Search Substrate

Discovery-first recommendations

Reuse resolved prompt/session/workspace choices; do not restart intake on handoff. Read az account show for caller/tenant/default subscription; explicit scope wins. Use that default unless overridden, never silently switch CLI context. Account context does not prove live access; missing CLI/sign-in or denied reads are blockers, not evidence that resources need creation. Pass chosen --subscription. Reuse supplied resource/intent, else ask USE EXISTING / FIND CANDIDATES / CREATE NEW before enumeration: resource intake. Only FIND lists, known RG before selected sub. Only a new group decision needs az group list --subscription <sub>. Do not fan out resource scans across all groups or subscriptions.

Inputs are required before execution, not necessarily questions for the user. Read IDs, roles, network state and inventories at the appropriate selected scope. Offer observed reuse/create/choose-another options and recommended names, region/capacity/costs; ask only the unresolved decision. Proposed owner/settings are not verified facts. Missing evidence blocks execution after discovery/handoff. Selection is not mutation consent; preserve valid concrete approval, never extend it to unknown future actions. Keep local roots and Blob container/folder boundaries explicit; discovery does not authorize uploads or whole-container ingestion. During source setup, not every retrieve request, source owners must learn content structure and answer needs before recommending unresolved Minimal/Standard: authorized representative samples or customer description, never filename/MIME inference or a silent Minimal default. Honor explicit choices and reuse supplied information; unavailable inspection means ask, not assume text-only. Scans/important visual or table layout can justify CU; a table alone does not require it. New unresolved search: recommend Hybrid; confirm tradeoffs. Hybrid needs selected embeddings. Neither choice determines KB reasoning effort or synthesis. Existing retrieval remains read-only; missing vectors: approved setup, never a silent configuration change.

Read exact dependency IDs first. Enumerate only FIND selections within the chosen scope, all pages. Offer choices rather than demanding remembered IDs/model names. Never scan unrelated resource types/subscriptions. Reuse resolved choices, not stale verification; retain complete selected-boundary inventories and fresh pre-write checks. Search owner separates operation-compatible reuse from new-service defaults; no existing keys/identity/network/tag changes to fit provisioning. Other selected dependencies still require material region/SKU/capacity, identity/network/auth/tags/model/version/access readback. Never alter shared configuration to fit. Minimal lexical File needs no model; vectorization adds embeddings; synthesis/nonminimal KB reasoning adds chat; standard File adds CU, with independently selected embeddings/chat, not a KB. Skip model discovery on minimal lexical/extractive paths without model dependencies.

Azure Policy diagnostics after creation failure

Do not enumerate or refresh policy as a preflight for plans, approvals, creation, delegation, retrieval or reuse. Ordinary resource, identity, quota, network, cost and definition checks still apply. Preserve known organizational requirements and user-approved security settings; no policy scan is not a compliance claim. Do not knowingly submit a noncompliant request just to obtain an error.

Only read policy when an Azure resource creation failure gives evidence that policy may be responsible, such as RequestDisallowedByPolicy, a policy assignment/definition ID, or policy evaluation details in deployment errors. First inspect the actual failure: a 403 alone is not proof of policy denial. Authentication, RBAC, network, quota and invalid-body errors use their own diagnostics, not a policy scan. Distinguish ARM from data-plane operations.

Preserve the original status/code/message/request ID, operation, exact target, approved body and completed writes. Follow referenced assignment/definition IDs and relevant initiative members first. Only if that evidence requires it, inspect assignments at the failed target’s scope and applicable ancestors, following every page within that narrowed scope. Never audit the whole subscription or management-group hierarchy by default; no global azqr scan. Resolve relevant parameters/defaults, effects, modes, enforcementMode, overrides/selectors, notScopes and exemptions/expiry only to explain this failure. Audit/disabled/non-enforcing policy is not an enforced denial.

Missing/unreadable policy details are diagnostic warnings, not new pre-creation blockers or proof of compliance. Keep the original creation failure unresolved if the cause or remedy cannot be established; hand off exact missing IDs to the resource/policy owner. Never turn an actual Azure denial into a warning/success.

Propose only a compliant correction, retaining exact required settings and tag values from authoritative input or one focused question. Never invent an owner or silently change region/SKU, permissions, networking or cost. Bind diagnostic evidence and body changes to a new plan_fingerprint and confirmation; pass them to delegated owners. A changed plan needs approval before retry. No unconditional policy refresh is required. Read back required settings and policy-applied changes after creation.

Known noncompliance (azure-policy-noncompliant) or a known required setting that cannot reach the request (azure-policy-setting-unsupported) still blocks. Never disable policy, create an exemption, alter an assignment, or retry through another name/surface to bypass it. Reconcile ambiguous writes read-only before any retry; never assume a failed command created nothing. If no writes occurred, use the shared blocked shape: blocked_at: verification, original code/message in first_blocker, HTTP status in first_blocker.status, request ID in first_blocker.request_id, and diagnostic warnings. Unavailable status/request IDs are null. After completed or ambiguous writes, return partial with remaining ownership. Cleanup remains separately approved.

Return dependencies/ownership and content assessment/uncertainty, answer needs and source mode for the calling KB owner; resume KB plus validated retrieval, never downscope to infrastructure/source-only. Technical-plan approval does not resolve unknown content structure; unavailable CU never implies Minimal. Preserve approval for its concrete unchanged Search artifact; never preapprove unknown bodies. Source/KB mutations and cleanup need their own approvals. Checkpoints are partial; preserve shared/reused resources.