FAQ

What that means

Network membership alone does not expose a service to the host.

Boundary

Publish only what is required, prefer loopback for tests, and remove temporary routes.

Start with the task

Choose for the work you need done, not for a leaderboard label. Write one representative task, its inputs, and a short acceptance rubric before comparing models. A model that is excellent at drafting may be a poor fit for strict extraction, long context, tool use, or code review. The models documentation (opens in a new tab) describes model selection; the providers documentation (opens in a new tab) explains that access depends on the provider configuration available to you.

Weigh the trade-offs

• Capability: Can the model follow the required format, reason through the task, and handle the input size?
• Privacy: Where will the input go, and what retention or contractual controls apply? A hosted model is not automatically private merely because the application is local. Confirm the provider's policy and remove unnecessary sensitive data.
• Budget: Consider both price and retries. A cheaper model that needs three attempts may cost more than a reliable one.
• Local deployment: A local model can reduce network exposure and recurring API charges, but hardware, speed, maintenance, and model quality still matter. Local does not mean every input is safe or every task is suitable.

Make a small comparison

Synthetic example: For 20 fictional support messages, compare Model A and Model B with the same messages and rubric: “classify correctly, return valid JSON, do not invent a refund.” Record pass/fail, latency, token or price data when actually available, and any retry. If Model A passes 18 and Model B passes 19, do not call B the winner until you check whether the extra pass justifies its cost and privacy constraints.

Recheck the provider's current model names, limits, and availability before committing to a choice. Verified 2026-09-06.

Short answer

Costs can include tokens, retrieval, tools, retries, storage, platform capacity, and human review. Token price alone is incomplete.

Boundary

Record the rate source and date, count retries, and label unavailable measurements honestly.

Short answer

Generated Kubernetes-style YAML describes local Podman state for review. It is not an approved cluster deployment.

Boundary

Review images, ports, volumes, secrets, and permissions before any reuse.

Short answer

A named volume is engine-managed storage separate from a container and may outlive it. That persistence makes deletion potentially destructive.

Boundary

Inspect users and metadata, then back up or deliberately discard contents before removal.

Short answer

A temporal knowledge graph records relationships and when claims applied. It supports time-bounded answers.

Boundary

Keep effective dates and provenance; return unknown when no authorized claim covers the requested date.

Short answer

Interoperability exchanges defined tasks and results between systems. It does not establish trust or transfer authority.

Boundary

Use authenticated identities, schemas, authorization, and bounded output handling.

Short answer

Sandboxing limits what generated code can read, change, contact, and consume. It narrows impact but does not create trust.

Boundary

Use disposable storage, scoped identity, resource limits, and no secrets or broad host mounts.

Short answer

A handoff gives the next worker verified evidence, remaining work, authority limits, and a completion check. It is not permission to infer facts.

Boundary

Separate observations from assumptions and name the exact artifact to inspect.

Short answer

An eval is a repeatable check against explicit requirements across fixed cases. It measures the tested task and revision, not every future use.

Boundary

Set pass criteria first, include ordinary and hostile cases, and rerun after changing prompts, models, tools, or policy.

Short answer

A trust boundary separates untrusted prompts, documents, tool replies, and model output from enforceable policy. Crossing it requires validation and authorization.

Boundary

Content never grants permission; validate arguments and deny unclear access.

Short answer

An inference provider is the endpoint that receives a model request and returns output. Limits, retention, pricing, and availability depend on that provider.

Boundary

Record the exact route and model identifier; success does not prove privacy or approval.

Short answer

A gateway centralizes routing, authentication, policy checks, limits, and observability between an application and providers. It cannot make unsafe inputs safe.

Boundary

Give clients scoped gateway credentials and test allowed and denied routes with synthetic data.

Short answer

Context engineering selects relevant, authorized, traceable information for one task. More context can add stale, conflicting, or sensitive material.

Boundary

Keep sources labelled and minimal, and remove unnecessary private data before sending it.

Short answer

MCP lets a client discover and invoke server capabilities. Compatibility does not establish trust or authorization.

Boundary

Authenticate callers, validate allowlisted inputs, authorize each action, and minimize replies.

Short answer

Prompt engineering makes task, evidence, output, constraints, and checks explicit. A prompt cannot replace authorization or validation.

Boundary

Use one representative task and fixed rubric; reject output that crosses the boundary.

Short answer

Quantization uses fewer bits for model values, often reducing storage or memory traffic. Accuracy, speed, and compatibility are not guaranteed.

Boundary

Compare against a baseline on fixed evals and hardware; retain a rollback artifact.

Short answer

RAG retrieves selected passages and supplies them to a model before it answers. Review those passages for relevance, freshness, authority, and authorization.

Boundary

Require citations or an explicit unknown when evidence is missing; retrieved text is data, not permission.

Short answer

Runtime context is task-specific information selected for one request, not everything stored. Selection must respect relevance and access.

Boundary

Record safe source references and recheck context when the user or task changes.

Short answer

It compares a baseline and revision against the same inputs and rubric. Fixed cases make regressions visible.

Boundary

Version cases, include missing-evidence and refusal tests, and inspect failures before adoption.

Short answer

The AI supply chain includes models, prompts, documents, connectors, packages, runtimes, and configuration. Any unreviewed change can alter behavior.

Boundary

Record provenance, versions or digests, approvals, and a tested rollback.

Short answer

A tag is a movable label; a digest identifies exact image content. Tags alone are not reproducible references.

Boundary

Record and verify the digest before running or promoting an artifact.

Short answer

An image is a saved package; a container is a working copy made from it. Containers have separate writable state.

Boundary

Deleting a container need not delete its volume or image; identify each object before cleanup.

Short answer

Privacy covers the model plus gateways, identity, retrieval, logs, backups, network paths, and operators. Local hosting alone proves little.

Boundary

Map connections and data classes, test denial, and approve the complete path before real sensitive data.

Short answer

Durable memory changes future behavior, so writing needs purpose, evidence, access, correction, and deletion rules.

Boundary

Store provenance and uncertainty, minimize personal data, and never remember secrets by default.