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/blob-contracts.md

Blob/ADLS contract

Blob progress: --no-progress disables stderr.

Planning

python helpers/blob_source.py --plan intent.json

Intent:

{"schema_version":"1.0","endpoint":"https://svc.search.windows.net","name":"manuals","owner":"[email protected]","storage_id":"/subscriptions/00000000-0000-0000-0000-000000000000/resourceGroups/Example/providers/Microsoft.Storage/storageAccounts/example","container":"documents","prefix":"Reports/","is_adls":false,"api_version":"2026-04-01","processing":"minimal-lexical","network_access":"public","identity":"system-assigned","permission_options":[],"ingestion_schedule":null,"description":null,"rbac":{"assignments":[{"id":"<observed-assignment>","principalId":"<observed-principal>"}]},"network":{"posture":"public","evidence":"<fresh-readback-reference>"},"inventory_limits":{"max_pages":100,"max_objects":10000,"max_requests":25000,"deadline_seconds":600},"poll":{"deadline_seconds":600,"max_requests":120,"interval_seconds":5}}

Content fit (no default): minimal-lexical=minimal/no vectors, minimal-vector=minimal/vectors, standard-cu=standard/optional vectors; not KB reasoning. Public/system-assigned; no schedule settings/permission ingestion. Blob: 2026-04-01 or 2026-08-01-preview; ADLS: preview, is_adls: true, directory prefix without trailing /. Root: explicit prefix: "", no widening. Description: text/null; other modes: planning-processing-unsupported.

No writes/auth changes/installs/polling/cleanup. Verify access/cost/data preapproval.

Two exact Search GETs bracket two-pass Storage observation. Reuse requires matching definition/ETag, Storage binding and four generated identities. reuse_input_file/reuse_result_file: retained approved input/successful creation result. Validate pairing/types before auth; load only for redacted binding. Required fingerprints/boundaries/source ETags/definitions/generated identities must match fresh readback. Unproven binding: source-binding-unverified. Old consent/readiness never authorizes new work.

Exit 0: status: planned, plan_fingerprint, execution_input, approval_summary, request IDs and writes_performed: []. Save only execution_input; its approval is false and the executor rejects it until actual unchanged-plan approval. Private artifacts preserve ResourceId connection strings/fingerprints; ordinary result redaction is unchanged. Summaries omit paths/ACLs/hashes; no readiness. Exact reuse sets execution_required and mutation_approval_required false: no executor or approval needed. Drift: rerun --plan, replace the artifact, discard consent; never hand-edit nested plans/hashes. Cleanup stays separate and run-owned-only.

For standard-cu, read CU choices/wire. content_understanding binds CU; optional embedding selects source vectors.

Discovery

python helpers/blob_source.py --discover <boundary.json> accepts this closed JSON (no approval):

{"boundary":{"storage_id":"<ARM-account-ID>","container":"<container-or-filesystem>","prefix":"Reports/","is_adls":false},"inventory_limits":{"max_pages":100,"max_objects":10000,"max_requests":25000,"deadline_seconds":600}}

Arbitrary partial-name prefixes: boundary-ambiguous. Public cloud/private DNS.

Read all pages twice including empty continuations. Exit 0 returns status: discovered, mutation: none, writes_performed: [], boundary/account, sorted objects (URL/path/size/ETag/version), ADLS path/property digests, request IDs, inventory_digest; not approval/ingestion. Incomplete/duplicate/out-of-scope/unreadable/unstable evidence blocks. ADLS HEADs path/ancestor ACLs; no raw ACL/content. Operator access is not Search access. Every path has an ACL HEAD; every non-root path also has a properties HEAD for type and matching ETag. Root has no type header. Preserve ancestors and normalize quoted/unquoted Blob/DFS ETags only for comparison; never skip required reads.

Inventory bounds apply per discovery: pages/objects per pass; requests/time across both passes, including empty pages and all HEADs. Maximum accepted limits are 1000 pages, 100000 objects, 200000 requests and 3600 seconds; exceeding any bound blocks, never truncates. Planning adds at most two Search GETs and shares its time budget with discovery. Bodies are capped at 8 MiB plus one overflow probe byte. Socket watchdog interrupts body reads; DNS/connect/header acquisition and scheduling are not hard wall-clock bounded. Redirects and unsupported deadline readers block. Limits are safety bounds, not spending consent.

Apply

python helpers/blob_source.py --input <approved.json> uses:

{"schema_version":"1.0","plan":{},"approval":{"confirmed":true,"fingerprint":"sha256:<canonical-plan-digest>"}}

Canonical JSON: UTF-8, sorted keys, ASCII-escaped/compact. operation: reconcile-and-monitor, owner, cleanup_approved: false, discovered boundary, inventory_digest, inventory_limits, source, poll: {"deadline_seconds":600,"max_requests":120,"interval_seconds":5}. Planner reuse: expected_generated, four typed/name/service-managed identities, checked before writes; legacy artifacts remain valid. Planner creation sets expected_source_absent: true; an intervening source blocks. Redacted GET requires the PUT ETag (body/header); conflicts block. ambiguous unproven ownership remains partial, never adopted or cleaned up.

source: operation: reconcile, resource_type: knowledge-source, exact endpoint/name/api_version, action: create or reuse, matching owner, cleanup_approved: false, approved desired, and source_evidence with verified: true and the same inventory_digest. ADLS also requires path_verified: true, acl_verified: true attest metadata, not effective Search permissions.

desired: matching name, kind: azureBlob; azureBlobParameters: connectionString: ResourceId=<exact-storage-resource-id>, containerName, exact folderPath (null for explicit root), boolean isADLSGen2, ingestionParameters. Optional shared controls. No keys/SAS/updates/credential environments or supplied createdResources/assetStore.

Recompute approval/evidence; block drift. Create conditionally/reuse zero-write; supplied ETag must match. Poll GET knowledgesources('{name}')/status within bounds. Only 408/429/500/502/503/504/transport-ambiguous reads retry; 403/404 block. Watch expiry pauses; it does not fail ingestion. Backoff and explicit GET-only continuation/compact output: recheck.

Require completed start/end times at or after the pre-write cutoff, nonnegative integer itemsUpdatesProcessed, itemsUpdatesFailed, and itemsSkipped, zero failures, no observed relevant errors or active cycle. Missing counts are not zero. Completed status may be absent; reject partialSuccess/failure/malformed/unknown. Never trigger indexers/edit schedules; stale success cannot pass. Reuse also excludes the first observed completion.

Results

--receipt-dir: PUT acknowledgement before GET. Binding/recovery: redaction, recovery, reuse; never reconstruct cutoffs. Timing differences: indexer checks.

Exit 0: completed; unverified zero-write reuse is exit 2, blocked with reconciliation: completed. After a write, exit 3 / partial retains approval/first failure/resources_remaining.run_owned. Conflicting children aren't owned. Return generated identities/readiness/digests/ warnings/cleanup; errors retain status/request ID, not sensitive text. No rollback/source writes/child edits.

Owner handles RBAC/private links. CLI Entra: ARM/Storage/Search audiences. Probes: ARM GET 2025-06-01, Storage 2025-05-05 listing, ADLS HEAD action=getAccessControl&upn=false.