Technology Participation in Enterprise Intelligence — browse technologies by architectural contribution.
From an analytical question to evidence fit for its purpose
Arqua architectural interpretation of selected public documentation, reviewed 28 September 2026. Not a deployment assessment or certification.
An analytical agent is most useful when it connects a business question to well-defined information and appropriate computation. Databricks Genie offers mechanisms for curating that connection rather than requiring every user to reconstruct the data model or write SQL.
This explanation follows a bounded progression: curated context → selected analytical logic → executed result. It focuses on structured-data Chat mode, with separately identified inspection, API and metric-view contributions. It is not an assessment of the entire Databricks platform, Agent mode, unstructured analysis, Genie Code or cross-platform federation. Unselected capabilities are not presumed absent.
Connect local context to accepted enterprise meaning
The Genie concepts reference describes selection from table metadata, agent-local knowledge, SQL examples and conversation history. Agent-local metadata does not alter the underlying Unity Catalog metadata, and token limits can exclude older conversation content. The configured knowledge, context selected for one question and resulting answer are therefore different records.
Unity Catalog metric views provide a complementary way to reuse analytical definitions across SQL, dashboards and Genie. They can reduce repeated implementation of a calculation; they are not a mandatory component of every Genie deployment.
The metric-view metadata reference describes names, formats and synonyms and specifies Runtime 17.3 and YAML 1.1 for that documented feature. Preserve the applicable feature version rather than generalise one prerequisite to all Genie functionality.
The enterprise can govern accepted definitions within these native artefacts. What matters is who accepts a definition, its purpose and effective revision, and how dependent consumers handle change—not a requirement to copy every definition into another repository. A shared calculation still needs an agreed population, grain, unit and time interpretation.
Match trusted logic to the actual question
Genie's trusted-asset explanation links the label to author-provided queries or functions and their parameters. It also warns that Genie can choose the wrong query or function for a question. Correct reusable logic and correct selection of that logic are separate assurances.
For example, a valid calculation for completed work orders does not answer a question about all initiated work orders. Review the selected definition and parameters, not just whether the result carries a reassuring label. An ambiguous question may need clarification rather than an immediately available total.
Context also depends on the entry point. The curation guidance distinguishes an existing agent linked to a dashboard, which does not inherit dashboard filters, from automatically created dashboard agents, for which those filters work. Do not assume the visible dashboard selection became the analytical question's population.
Bind source scope, identity and shared helper information
The newer official indexed setup description, dated 24 September 2026, restricts queries to explicitly attached sources while retaining each user's Unity Catalog permissions. An older directly retrieved rendition described broader source access. The newer account is evidence of a documented boundary change, not proof of rollout in a particular workspace. Confirm the effective configuration before relying on either behaviour.
The setup documentation also distinguishes author-provided warehouse access from the user's data permissions. Knowing which compute can run does not establish which information a caller may receive. A source attachment is not, by itself, a permitted-purpose declaration.
The quality-tuning reference describes representative values generated with author permissions and shared as helper context. It excludes filtered or masked base tables from prompt matching and requires authors to disable entity matching for views over those tables and for dynamic views. Hiding a column from context is not an access-control change. These are documented configuration responsibilities, not findings from an inspected tenant.
The AI trust reference identifies prompts, metadata, relevant values and authored guidance among model inputs. Review the eligibility of that material separately from the final SQL result. A query-time control should not be assumed to explain every earlier helper-context or model-input path.
Use inspection and benchmarks as scoped evidence
The concepts reference labels Inspect as Public Preview. It describes checks of selected SQL details, including filters, dates, joins and aggregations. This is useful native query-review support, not evidence that every condition for relying on an answer has been assessed.
Chat-mode benchmarks compare generated SQL or results against an authored SQL answer under documented comparison rules. Questions run as new conversations, so they do not reproduce threaded context. Generated expected SQL needs review; questions without it require manual assessment. A configured benchmark is not a completed evaluation.
Establish the expected business meaning independently of the answer being tested. A candidate and its expected SQL can agree while both omit cancelled records or apply the wrong reporting calendar. Preserve what was tested, the actual result, unresolved checks and the intended use of the score.
Within Arqua's Runtime Context Assembly and Context Qualification definitions, formation and qualification remain distinct responsibilities within RCA. Native inspection, benchmarks and domain review may contribute to qualification. A benchmark label is not a universal qualification outcome or permission for an operational action.
Keep the interface and returned evidence explicit
The Genie API reference guide separates user and service-principal authentication, progressively populated message attachments, reply completion and query-result retrieval. A client should bind the conversation, message, generated statement, parameters and returned result to the principal actually used. An intermediary service does not automatically retain an original end user's identity.
The documented Conversation API does not reproduce the UI's two-phase preliminary-answer-then-inspection presentation. That is a specific interface distinction, not a finding that every API path lacks all checking. Assess the actual route rather than assume the same evidence is available through every surface.
Configuration export can support migration and review, but exporting instructions is not proof of the context selected for an individual response. A completed message likewise does not establish that the consuming application retrieved all the result evidence it requires.
Preserve the difference between rerunning and explaining history
The user guide distinguishes revisiting a conversation from refreshing its result. It documents seven-day query-result availability; rerunning SQL or regenerating an answer is a new operation, not guaranteed recovery of the earlier result.
Where a past answer informed a decision, retain the evidence needed to explain that use under an appropriate retention and access policy. This can include the accepted definition, source basis, relevant times, statement, parameters, returned population and review outcome. Do not assume either that conversation history preserves everything or that all sensitive results should be retained indefinitely.
A saved answer does not automatically become accepted Institutional Memory. A correction should preserve what was relied upon previously and identify the revised interpretation. Likewise, updated helper values, changed metric definitions and newly loaded records can affect an answer in different ways even when the agent's name is unchanged.
Begin with one bounded service-review question
A proposed read-only test asks for a service-performance measure over an agreed period and population. Establish a reviewed calculation and expected exceptions before asking Genie. Then vary the wording, introduce a closely related but inapplicable trusted query, change a dashboard filter and supply a delayed source update. Assess whether the answer remains bound to the intended meaning and actual evidence.
Test the selected user and service-principal routes separately in an authorised environment, including an unattached source and restricted helper-context case. Retain useful independently supported findings while withholding conclusions whose required evidence is missing. Any downstream remediation requires its own current authority and target-system controls.
Start with existing governed datasets, definitions and SQL. A fixed report or reviewed query may be sufficient; add conversational selection where it creates value. No Databricks account, query, benchmark, API client or installed Guide was operated for this explanation. The documentation has mixed retrieval ages and does not verify a tenant's features, permissions, performance or residency.
The objective is a traceable connection between the question intended, the logic selected and the evidence returned—so that a trusted analytical asset is not mistaken for an automatically qualified answer.