Intent Routing

Route by the user’s end result, not the first connector/resource keyword. Within Foundry IQ, “make these documents searchable” or “ask questions over these documents” means a usable KB with validated retrieval. Default to Create KB, not a source-only operation. Words such as files, Blob, index or embeddings identify prerequisites; they do not downscope a broader goal.

Confirm the intended finish once in normal intake, in user terms: “I’ll make these documents searchable through a knowledge base and check retrieval against them. Do you want that full outcome, or only the knowledge source?” Do not re-ask an already explicit or confirmed goal. Use source-only only when the user explicitly requests or confirms that narrower outcome; never infer it from an intermediate step. Goal confirmation is not approval for Azure writes.

Retain the confirmed goal, selected data, explicit exclusions and remaining steps across handoffs. A child’s completed result is not overall completion. For a KB goal, resume the KB owner after Search/model/source prerequisites with returned IDs and fresh source/readiness evidence. Present and approve the concrete KB plan, create/reuse and read back the KB, then run supported-question and unrelated-question retrieval checks under the selected API/output mode. Require source-backed results and original-source citations; source ingestion, KB existence or HTTP success alone is not validated retrieval.

Finish only when every requested outcome is verified. Otherwise report completed prerequisites, the blocker and remaining steps; never mark the whole goal complete or silently downgrade it to KS-only. Resume after an approved prerequisite rather than asking the user to restart. Never preapprove an unknown KB mutation. Explicit source-only, infrastructure-only and read-only requests remain bounded; never create an unrequested KB/agent. Explicit classic Search work still routes out.

Search-mode choice

For new unresolved search-mode intent, recommend/preselect Hybrid (keyword + vector), with Keyword-only (disable vectors) as the explicit opt-out. This recommendation is not consent or a helper default. Honor existing/informed keyword-only choices and exact reuse unchanged: do not re-ask settled choices, silently retrofit vectors or choose/deploy a model. An explicit change to an existing source needs a separate supported plan/approval; read-only retrieval never changes its source.

Ask in the grouped intake: “I recommend Hybrid (keyword + vector) for keyword precision plus semantic/paraphrase matching. Would you like Hybrid, or Keyword-only without vectors?” Where selectable defaults are supported, preselect Hybrid, but require an explicit customer choice; silence or an unanswered field is not consent. Reuse supplied informed choices and group compatible unresolved decisions rather than adding separate confirmations for each disclosure.

Before plan approval explain:

Keyword-only: no embedding discovery. Discover only missing dependencies for the confirmed hybrid choice, using the bounded scoped helper; select an account before listing its deployments. Known compatible model choices need exact readback, not rediscovery. Never auto-select a deployment, deploy models, call models or upload content during intake. Missing/unavailable hybrid dependencies require supported scoped discovery, concretely approved setup, or explicit customer keyword-only opt-out; no silent keyword fallback. Report readback/access/capacity uncertainty precisely; metadata is candidate evidence, not effective readiness. Model-free minimal+extractive routes do zero model/CU discovery; keyword-only does not suppress independently required chat/CU discovery.

Retrieval keyword+vector is independent of text-only versus multimodal extraction. Neither search choice selects CU, KB reasoning/synthesis or semantic reranker automatically. A text-only PDF can use hybrid; multimodal content can use keyword-only. Assess extraction/answer dependence separately without defaulting Minimal or selecting Standard merely because an image/table exists.

Keep helper inputs explicit: File vectorization: none versus azureOpenAI; Blob minimal-lexical versus minimal-vector, or standard-cu with independently selected optional embeddings. File/Blob JSON examples are explicit keyword-only choices, not UX defaults. Omitted/null embeddings never become vectors; a selected vector mode with missing model choices blocks, not downgrades. Preserve separate sample-content consent and actual resource/permission/CU/source/KB approvals; do not approve unknown future bodies. Approved source creation/ingestion is not validated hybrid retrieval; use the existing vector/hybrid verification contract with explicit query approval, not a billable intake probe.

Scenario examples

Read the whole request; broader goals and explicit limits take precedence over a mentioned child step.

References

Authorities: failure/conflict/uncertainty only.

• Hybrid search (opens in a new tab)
• Integrated ingestion/query vectorization (opens in a new tab)