watsonx AI Assistant Engineer flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

Assistant object hierarchy

An instance contains assistants; each assistant has draft and live environments plus channels, knowledge bases, and extensions. Actions and dialog are edited in draft and published to live as a snapshot.

Intent versus entity

An intent captures a user goal, defined by example utterances. An entity captures a typed parameter such as date or product, with synonyms and patterns used to extract values from user input.

Session state

The runtime stores conversation context under a session ID. Stateless message calls must resend context each turn because no server-side session is retained between requests.

Draft versus live

Draft is the editable working copy visible in preview. Live serves published content to real channels and changes only when you explicitly publish a snapshot.

Action naming rule

Name an action as a short user request, verb plus object, because the name and description feed the model that decides which action matches the user's utterance.

Two-stage fallback design

First attempt a knowledge base search for unmatched input; if confidence stays below threshold, escalate to a human or offer contact options. Log the failed utterance for review.

Slot filling discipline

Collect one required slot per turn, mark optional slots explicitly, and confirm captured values before irreversible actions. This reduces correction turns and stale data.

Action disambiguation

When several actions plausibly match, the assistant asks a clarifying question listing candidates. Keep candidate names short and clearly distinct to make the choice easy.

Action step types

Steps are conversational (collect or confirm), advanced (conditions, variable assignment, expressions), extension calls, or end-with-response. The first step is model-generated unless set to fixed.

Variable scope

Action-scoped variables exist only during that action; session-scoped variables persist across actions in the conversation. Reference them in text and JSON with the dollar-brace syntax.

Expression language

Action expressions use Spring Expression Language. Helpers expose input text, system entities such as date and number, and string or date utility methods for transformations.

Digressions

A digression lets a dialog node interrupt the current flow, handle a side request, then return to the original node. Use sparingly to avoid confusing multi-turn state.

Model size trade-off

Smaller, faster models suit intent routing and slot extraction. Larger models suit generative answers and multi-step reasoning, at higher latency and cost per turn.

Structured output prompting

Instruct the model to return JSON with named fields when a later step parses the result. Include an explicit instruction to say unknown rather than invent values.

Temperature setting

Low temperature, roughly 0 to 0.3, gives deterministic extraction and routing. Higher values, roughly 0.5 to 0.7, give more varied conversational phrasing for open answers.

Prompt injection defense

Treat user text as data, not instructions. Separate system instructions from user content, and validate every argument before passing it to an extension or tool call.

Knowledge base chunking

Documents are split into chunks before indexing. Smaller chunks improve retrieval precision; overlap prevents answers from being cut at chunk boundaries.

Hybrid search

Hybrid search combines lexical keyword matching with semantic vector similarity. It usually outperforms either alone when users mix product codes with natural phrasing.

Relevance threshold

A minimum retrieval score below which the assistant declines to answer from the knowledge base. Set it so weak matches route to fallback instead of producing wrong answers.

Source synchronization

Each knowledge source has a sync schedule and last-sync status. Uploaded and crawled content is not searchable until indexing completes, and edits require re-sync.

RAG pipeline stages

Query rewrite, retrieval of top-k chunks, rerank and filter, prompt assembly with context, then grounded generation with citations. Each stage has its own quality and latency cost.

Grounding constraint

The model may assert only facts present in retrieved context; otherwise it must state the information is unavailable. This constraint is what separates grounded answers from hallucination.

Conversational query rewriting

Follow-up questions like what about the second one lack standalone meaning. Rewrite them using dialogue history into a self-contained query before retrieval, or results will be poor.

Groundedness metric

Groundedness, or faithfulness, measures whether each claim in an answer is supported by retrieved sources. Score it with human review or a model-based judge on sampled answers.

Custom extension definition

An OpenAPI 3.0 specification imported into the assistant. Its operations become callable steps with typed parameters, and authentication secrets live in the extension configuration.

Message API call

Send a POST to the assistant message endpoint with assistant ID and session context. The response returns output text, updated context, and any user-defined actions to perform.

Idempotent external calls

Channels may retry on timeout, so extension calls that create or change records should be idempotent or guarded by a request key to prevent duplicate side effects.

Containment rate

The share of conversations resolved without human escalation. It is the primary outcome metric, best read alongside fallback rate and action completion rate.

Fallback utterance backlog

Unmatched user inputs are the main improvement queue. Cluster them by theme, then add actions, entity synonyms, or knowledge content, and re-measure after release.

Regression test set

A fixed list of utterances with expected action or knowledge answers, run before each publication. It catches prompt, model, and content changes that break previously working flows.

Least privilege credentials

Scope API keys and service credentials to the minimum role needed, rotate them regularly, and never place them in client-side web chat code where users can read them.

PII redaction

Mask or remove personal data before sending text to a model or writing logs. Configure log redaction and retention to match your organization's data policy.

Rollback via export

Export action and dialog JSON before publishing, or republish a prior snapshot. After rollback, verify channel credentials and extension endpoints still match the restored configuration.