watsonx AI Assistant Engineer key terms

The top-level container in watsonx Assistant that holds actions, dialog, knowledge sources, integrations, and channel configurations. It is versioned and published as a whole, and each environment points at a specific published version.

A discrete unit of conversational capability that defines the steps the assistant takes to complete a user goal, including clarifying questions, variable collection, tool calls, and response generation. Actions replaced the older dialog-node-centric modeling approach.

An individual instruction inside an action, such as asking a question, setting a variable, calling an extension, or producing a final answer. Steps execute in order and can branch on conditions, so action logic is read as a sequence rather than a graph.

A typed slot that stores information collected or computed during an action, such as a customer ID or order number. Variables can be scoped to a single step, the whole action, or the session, and their types drive validation of user input.

The technique of collecting several required pieces of information across multiple turns before acting, rather than asking one question and stopping. The assistant tracks which slots remain empty and re-prompts only for missing values.

The behavior that occurs when user input matches more than one action or knowledge topic with similar confidence. The assistant presents the candidates as options so the user selects the intended intent instead of the system guessing.

The configured response path taken when no action, dialog branch, or grounded answer meets the confidence threshold. Good fallback design offers search, escalation to a human agent, or a clarifying question rather than a dead end.

A tunable cutoff that determines how certain the classifier or retrieval step must be before the assistant acts on a match. Raising it reduces wrong answers but increases fallbacks; lowering it does the reverse.

The user goal inferred from an utterance, such as checking an order status or resetting a password. Intents are the unit of routing in classic dialog design and remain relevant when mapping utterances to actions.

A structured piece of information extracted from user input, such as a date, product name, or account number. Entities convert free text into typed values that actions can branch on and pass to backend systems.

An example phrase a designer supplies to teach the assistant how users express a given intent or action. Coverage across paraphrases, slang, and misspellings matters more than sheer volume of near-duplicate examples.

A neural network trained on broad text corpora that generates and interprets natural language. In watsonx Assistant it powers action step generation, conversational search answers, and utterance classification helpers, subject to grounding and guardrails.

The instruction and context supplied to a large language model for a single request. Effective prompts state the role, task, constraints, output format, and relevant context, and are iterated against real examples rather than written once.

Including a small number of worked input-output examples in a prompt so the model imitates the desired pattern. It is useful for enforcing response tone, JSON structure, or classification labels without fine-tuning.

Constraining a model response to a machine-readable shape such as JSON with named fields, so downstream steps can parse it reliably. It is typically enforced by explicit schema instructions plus validation and retry logic.

A fluent but unsupported or incorrect model statement, often caused by gaps in retrieved context or ambiguous prompts. Mitigations include grounding, citation requirements, refusal instructions, and post-hoc validation of factual claims.

A curated collection of documents or web content indexed so the assistant can retrieve passages relevant to a user question. Content quality, chunking, and freshness of the source material dominate answer quality.

An architecture that retrieves relevant passages from a knowledge source and supplies them to a language model as context for the answer. It improves factuality and traceability compared with relying on model parameters alone.

Tying a generated answer to specific retrieved source content so the response reflects organizational knowledge rather than model guesswork. Grounded answers typically include citations or links that let users verify the claim.

Splitting source documents into smaller passages before indexing so retrieval can return focused context. Chunk size and overlap trade precision against losing surrounding meaning, and structure-aware splitting usually beats fixed-length cuts.

A numeric vector representation of text that places semantically similar content near each other in vector space. Retrieval compares the user question's vector against indexed passage vectors to rank candidates.

Retrieval by similarity in embedding space rather than keyword overlap, allowing matches on meaning even when wording differs. Hybrid approaches combine it with lexical search to catch exact identifiers and product codes.

A reference in an answer pointing to the source document or passage it was drawn from. Citations build user trust, support auditing, and make it easier to spot knowledge gaps during content review.

The designed behavior when retrieval finds nothing sufficiently relevant or the model cannot ground a response. A good pattern acknowledges the gap, offers related topics, and routes to a human rather than inventing content.

A registered external service, described by an OpenAPI specification, that an assistant can call during an action to fetch or update data. Extensions decouple conversational logic from backend implementation and are configured per environment.

A standard, machine-readable description of an HTTP API's endpoints, parameters, and schemas. watsonx Assistant imports it to know which operations an extension exposes and how to build requests from action variables.

An outbound HTTP call the assistant makes at a defined point, such as after collecting input or before sending a response. Webhooks carry conversation context to custom code that can transform or enrich the payload.

A programmatic interface for interacting with the assistant outside the web chat, used to send messages, manage workspaces, or retrieve logs. Stateless and stateful endpoints differ in whether the server tracks session context.

A bounded conversation thread identified by a session ID, holding context such as collected variables and recent turns. Session lifetime and statefulness affect how much history the assistant can use when answering.

The set of variables and metadata carried through a conversation and shared with actions, extensions, and the model. Deliberate context design prevents leakage of sensitive values and keeps prompts within token limits.

A surface where the assistant is deployed, such as a website chat widget, a contact-center platform, a messaging app, or a voice gateway. Each channel imposes its own formatting, latency, and authentication constraints.

A configured connection between the assistant and an external platform, such as a contact-center provider, messaging service, or search backend. Integrations handle transport and authentication so action logic stays platform-neutral.

An isolated deployment target such as draft, staging, or production, each with its own credentials, extension settings, and published assistant version. Promotion moves a tested version between environments without rebuilding it.

The separation between the editable working copy of an assistant and immutable published snapshots. Only published versions serve live traffic, so edits in draft never affect users until explicitly published and promoted.

Dashboards and logs reporting conversation volume, containment, fallback rates, top intents, and unresolved questions. They reveal where users struggle and which actions or knowledge gaps deserve redesign.

A held-out, labeled collection of realistic utterances and expected outcomes used to measure routing accuracy and answer quality. It should be refreshed with production-like examples and never overlap with training utterances.

The share of conversations resolved without escalation to a human or an unresolved fallback. It is a headline success metric but must be read alongside satisfaction and accuracy, since refusing to escalate can inflate it.

A policy or filter that constrains model and assistant behavior, blocking harmful content, restricting topics, or preventing disclosure of sensitive data. Guardrails are configured and tested like any other functional requirement.

Detecting and masking personal data such as names, card numbers, or identifiers before it is logged, sent to a model, or stored. It reduces regulatory exposure and keeps prompts free of unnecessary sensitive content.

Assigning permissions by role so that only authorized team members can edit actions, change knowledge sources, or publish to production. It supports separation of duties between builders, reviewers, and operators.

A recorded history of configuration changes, publications, and administrative actions with actor and timestamp. It supports incident investigation, compliance evidence, and accountability for automated or model-driven behavior.