Before You Tune Maximo Assistant, Define What a Correct Answer Includes

A faster Maximo Assistant response is not an improvement if it loses the records, business context or evidence users need. IBM's documented configuration arguments provide a practical starting point for testing response time and correctness together.

Share
Before You Tune Maximo Assistant, Define What a Correct Answer Includes

Before You Tune Maximo Assistant, Define What a Correct Answer Includes

When a conversational assistant feels slow, the easiest response is to blame its model. In a Maximo environment, that explanation can skip the more important question: what work did the request actually ask the system to perform? Finding operational records, selecting relevant business context, generating an explanation, searching documentation, and producing condition insights are not interchangeable tasks. Measuring them as if they were the same workload can produce a reassuring average and an unreliable service.

IBM's current Maximo Assistant documentation gives configuration owners a concrete way to examine that boundary. The enabling guide describes arguments that influence domain-rule selection, schema selection, relevance screening, and whether a response includes the retrieval query. It explicitly warns that some settings can affect performance and says that schema filtering might improve response time. Those are documented possibilities, not measured improvements for your environment.

The architecture guide adds a necessary distinction. The non-agentic assistant uses data retrieval, while the agentic assistant coordinates tools through an ALM agent, AI Service, and the Maximo MCP server. A timing comparison that ignores which assistant path was active can confuse an architectural difference with a configuration regression.

This article proposes a controlled acceptance method built around those documented behaviors. It does not offer a benchmark, a guaranteed tuning recipe, or a claim that a particular setting should be enabled everywhere. The objective is to measure a useful answer: the right records, within the right access boundary, with enough context to support the intended decision. Once that contract is explicit, response time becomes something a team can investigate without quietly lowering the standard of correctness.

Identify the Request Path Before Measuring the Model

IBM's architecture guide describes the non-agentic assistant as using a single data retrieval tool without the ALM agent. A valid prompt is processed into a SQL representation and then converted into a REST OSLC query against a Maximo Manage object structure. AI Service passes the query and does not directly access the database. Results return as JSON and are displayed through a schema-backed dynamic table. Those details matter when somebody describes the assistant as simply writing SQL against production.

The same guide describes a different path for the agentic assistant. A prompt reaches the ALM agent, AI Service connects to the model, the agent selects tools, and the MCP server routes tool calls to the appropriate Manage API. Results return through the server and agent before the natural-language answer reaches the chat window. A user-facing pause can therefore include activities beyond generating the final prose. The documentation does not publish a latency breakdown for those activities.

Start an evaluation by recording which path the deployed assistant uses. Also record the enabled capabilities, relevant AI configuration, object structures, user role, and the operational question being tested. These are recommended test metadata, not a claim that IBM exposes every field through a single diagnostic interface. If a team cannot establish a particular detail from supported administration or diagnostic evidence, mark it unknown rather than inferring it from the answer's style.

Keep request families separate. An operational record lookup should be compared with an equivalent lookup. A documentation question should not quietly join the same performance sample as an asset condition investigation. A compound request may be useful to users, but it needs its own expected result. Otherwise, an apparent slowdown could simply reflect that one request required additional work that another request did not.

IBM's user guide offers a practical discovery step: click the AI icon to understand what the assistant is configured to do and, when tools are mentioned, ask what it can do. Treat this as an initial user-level check, then confirm the configuration through the administration process. The assistant's own description is helpful context, not a replacement for an approved configuration record.

For every test, define where the user-visible measurement starts and ends. A team might measure from submission to a complete usable answer and separately record visible progress or completion notification. That is a proposed measurement convention, not a built-in IBM metric. Apply it consistently and avoid claiming internal stage timings unless supported observations actually establish them. The result is a workload description that makes a later tuning comparison interpretable instead of merely numerical.

Treat Configuration Arguments as Testable Hypotheses

The enabling guide documents use_domain_rule_filter as a way to limit the prompt-tuning rules sent to the model. When set to true, the model performs a hybrid RAG search over those prompts and extracts the most relevant domain rules, up to the documented limit of 50. IBM says this might increase response relevance and accuracy. It does not promise a universal response-time improvement, and the rule limit should not be mistaken for a guarantee that the right business rule will always be selected.

For a controlled evaluation, identify questions that depend on local terminology or a specific domain rule. Define the expected interpretation before changing the argument. After the change, compare not only elapsed time but also whether the answer preserved that interpretation. A faster response that stops recognizing a locally significant code is not a successful optimization. This is an acceptance recommendation derived from the documented selection behavior, not a reported product defect.

The same guide describes use_schema_filter as limiting review to the most relevant business objects and attributes, selected based on the user's prompt. IBM says setting it to true might improve response time. Test that possibility with narrow questions and with questions whose meaning depends on a relationship between objects. A useful result must preserve the required relationship, not merely return a plausible table more quickly.

use_explanation determines whether a response includes the query used to retrieve data. IBM notes that including the query can affect performance. Turning it off may remove evidence that technical reviewers rely on during acceptance. Decide explicitly whether the production audience needs that evidence, whether a separate approved review process provides it, and whether the performance comparison uses the same evidence requirement on both sides.

use_router_filter concerns screening prompts for relevance before processing. The documentation identifies false as the default and best-practice setting, and warns that enabling screening can affect performance. That is not an invitation to change it merely because the argument exists. Establish the reason for testing it, the questions it should accept, and the questions it should reject. Do not generalize a result beyond the configuration and workload actually evaluated.

The guide also notes that some earlier assistant versions do not expose configurable arguments. Check whether Set arguments is available before preparing a change. Record the existing values, follow the documented deactivation and reactivation process for configuration edits, and make one deliberate comparison at a time. The objective is a reversible experiment with an explainable outcome, not a bundle of simultaneous changes followed by a hopeful interpretation of whatever becomes faster.

Include a deliberate negative case in that experiment. Choose a question outside the approved scope and record what an acceptable rejection or limitation should communicate. This checks that an apparently favorable configuration has not achieved its result by treating every question as answerable. Preserve the same negative case when repeating the comparison, and have an operational reviewer approve its expected handling before interpreting the result.

Protect Data Meaning While Narrowing the Workload

IBM advises users to start small and scale up because the assistant can analyze a limited number of records at a time, with the limit set by the administrator. Its user guide recommends breaking broad requests into smaller, more specific questions. That is useful guidance for performance testing, but it introduces an acceptance trap: a narrow answer can appear faster simply because it no longer answers the original operational question.

Define the scope in business terms before reducing it. Specify the site, object type, relevant state, and decision the user needs to make. Then identify the expected records through an authorized, independently reviewed application or query path. This reference set is a recommended testing control. It should not depend entirely on another assistant answer, because that would make the assistant both the system under test and the authority used to judge itself.

Consider an illustrative request for overdue preventive maintenance at a named site. A comparison should establish what overdue means locally, which records the chosen user can access, and whether the response actually includes the relevant population. A short answer that silently narrows the date range or omits a class of records may be less useful even when its wording appears confident. This scenario is hypothetical and describes a test to run, not a known behavior in a particular deployment.

Object-structure configuration is another important boundary. The enabling guide distinguishes the data access available in 9.1 from the configurable searchable data available in the 9.2 feature channel. It describes enabling and disabling object structures and selecting default query templates. Those templates determine the table column headers returned with data. Confirm the deployed channel and supported configuration before applying instructions from a different version's example.

Columns deserve business review. Removing information that users need to interpret a result can shift work out of the measured chat response and into subsequent manual investigation. The response becomes shorter, but the user's task may become longer. Include enough context in the acceptance contract to establish which asset or work record a row represents and what action the recipient can reasonably take with it.

Permissions must remain part of the test. IBM states that users can receive only information they are authorized to access and that different users can receive different answers to the same question. Compare like-for-like roles when evaluating a configuration change, then run separate access-boundary tests. A smaller answer from a restricted account is not evidence that the assistant has become more efficient. It may simply reflect the correct security boundary. Performance acceptance should preserve that boundary, not treat it as inconvenient variation to remove from the analysis.

Separate Retrieval, Documentation, and Insight Freshness

IBM's architecture guide describes documentation search as a tool using a static, curated set of IBM Documentation articles. That is not a statement that the assistant searches every current support page, community post, local procedure, or newly published release note. When a user asks an operational question that depends on product behavior, acceptance should establish what kind of source the answer actually used and whether that source supports the proposed action.

The user guide recommends explicitly saying when a request should search IBM Documentation or use general knowledge. Use that distinction in test prompts. A request to retrieve work records and a request to explain a documented product feature should have different expected evidence. Do not reward an elegant general explanation when the acceptance case requires records from the local system, and do not treat a returned record table as an adequate answer to a configuration question.

Documentation references have a deployment-specific limitation that deserves explicit testing. IBM says the reference links typically returned by documentation search do not work in on-premises or air-gapped environments. This does not by itself establish that the assistant cannot produce a documentation-based response there. It establishes that a reference link is not a complete evidence-delivery mechanism for those environments. Reviewers need a supported way to inspect the relevant documentation before approving an operational instruction.

Condition insights introduce a different freshness question. IBM's user guide says asset insights are not automatically updated. They update when a user asks the assistant to generate insights for a specific asset or refreshes them from the asset dashboard's Insights card. If an evaluation asks about asset condition without accounting for that behavior, it can mistake an older insight for an answer derived from the latest operational evidence.

Separate an existing-insight question from an explicit insight-generation request. Record what each test is intended to prove, and establish the state of the relevant asset information through authorized application evidence. A faster answer that reuses an older insight should not be compared with a request that generates new analysis as though the workloads were identical. The documentation supports the freshness distinction, not a specific timing relationship between the two requests.

Finally, preserve the boundary between advice and action. IBM's user guide says the assistant cannot automate tasks or create or edit system data unless an administrator has configured a custom tool for that purpose. A fast answer that describes a proposed action is not evidence that the action occurred. If the local deployment includes such tools, evaluate their authorized effects separately. For this response-quality exercise, the essential discipline is to identify which evidence was retrieved, which knowledge was consulted, which insight was refreshed, and which changes, if any, were actually confirmed.

Practical Implications

Create an acceptance sheet before opening a tuning change. For each representative question, record the assistant path, user role, relevant configuration, expected scope, required evidence, and definition of a complete response. Include narrow lookups, relationship-dependent questions, documentation requests, and condition-insight questions only when those capabilities exist in the deployed environment.

Preserve the baseline configuration and change a single relevant argument for each comparison. Use repeated observations rather than selecting the quickest successful example. Report both correctness outcomes and user-visible timing under a consistent measurement convention. If diagnostic evidence cannot isolate internal stages, state that limitation rather than attributing the whole delay to the model, the network, or Manage.

Include a usability check outside the chat window. Can the user identify the correct asset, verify a documentation claim, and complete the intended review without hunting for missing context? A change that moves effort into a separate application may improve the measured response while worsening the actual task. Let operational reviewers judge that tradeoff explicitly.

Keep restoration simple. Record current argument settings before changes, use the supported configuration lifecycle, and retain the test evidence needed to explain the decision. A tuning change is ready for adoption when it provides a useful, authorized answer more consistently within the agreed service expectation. Faster text alone is not that result, and a single convincing demonstration does not establish it.

Bottom Line

Maximo Assistant performance is an application acceptance problem before it is a model selection argument. IBM documents distinct request paths, configurable context and evidence choices, record limits, access boundaries, and insight-refresh behavior. Each can change what work a request represents or what evidence a user receives.

The safe tuning method is straightforward: define the operational question, establish an independently reviewed expected result, record the actual configuration, and compare equivalent requests. Treat arguments such as schema and domain-rule filtering as hypotheses to test, not universal acceleration switches. Keep the query evidence, documentation accessibility, and insight freshness that the decision requires.

The outcome to optimize is a complete task supported by an answer the user is allowed to see and can reasonably verify. If a configuration produces that outcome sooner, the improvement is meaningful. If it merely produces less evidence or a smaller population of records, the stopwatch has measured a compromise rather than a better assistant.

Sources: Maximo Assistant architecture, Enabling Maximo Assistant, and Using Maximo Assistant. Test design and acceptance recommendations are editorial analysis; no benchmark or production tuning result is claimed.