Prerequisites and scoped previews
Leakage, splits and evaluation contamination
Prevent future information, entity overlap, preprocessing leakage and test-set reuse.
Candidate retrieval and ANN indexes
Separate eligible-corpus recall from latency/memory and filtering effects.
Retrieval and ranking evaluation
Define recall/precision at K, MRR, DCG/NDCG, relevance labels and query/user weighting.
Registry, validation, rollout and rollback
Promote compatible artifact bundles through checks, shadow/canary and rollback gates.
Scoped prerequisite preview. The complete chapter is planned; this preview does not count as authored coverage.
Governance, privacy, lineage and deletion
Trace permitted data use through snapshots, features, indexes and artifacts; define deletion obligations.
Compute, memory and unit economics
Estimate feature, retrieval, training and inference costs with clear units and assumptions.
Scoped prerequisite preview. The complete chapter is planned; this preview does not count as authored coverage.
RAG and generative-system evaluation
Separate retrieval, grounding, generation, tool/action and safety evaluation.
Scoped prerequisite preview. The complete chapter is planned; this preview does not count as authored coverage.
What, when, why and how
Two experiences, one governed knowledge core
- what
- Customer self-service and agent assist share retrieval infrastructure but have different identity, visibility, context and action policies.
- when
- Use separate orchestration contracts when a customer asks about their own account versus an employee assisting them.
- why
- Agent-only runbooks and internal incident notes must never leak into customer answers.
- how
- Derive actor, channel, customer/case scope and allowed tools server-side; reuse only the policy-governed retrieval core.
Knowledge is not customer state
- what
- Knowledge articles explain general product behavior; billing, order, subscription and ticket systems hold live customer facts.
- when
- Use this distinction for account-specific or time-sensitive questions.
- why
- Indexing live customer records creates stale state and cross-customer exposure risk.
- how
- Retrieve approved reusable knowledge; fetch narrowly scoped customer facts from authoritative APIs after object-level authorization.
Conversation context needs provenance
- what
- History and summaries are derived context, not identity proof or authoritative account state.
- when
- Use when conversations resume across sessions, channels or human handoff.
- why
- Summaries can omit negation, mix speakers or retain stale facts.
- how
- Persist attributed message events and source IDs; reauthenticate and re-fetch live state for sensitive decisions.
Grounding and abstention
- what
- A correct answer must be supported by current approved evidence and fit the channel disclosure policy.
- when
- Use when articles conflict, evidence is stale or policy is ambiguous.
- why
- Fluency and retrieval score do not establish correctness or permission.
- how
- Validate article version, citation and claim support; clarify, partially answer or abstain when evidence is insufficient.
Escalation is a durable workflow
- what
- Handoff transfers a case with a compact evidence-backed summary and confirmed ticket ownership.
- when
- Use when a person is requested, risk is high, repeated failure occurs or approval is needed.
- why
- A dropped handoff loses trust; duplicate ticket creation creates operational noise.
- how
- Persist idempotent handoff intent, attach approved context and confirm ticket/queue before claiming success.
Measure resolution, not just containment
- what
- Quality includes verified resolution, repeat contact, customer effort, escalation appropriateness and harm.
- when
- Use when comparing prompts, retrieval, model versions or automation.
- why
- Optimizing deflection alone can trap customers in loops.
- how
- Join turns to mature ticket outcomes; evaluate retrieval, grounding, policy, escalation and outcomes separately.
Frame the support experience
We design SaaS support for customer web chat and a separate agent copilot in the ticket console. The system answers common product/policy questions, reads narrowly scoped account state through approved APIs and escalates cases requiring judgment. It does not replace identity, CRM, billing, order or ticket systems.
Synthetic interview assumptions: 100 tenants, 50,000 approved knowledge articles, 80 peak turns/s, 2.5 s p95 answer target, knowledge freshness within 15 minutes and urgent policy withdrawal denied within one minute.
Separate channels and trust boundaries
Customer and agent requests use different identities, knowledge visibility and tool permissions. Customer answers use published customer-safe articles; agents may access role-approved internal runbooks only after workforce and case-scope checks.
The browser, customer messages, uploads and retrieved text are untrusted. The LLM is not an authorization authority. A conversation ID, email address or model-generated summary never proves customer identity.
Start with deterministic search
Begin with help-center keyword search, stable article versions and direct links. Add deterministic intent routing for known flows such as password reset, billing status and outage notices. Introduce generation only when synthesis across approved sources adds measurable value.
Compare against the baseline on held-out support questions. Measure correctness, evidence support, policy violations, escalation and resolution—not style or containment alone.
Model conversations, evidence and handoff
Conversation events are append-only and attributed by speaker, channel and time. Summaries are versioned, bounded and rebuildable with source message IDs. Knowledge articles have immutable versions, locale, product version, effective dates and audience visibility.
Customer/account/order state remains in authoritative systems. Evidence references exact article versions and source spans. Customer binding, tool execution and handoff are durable auditable records.
Ingest approved support knowledge
Connect help-center pages, release notes and runbooks through stable source IDs and version cursors. Parse, normalize and chunk with title, locale, product version, effective dates, audience visibility and source offsets.
Use durable idempotent jobs and staged indexes. Publish only after metadata and retrieval checks. An urgent withdrawal writes a deny barrier immediately; physical index/cache cleanup is asynchronous. Reconcile missed webhooks.
VISUAL WALKTHROUGH
See the system in motion
Follow approved support knowledge from versioned connector event through validation and publication.
Capture stable source identity and revision.
Create a durable idempotent job.
Parse and classify audience visibility.
Stage chunks and indexes.
Validate locale, product version and policy metadata.
Publish the active version conditionally.
Apply urgent deny barriers before cleanup.
Loading interactive walkthrough…
Authenticate actor and assemble bounded context
Resolve tenant, actor, channel and authentication state server-side on every turn. Customer binding is an explicit verified relationship; account-specific requests require appropriate authentication strength and object-level authorization.
Load recent events plus a bounded attributed summary. Preserve speaker labels, timestamps, unresolved questions and source IDs. Re-fetch live state and reauthorize every sensitive read; redact secrets and minimize retention.
VISUAL WALKTHROUGH
See the system in motion
Trace verified channel identity, explicit customer binding and bounded attributed conversation history.
Validate channel session.
Derive actor and tenant from trusted identity.
Load bounded messages and summary provenance.
Bind customer only after verification.
Check object-level scope.
Reauthorize sensitive tool reads.
Loading interactive walkthrough…
Retrieve only channel-authorized knowledge
Build the query from the current turn and bounded history. Apply tenant, audience, locale, product/version, effective-date and deny-state filters. Run lexical and dense retrieval in parallel, fuse/deduplicate, then rerank a bounded authorized set.
Customer and agent visibility must be explicit. Recheck current article policy before prompt assembly. Cache keys include tenant, actor/role, channel, locale, product version, policy epoch and index generation. Fail closed if policy is unavailable.
VISUAL WALKTHROUGH
See the system in motion
Retrieve customer-safe or agent-authorized evidence without allowing semantic relevance to grant access.
Resolve channel and actor policy.
Filter tenant, audience, locale and product version.
Run lexical and dense retrieval.
Fuse and deduplicate candidates.
Check current policy before hydration.
Rerank only authorized evidence.
Loading interactive walkthrough…
Fetch customer facts from systems of record
A question such as 'Where is my refund?' needs a live billing/order lookup, not semantic retrieval over old tickets. Select a typed read tool, validate parameters, authorize the exact resource and call the authoritative service.
Fetch minimum necessary fields, attach source timestamp and provenance, redact unrelated data and apply per-customer limits. If the source is unavailable or too stale, explain the limitation and escalate rather than guessing.
VISUAL WALKTHROUGH
See the system in motion
Fetch customer-specific facts from source-of-truth APIs using narrow typed tools and object authorization.
Classify live-state intent.
Validate verified customer binding.
Authorize exact account/order/case.
Call a least-privilege read tool.
Attach source timestamp and provenance.
Redact unrelated sensitive fields.
Loading interactive walkthrough…
Generate and validate grounded answers
The prompt contains channel policy, bounded context, authorized knowledge evidence and separately marked structured tool results. Retrieved text and customer messages are untrusted data; their embedded instructions cannot override policy.
Return structured claims with evidence references. Validate schema, citation IDs, article versions, visibility and support for material claims. For live facts attach API freshness. If evidence conflicts or fails validation, clarify or abstain.
VISUAL WALKTHROUGH
See the system in motion
Generate a structured answer and validate evidence, article version, channel visibility and live-data freshness.
Assemble authorized evidence and structured facts.
Treat source text as untrusted.
Generate a bounded structured draft.
Validate citation IDs and versions.
Check claim support and policy.
Answer, clarify or abstain.
Loading interactive walkthrough…
Keep consequential actions confirmed and auditable
Separate read tools from write tools. A refund, cancellation or address change follows a deterministic workflow: explain details, obtain explicit confirmation, step-up if needed, reauthorize, execute with idempotency and verify resulting state.
The model may propose a typed call but cannot grant permission, invent resource IDs or skip approval. Enforce allowlists, schema validation, replay protection, rate limits and audit events outside the model.
VISUAL WALKTHROUGH
See the system in motion
Keep consequential actions outside model authority; require confirmation, reauthorization and idempotency.
Propose a typed action.
Validate parameters and eligibility.
Explain and obtain explicit confirmation.
Reauthorize the resource.
Execute with an idempotency key.
Verify state from the system of record.
Loading interactive walkthrough…
Create a reliable human handoff
Escalate when requested, evidence is insufficient, risk/policy requires review, repeated attempts fail or a workflow needs approval. Persist a deduplicated handoff with summary, identity state, attempted steps, evidence references and reason code.
The ticket system owns ticket ID, queue, priority, SLA and assignment. Keep the conversation pending until creation is confirmed. If ticketing is down, persist intent and retry safely; do not claim an agent accepted it.
VISUAL WALKTHROUGH
See the system in motion
Persist a deduplicated handoff and report success only after ticket and queue confirmation.
Select escalation reason and queue.
Build a privacy-minimized evidence summary.
Persist handoff intent with idempotency key.
Create or recover the ticket.
Confirm ticket ID and queue.
Keep pending state if ticketing is unavailable.
Loading interactive walkthrough…
Evaluate retrieval, policy and resolution
Build held-out questions with source/version judgments, expected facts, acceptable abstentions and escalation labels. Split by issue family, customer/account and time to avoid near-duplicate leakage. Review claim support, policy, tone and next-step correctness.
Join turns to ticket outcomes with a maturity window. Track verified resolution, reopen/repeat contact and customer effort alongside containment. Segment by language, product version, channel and customer tier; thumbs-up is noisy feedback.
Bound capacity and dependency work
Synthetic peak: 80 turns/s × 1,200 input tokens = 96,000 input tokens/s; at 180 output tokens, 14,400 output tokens/s before retries. If 20% require a live read, that is 16 tool reads/s. These are arithmetic assumptions, not measured capacity.
Set per-tenant admission control, retrieval caps, token budgets, tool deadlines and an overall request deadline. Batch embeddings offline, isolate ingestion from interactive traffic, protect ticketing from retry storms and never blindly retry non-idempotent writes.
Trace the end-to-end request
The gateway receives a customer or agent turn. Identity middleware derives actor, tenant, channel and scope. The orchestrator loads bounded context and routes to knowledge answer, authorized live-data lookup or handoff/action workflow.
Knowledge flow performs policy-filtered hybrid retrieval, reranking, generation and citation validation. Live-data flow calls an allowlisted source-of-truth API. Handoff persists an idempotent ticket request. Every path emits privacy-minimized audit and telemetry.
VISUAL WALKTHROUGH
See the system in motion
Trace the complete customer and agent support architecture across identity, RAG, live tools, validation and handoff.
Authenticate actor and channel.
Load attributed conversation context.
Route to knowledge, live data or handoff.
Retrieve only authorized evidence.
Call live tools after object authorization.
Validate answer claims and citations.
Confirm actions or durable handoff.
Record privacy-minimized outcomes.
Loading interactive walkthrough…
Degrade safely
Model timeout returns approved search results, retry guidance or handoff—not invented facts. Vector failure can fall back to lexical search. Policy outage fails closed for protected content. Stale billing/order API is labeled unavailable. Ticket outage leaves a visible pending handoff.
Recheck article deny state before final response. Session expiry stops sensitive reads/writes. Duplicate tool calls return the existing idempotent operation. Tag fallbacks so analytics do not count them as grounded resolutions.
Roll out with independent rollback controls
Replay held-out cases, then shadow retrieval/generation without showing output. Launch agent copilot first with citations and human control; use a small cohort. Customer self-service follows after channel-specific safety and escalation tests pass.
Version prompt, retrieval config, embedding model, index generation, policy bundle and tool schemas as a compatible release. Rollback cannot override current authorization, deny lists or account permissions.
Senior and staff-level trade-offs
A shared index is simpler, while separate visibility partitions reduce disclosure risk. More conversation history can improve continuity but raises cost, privacy exposure and stale-context risk. Agent copilot is safer initially than autonomous customer actions.
Choose the least autonomous system that meets the verified customer outcome. Physical isolation, separate indexes and action automation should be justified by risk, scale and operational evidence.
Data model, schema and relationships
Logical model separates conversation events, rebuildable summaries, immutable approved knowledge, exact evidence, verified customer binding, tool executions and durable handoff. Customer/order/account state remains in external systems of record.
Implementation boundary: Educational architecture contract only; no migrations, live connectors, authorization or ticketing integration are installed.
These are proposed logical schemas with field contracts, not executable database migrations. PK identifies a row, FK references an existing parent, UK is a uniqueness rule. A composite PK identifies the combination. Nullable references are marked with ?. A foreign key protects identity; cross-row limits, timing and state transitions also require the transaction rules below.
Connectors name the relationship; multiplicities state parents per child and children per parent. 0..1 means optional, 1 required, 0..* any number. Self-links show two roles of the same entity.
Connectors name the relationship; multiplicities state parents per child and children per parent. 0..1 means optional, 1 required, 0..* any number. Self-links show two roles of the same entity.
Connectors name the relationship; multiplicities state parents per child and children per parent. 0..1 means optional, 1 required, 0..* any number. Self-links show two roles of the same entity.
Connectors name the relationship; multiplicities state parents per child and children per parent. 0..1 means optional, 1 required, 0..* any number. Self-links show two roles of the same entity.
Entity definitions and field contracts
Conversation
Channel-scoped interaction lifecycle.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
tenant_id | opaque_id · VALUE | tenant id field. |
channel | enum · VALUE | channel field. |
status | enum · VALUE | status field. |
Code mapping: Logical architecture schema; not an installed migration.
Message event
Append-only actor-attributed message.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
conversation_id | opaque_id · FK | conversation id field. |
speaker | enum · VALUE | speaker field. |
created_at | timestamp · VALUE | created at field. |
Code mapping: Logical architecture schema; not an installed migration.
Conversation summary
Derived bounded context with provenance.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
conversation_id | opaque_id · FK | conversation id field. |
source_message_ids | json · VALUE | source message ids field. |
expires_at | timestamp · VALUE | expires at field. |
Code mapping: Logical architecture schema; not an installed migration.
Knowledge article
Stable connector-scoped source identity.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
tenant_id | opaque_id · VALUE | tenant id field. |
source_id | string · VALUE | source id field. |
active_version_id | opaque_id? · VALUE | active version id field. |
Code mapping: Logical architecture schema; not an installed migration.
Article version
Immutable content and audience contract.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
article_id | opaque_id · FK | article id field. |
revision | string · VALUE | revision field. |
visibility | enum · VALUE | visibility field. |
locale | string · VALUE | locale field. |
Code mapping: Logical architecture schema; not an installed migration.
Retrieval evidence
Exact source/version/span used for a turn.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
conversation_id | opaque_id · FK | conversation id field. |
article_version_id | opaque_id · FK | article version id field. |
source_span | json · VALUE | source span field. |
Code mapping: Logical architecture schema; not an installed migration.
Customer binding
Verified conversation-to-customer relationship.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
conversation_id | opaque_id · FK | conversation id field. |
verified_subject | string · VALUE | verified subject field. |
customer_id | opaque_id · VALUE | customer id field. |
expires_at | timestamp · VALUE | expires at field. |
Code mapping: Logical architecture schema; not an installed migration.
Tool execution
Auditable typed operation.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
conversation_id | opaque_id · FK | conversation id field. |
idempotency_key | string · VALUE | idempotency key field. |
approval_id | opaque_id? · VALUE | approval id field. |
status | enum · VALUE | status field. |
Code mapping: Logical architecture schema; not an installed migration.
Handoff
Durable transfer to human queue.
| Field | Type / key | Meaning and reference |
|---|---|---|
id | opaque_id · PK | id field. |
conversation_id | opaque_id · FK | conversation id field. |
ticket_id | string? · VALUE | ticket id field. |
idempotency_key | string · VALUE | idempotency key field. |
status | enum · VALUE | status field. |
Code mapping: Logical architecture schema; not an installed migration.
Relationship cardinalities and their reasons
| Parent → child | Multiplicity | FK / why |
|---|---|---|
| conversation → message_event Message events | 1 parent per child; 0..* children per parent | message_event.conversation_idEach record belongs to one conversation and remains explicitly scoped. |
| conversation → conversation_summary Summary revisions | 1 parent per child; 0..* children per parent | conversation_summary.conversation_idEach record belongs to one conversation and remains explicitly scoped. |
| conversation → retrieval_evidence Evidence records | 1 parent per child; 0..* children per parent | retrieval_evidence.conversation_idEach record belongs to one conversation and remains explicitly scoped. |
| conversation → customer_binding Verified bindings | 1 parent per child; 0..* children per parent | customer_binding.conversation_idEach record belongs to one conversation and remains explicitly scoped. |
| conversation → tool_execution Tool executions | 1 parent per child; 0..* children per parent | tool_execution.conversation_idEach record belongs to one conversation and remains explicitly scoped. |
| conversation → handoff Handoff requests | 1 parent per child; 0..* children per parent | handoff.conversation_idEach record belongs to one conversation and remains explicitly scoped. |
| knowledge_article → article_version Article versions | 1 parent per child; 1..* children per parent | article_version.article_idA source article has immutable revisions. |
| article_version → retrieval_evidence Cited version | 1 parent per child; 0..* children per parent | retrieval_evidence.article_version_idEvidence points to the exact article version. |
Transaction boundaries
- Conversation identity binding is appended only after verified authentication; tenant and customer scope are checked transactionally.
- Article publication advances active_version_id only after staged chunks, audience metadata and retrieval checks pass.
- Tool execution requires a validated allowlisted operation, explicit approval where consequential, and a unique idempotency key.
- Handoff intent is persisted before ticket creation; ticket ID and queue status are recorded before the conversation reports a confirmed handoff.
Indexes and uniqueness
- UNIQUE(tenant_id, source_id) on knowledge_article.
- UNIQUE(article_id, revision) on article_version.
- INDEX(conversation_id, created_at) on message_event.
- UNIQUE(idempotency_key) on tool_execution and handoff.
- Search indexes partition/filter by tenant, audience visibility, locale, product version and active article generation.
Schema traps and failure cases
- Expired or missing customer binding blocks sensitive reads and writes.
- Policy service outage fails closed for protected articles and customer objects.
- Duplicate action or ticket requests return the existing idempotent execution.
- Ticket API outage leaves a durable pending handoff and does not claim agent acceptance.
- Article withdrawal denies immediately while asynchronous cleanup reconciles every index and cache.
Trace the records through one operation
- Customer asks for refund status; authenticate, bind the customer, authorize the exact order and fetch live billing state.
- Assistant retrieves customer-visible refund policy, validates article version and cites both policy and live status.
- Customer requests a refund; show the proposed amount and reason, obtain confirmation, reauthorize, execute idempotently and verify the ledger.
- Evidence is insufficient; persist a handoff with the approved summary and citations, create/recover a ticket and confirm queue ownership.
Diagrams and relationships
Authorization and source-of-truth systems remain authoritative; the model coordinates a bounded workflow.
Scroll horizontally on smaller screens, or enlarge the diagram.
Urgent withdrawal denies immediately; physical index cleanup is asynchronous.
Scroll horizontally on smaller screens, or enlarge the diagram.
Conversation text never authenticates a customer; sensitive tools reauthorize each resource.
Scroll horizontally on smaller screens, or enlarge the diagram.
Evidence validation and ticket confirmation determine what the customer is told.
Scroll horizontally on smaller screens, or enlarge the diagram.
Engineer, Senior and Staff follow-ups
These are teaching lenses. Each answer changes the scope of reasoning; they are not official employer leveling criteria.
Engineer: Why not put all customer tickets into the vector database?
Tickets contain sensitive customer-specific facts, stale state and agent-only notes. Keep live customer state in authoritative systems and fetch only authorized fields. Index only approved reusable knowledge with explicit audience visibility.
Engineer: A user asks for a refund. What does the LLM do?
It may propose a typed workflow. The service validates identity, eligibility and policy; obtains explicit confirmation; reauthorizes; executes with an idempotency key; and verifies the ledger state. The model never grants permission.
Senior: How do you prevent a summary from mixing customers or losing a critical negation?
Bind summaries to conversation and customer-binding versions, retain source message IDs and speaker/time metadata, reauthenticate across sessions, and confirm consequential intent against original messages and live state.
Senior: What if a help article is withdrawn during generation?
Recheck deny state before final response. Discard disallowed evidence and regenerate from remaining authorized sources or abstain. Purge all indexes and caches asynchronously with acknowledgements.
Senior: Why launch agent copilot before customer self-service?
An agent remains accountable and can correct suggestions while reviewed evidence accumulates. It reduces initial customer-facing risk, but still requires least-privilege authorization and provenance in the UI.
Staff: How do you measure actual resolution?
Use mature ticket outcomes: verified resolution, reopen/repeat contact, customer effort, escalation appropriateness and harm. Keep retrieval, claim support, policy violations, latency and cost as separate diagnostics.
Staff: How do you stop rollback restoring withdrawn content or unsafe actions?
Version model, prompt, index and tool schema as a compatible bundle, but keep authorization and deny state authoritative/current. Rollback cannot override policy epochs or deletion barriers.
Practice and self-review
Your practice draft
Design customer-support RAG for web chat and agent copilot under the synthetic 100-tenant, 50,000-article, 80-turn/s workload. Draw ingestion, identity, knowledge/live-data boundaries, retrieval, grounding, action confirmation, handoff, evaluation and rollout. Then change one constraint: session expires mid-tool call, article is withdrawn during generation or ticketing is unavailable.
Saved locally on this device. The worked answer does not score your draft. Existing lesson notes and completion are separate.
Progressive hints
Hint 1
Separate customer and agent channels, identity, audience visibility and allowed tools.
Hint 2
Keep reusable approved knowledge in RAG; fetch customer-specific state from source-of-truth APIs after object authorization.
Hint 3
Trace a normal answer, live lookup, sensitive action and escalation; include idempotency and safe failure.
Reveal worked answer guidance
Use governed versioned knowledge ingestion with channel-specific retrieval policy. Derive actor, tenant and scope from verified identity; load bounded attributed context; route to knowledge, live-data or handoff. Retrieve only authorized evidence, fetch customer facts from authoritative APIs after object-level authorization, and validate structured claims/citations before responding. Sensitive actions require confirmation, reauthorization, idempotency and post-write verification. Persist handoff intent and report success only after ticket/queue confirmation. Evaluate retrieval, grounding, policy and mature resolution separately; launch agent copilot before customer automation.
Evidence to check in your answer
- Channel-specific identity and disclosure policy.
- Immutable article versions, attributed conversation events and evidence provenance.
- Object-level authorization and least-privilege tools.
- Hybrid retrieval, citation validation, grounding and abstention.
- Confirmed idempotent actions and durable handoff.
- Held-out evaluation and mature customer outcomes.
- Bounded capacity, safe fallback, rollout and honest synthetic assumptions.
For each item, record Missing, Prompted, Independent or Adaptive evidence. Adaptive means you can revise the reasoning under a changed constraint. Reading completion and readiness are separate.
Sources and verification scope
Original synthetic examples and reference calculations. Sources support narrowly scoped background; they do not establish production performance or independent chapter review.
SRE workbook: data processing pipelines ↗ · Pipeline SLOs and operational ownership; do not copy source workload numbers as course production results.
NIST AI Risk Management Framework ↗ · Risk vocabulary and governance; not a certification or blanket legal-compliance guarantee.
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks ↗ · Historical research foundation for retrieval/generation; not a complete modern permission/security architecture.
Hidden Technical Debt in Machine Learning Systems ↗ · System-boundary and lifecycle risks; source lessons motivate original examples.
Primary background links checked during authoring on 2026-10-19. Arithmetic, temporal eligibility, maturity and split reference functions are checked by the bounded local fixture suite. SQL is explanatory, not a migrated production query. No GPU, external dataset, model API or training job is required.