AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

foundry-iq

Foundry IQ knowledge bases. WHEN: make local or Blob documents searchable; create/diagnose KBs; triage unsupported connectors or multi-source KB creation/reconfiguration (even with classic-search terms); connect existing KB to agents (including multi-source KBs); create/reuse a Search service; retrieve from an existing knowledge base with citations; no brand words needed. Read its procedure before query/target questions. NOT: other KB providers, repository-file search, classic Azure AI Search index/query/app work with no Foundry IQ KB or connector request, generic agent creation.

Original source
microsoft
License
MIT
Pinned revision
3f5b461b2314e01c400564051e92469114c5b68f

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read microsoft--foundry-iq at revision 3f5b461b2314e01c400564051e92469114c5b68f, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 45

helpers/bootstrap-contracts.md

Search bootstrap

New Basic/standard Search in an existing group. Other prerequisites/cleanup: native bootstrap.

Progress: stderr; --no-progress disables.

Choices

Python 3.12+, signed-in CLI on Windows/POSIX; no login/install/context switch. No preflight policy reads. Prerequisites: owner-verified references, not helper-verified. 1–256 printable characters; punctuation allowed. Errors omit values.

Creation: search/create. Existing operations: intake; legacy reuse stays strict, never hardens. New locations fold ASCII case/whitespace (East US → eastus); Not availability proof; bodies/hashes stay unchanged. Basic: 1–3 replicas/one partition; standard: 1–12 replicas/1/2/3/4/6/12 partitions. Fixed SystemAssigned, keyless auth and public networking; no role changes. Non-secret tags; creation adds operation/owner tags.

{"schema_version":"2.0","resource_kind":"search","action":"create",
"subscription_id":"00000000-0000-0000-0000-000000000001","tenant_id":"00000000-0000-0000-0000-000000000002",
"resource_group":"rg1","name":"mysearchservice","location":"westus","sku":"basic","replicas":1,"partitions":1,
"public_network_access":"Enabled","tags":{"project":"knowledge"},"owner":"operator",
"prerequisites":{"quota":"checked; enough","pricing":"approved","network":"approved public","requirements":"settings","exclusive_name_authority":"exclusive name"},
"limits":{"command_seconds":5,"wait_seconds":10,"wait_interval_seconds":1},"receipt_dir":"<absolute-private-directory>"}

Replace examples/limits. Reuse: exclusive_name_authority: null; required tags match; no ownership transfer.

Region scope (closed):

{"schema_version":"2.0","subscription_id":"<guid>","tenant_id":"<guid>","receipt_dir":"<private-dir>","limits":{"command_seconds":5}}

No RG/name/SKU: account check plus one az rest GET: /subscriptions/<sub>/providers/Microsoft.Search?api-version=2021-04-01. searchServices.locations supplies canonical available_locations (1–256). Exit 0: discovered, no approval/artifact. Create plan/apply: one fresh region GET each, including retained artifacts; no cache. Unknown: bootstrap-region-unsupported plus choices; denied/malformed: blocked. Reuse skips catalogs. Not SKU/quota/capacity/model/residency proof.

python "<skill-root>/helpers/bootstrap_azure.py" --regions scope.json
python "<skill-root>/helpers/bootstrap_azure.py" --plan choices.json
python "<skill-root>/helpers/bootstrap_azure.py" --apply "<private-directory>/<artifact_id>.plan.json" --approve

--plan uses Azure reads/local private writes: <artifact_id>.plan.json body, before-state, choices, time, unapproved envelope. Stdout: artifact_id/approval_summary, no paths/errors/hashes. Labels: caller attestation, not helper verification; exact references stay private. Show resource/network/capacity/tags/cost/limits; --approve consents once. No cached consent or user checksums.

15-minute expiry; submission markers prevent retries. On drift/failure/expiry/changed choices, rerun --plan; discard consent. Reconcile uncertainty; no suffix-create/body/hash edits. Schema 1.0: regenerate (bootstrap-contract-version). Removed inputs max_polls/poll_seconds/poll_interval_seconds are rejected.

Execution

Argv: az account show --subscription, az group show --name --subscription, Search ARM GET/PUT API 2025-05-01. PUT uses --body then one "@" + str(body_path) argument, Content-Type/client-request-ID headers. POSIX runs az directly; Windows MSI/ZIP uses bundled python.exe -IBm azure.cli. Other Windows layouts block.

Only normally returned target ResourceNotFound proves absence. Fresh principal/tenant/subscription/group/exact-name absence precede one PUT. GET/PUT is not atomic: require exclusive name authority. Mandatory atomicity: bootstrap-concurrency-unresolved, no invented conditionals.

Reuse: two fresh material readbacks, required configuration/tags, keyless identity/principal/tenant, endpoint, provisioning succeeded/status running. reused: no mutation approval/executor or ownership adoption. Wait uses the same identity; existence is not readiness.

After PUT: az resource wait --ids <id> --api-version 2025-05-01 --custom <generated-condition> --interval <wait_interval_seconds> --timeout <wait_seconds>. Fixed properties: succeeded/running or terminal states, lower/title/uppercase. No caller expressions/custom polling. Final raw GET establishes material match/ownership before readiness. Wait failure remains primary even with ready GET; provider evidence is secondary.

Limits: command 1–60 seconds, wait 1–900, interval 1–30. Wait/final GET: separate wait_seconds/command_seconds; sum is not an end-to-end deadline. subprocess.run: argv, shell=False, best-effort timeout/cleanup. CLI/OS/filesystems prevent hard time/process-tree/request-count guarantees. No PID or confirmed-cleanup claim. Captured output is rejected after capture above 1 MiB per stream: not a streaming cap or memory bound. Input/body/receipt limits remain 1 MiB. Timeout/execution/receipt failure after attempted PUT is partial.

Checks: target, principal/tenant, capacity/SKU, auth/network, identity, required tags, ARM readiness. Metadata/ordering/unrelated tags, ID/enum casing, equivalent optional defaults do not need another approval. Never default missing material fields or return extra tags. Material conflict, unknown ownership or failed readiness blocks/returns partial.

Evidence and failure

Existing absolute private directory outside the plugin. No symlinks/reparse paths. POSIX: directory-FD creation, identity/owner/private-mode rechecks; exclusive safe leaves, mode 0600. POSIX semantics required; later pathname uses are not pinned. Windows checks owner/DACL (owner, SYSTEM, Administrators), not handle-relative creation; chmod does not establish Windows privacy. No ACL changes. Same-user/admin tampering is possible; unprovable privacy blocks. Mocked/DrvFS is not native-filesystem or Azure evidence.

Flushed receipts: operation/target/body/argv, client-request ID, readbacks/digests, first error code/status/request ID, ownership. No raw errors/output/policy values/tokens/documents. Message digest/message_withheld; secret-free labels/tags. CLI file logging/dynamic installation: process-disabled.

Creation-only policy diagnostics: RequestDisallowedByPolicy, supported policy IDs or policy-violation details. Provider error/statusDetails, never tags/configuration. A bare 403/RBAC/network/quota failure triggers no policy calls. Preserve first failure; diagnostic/receipt errors are secondary warnings. At most eight referenced assignments/selected-subscription definitions within 60 seconds: az policy assignment show --name --scope and az policy definition show --name --subscription. No lists or global scans. Built-ins/management groups/initiatives/exemptions/missing references need the policy owner. No compliance/applicability proof, policy changes, write retries or correction approval.

Exit 0: planned/fresh reused/applied completed (ARM only). Data-plane access/ingestion/retrieval stay unverified. Exit 2: blocked, no confirmed/ambiguous Azure writes. Exit 3: partial, first_failure, attempted_writes, resources_remaining; unproven ownership is unverified, never run-owned. Receipt failure: block before submission, partial handoff afterward. No automatic rollback/deletion. Cleanup always needs separate consent and fresh ownership/configuration/children/scoped-role evidence.

Authorities: failure/conflict/uncertainty only.

Native wait; location metadata; provider GET.