion request: A reviewed ModelPromotionRequest (under
platform/model-catalog/promotion-requests/) that authorizes moving a model to a target status. It
records targetStatus, requestedBy, approvers, and a businessJustification. A model cannot reach
approved or a gateway allowlist without a matching promotion request, provenance digest, and evidence.
See Model governance.
Prompt secret detection: An input-path guardrail that rejects requests appearing to carry
credential material (e.g. private_key, github_token, bearer_token), protecting coding-agent
prompts that may accidentally include repo files or env output. PII detectors (email, us_ssn,
credit_card) are built in but opt-in. A match returns HTTP 400 with prompt_secret_detected and
names the pattern without echoing the matched text. See
Guardrails.
Q¶
Quantization (FP8/AWQ): Serving model weights (and optionally the KV cache) at reduced precision
to fit a model onto fewer GPUs, exposed via server.quantization, server.kvCacheDtype, and
server.gpuMemoryUtilization in the vLLM chart. FP8 (Hopper/Ada) gives ~2x memory saving at small
quality cost; AWQ is 4-bit weights for Ampere (point model.name at a pre-quantized …-AWQ
checkpoint). Ready profiles ship at deploy/clusters/customer/values/vllm-nvidia-fp8.yaml and
vllm-nvidia-awq.yaml; re-validate quality with make eval after changing it. See
GPU capacity.
R¶
Release gate: A machine-checked readiness gate declared in platform/slo/release-gates.yaml that
verifies eval, load, restore, toolchain, SLO, quota, provenance, supply-chain, egress/retention, and
evidence-pack evidence before a customer handoff. Run the default gate with make release-gate. A
failed gate means the handoff evidence is incomplete or below threshold. See
Release gates.
Response cache: An exact-match, per-sandbox cache of non-streaming chat completions, keyed by
(sandbox_id, canonical-payload) so a repeated identical request skips the runtime and one tenant's
cached answer is never served to another. TTL + LRU bounded; stream is excluded from the key and
streaming responses are never cached. Off by default (responseCache in
deploy/charts/inference-gateway/values.yaml);
use the Redis backend for a cache shared across replicas. See
src/inference-gateway/app/cache.py.
riskTier: A model catalog field recording the potential harm if a model misbehaves, validated
against {low, medium, high} in
scripts/model-catalog.py.
Distinct from dataClassification: riskTier is about model behavior/impact, dataClassification is about
the sensitivity of the data the model touches. A model carries both, and they must match its provenance
record.
RWX weight cache: A persistent volume that stores downloaded vLLM model weights so they survive
pod restarts and scale-up instead of re-downloading a 100+ GB model into an emptyDir on every cold
start. Configured under cache.persistence in the vLLM chart; sharing weights across replicas requires
a ReadWriteMany (RWX) storage class, whereas ReadWriteOnce requires keeping replicaCount at 1 or
pre-baking weights into the image. Disabled by default. See
GPU capacity.
S¶
Sandbox-id: A lowercase, DNS-label-style identifier (e.g. local-lab, team-a-lab) that scopes a
request for tracing, budgets, and attribution. Callers send it as the X-Sandbox-ID header; the
gateway echoes it and forwards it to the runtime, and it keys sandbox budgets and the response cache.
For a tenant lab it equals the namespace sandbox id. See
Traceability sandbox.
Strict release gate: The stricter form of the release gate, run with make release-gate-strict
for customer demos, release reviews, restore-drill reviews, and production-readiness handoff. It fails
when a required gate falls back to checked-in sample-* evidence or when selected evidence is older
than RELEASE_GATE_MAX_EVIDENCE_AGE_HOURS (default 24h), so the report is based on the current build
rather than only validating sample report shapes. It does not establish production readiness by
itself. See
Release gates
and Evidence and validation.
T¶
Tamper-evident audit chain: A per-process SHA-256 hash chain linking the gateway's redacted audit
events (h_i = SHA-256(h_{i-1} || canonical(record_i))) so any edit, insertion, deletion, or reordering
is detectable by recomputation. The live construction matches the offline auditor/verifier reference in
paper/evidence-model/audit_chain.py byte for byte. The chain is per replica; cross-replica and
long-horizon integrity depend on log shipping and an external head commitment. See
ADR 0006.
Tensor-parallel: Splitting a single model's tensors across the GPUs of one node via
--tensor-parallel-size (set through the vLLM chart's extraArgs/GPU profiles). Keep
--tensor-parallel-size equal to accelerator.count. For models too large for one node, combine it
with pipeline parallelism across nodes under a LeaderWorkerSet topology. See
GPU capacity.
Tenant onboarding: Generating a repeatable, reviewed package for a customer team: both the tenant
sandbox controls (Namespace, quota, LimitRange, NetworkPolicy, trace contract, RBAC) and the
coding-agent workspace values. Driven by a spec under tenants/onboarding/ via make tenant-onboard
(or make tenant-onboard-regulated for the offline profile), with output written under
.out/tenants/<sandbox-id>/. See
Tenant labs.
Traceable sandbox: The ai-sandbox namespace deployed by the traceable-sandbox Application:
quota, default resource limits, restricted Pod Security Admission labels, and default-deny network
policy, proving that a request can be traced end to end without leaking prompt text. Exercised by
make trace-smoke (renamed from the retired sandbox-smoke target in
ADR 0010). Distinct from the agent sandbox workspace
runtime (see above). See
Traceability sandbox.