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

references/blob-readiness-recheck.md

Blob readiness recheck

Retained evidence

New minimal/Standard/CU, lexical/vector and admitted legacy modes/schedules: Blob GA 2026-04-01 or preview 2026-08-01-preview; ADLS requires preview/ACLs. Keyless/system-assigned only; no asset stores, new settings or other connectors. Retain original approved input; --receipt-dir requires an existing private directory outside plugin. Check privacy/context before cloud work. No mkdir/ACL changes/overwrite. Conditional PUT acknowledgement precedes post-write GET. Binding/recovery owns acknowledgement recovery. Immutable <operation_id>.blob-readiness.json binds approval/owner, pre-write cutoff, context, generated identities/ETags/digests and request IDs. No credentials/text/ACLs. Persistence failure retains partial ownership, without polling/write retry/cleanup.

Fresh reuse capture

Retain unchanged private reuse execution_input (approval.confirmed: false):

python helpers/blob_recheck.py --capture <reuse-input.json> --receipt-dir <absolute-private-directory>

Capture checks source/configuration/Storage/context and retains cutoff/first cycle. Exit 0, blob-readiness-capture: checkpoint retained, not readiness. No writes/new ownership/polling. Creation artifacts/ambiguous PUTs cannot use capture.

Checkpoints: 3.0 creation; 3.1 reuse; 3.2 paired binding_digest. Legacy 1.0/1.1/1.2 and 2.0/2.1/2.2 remain readable, never upgraded; acknowledgements stay 1.0. V3 retains full digest/ETag for cleanup, independent core_digest, allowlisted field_digests, and datasource credential_digest. Datasource binding_proof is observed ResourceId or unverified; hidden initial snapshots cannot grant readiness. No secrets/skill text; unknown core stays protected. Reuse retains excluded_cycle and no owned resources.

Recheck only

python helpers/blob_recheck.py --input <original-input.json> --receipt <operation_id.blob-readiness.json> --compact

Validate private UTF-8 records: no links/duplicates; bind integrity/approval/owner/ plan/ownership before reads. Public-cloud CLI tenant/subscription/principal must match; no login/switching/grants. Tokens are ephemeral, not access-proof/locks.

Cloud reads:

  • Source GET knowledgesources('{name}') and /status, original API version.
  • Four exact returned datasources, indexers, skillsets, indexes: GET {collection}('{name}')?api-version={original-version}. Concealed datasource credentials permit one extra GET per verification pass, adding only &includeConnectionString=true: sanitized selected keyless ResourceId plus matching current plain/sanitized ETag/core is proof. Null/masked keys/SAS cannot prove binding. No redirects/retries; 60-second/8-MiB read bounds.
  • Existing bounded two-pass Storage inventories plus ADLS metadata HEADs, before and after monitoring; both must match the original inventory digest.

Configuration brackets inventory/status: nonempty ETags, exact source/generated identities and indexer datasource/skillset/index relationships. For timing read indexer observations: equivalent formats are INFO, unconstrained timing WARNING, explicit violations block. V3 core-equal ETag-only changes are INFO, not readiness failures. Datasource type/container/prefix/system identity must match. Concealment requires independent current child Storage binding proof. Root identity/old receipts cannot prove mutable child credentials. Missing proof is datasource-binding-unverified, not misconfiguration. Visible conflicts/core drift block, even during ingestion. CU/source-vector endpoint/auth checks still apply. Masked/nonempty/malformed keys or changed identities block; no model calls or query claims.

Reuse _blob_source_read.monitor with retained cutoff and bounded watch limits. Reuse excludes its retained first completion (normalized times). Require nonempty zero-failure/zero-skip completion, valid times/counters and no errors/active cycle/future completion. Stale/denied/malformed/missing evidence blocks. Active work at a watch limit is paused/in-progress, not failed ingestion. Legacy ingestion-timeout uses readiness.watch to distinguish pauses. Primary knowledge-source status drives progress interpretation. Missing endTime waits; missing counts stay unknown. Current runs supersede older last runs; replace counters without cross-run sums or file arithmetic. Supplemental indexer lastResult reports attempted/failed items within the shared status-request/time budget, never primary readiness or unique file counts. No run/reset, queries/models, PUT/PATCH/DELETE or data/RBAC changes.

Bounded continuation

python helpers/blob_recheck.py --input <original-input.json> --receipt <checkpoint.json> --watch-seconds 1800 --watch-max-requests 300 --watch-interval 5 --compact

Three overrides together, only with --input: maxima 3600 seconds, 1000 status requests, 60-second initial interval. Otherwise use original limits. One bounded window per invocation; no daemon/renewed approval. Unchanged/transient reads back off exponentially to 60 seconds; progress resets. Retry-After beyond the window/cap pauses. Socket timeout ≤30 seconds; windows aren't hard DNS deadlines. Ctrl+C during polling pauses; library cancelled() stops before the next request. Resume the same receipt after a pause; never recreate/run/reset/delete. watch.latest: actual synchronization start/state/counters, not independently ready documents. No ETA/percentage/assumed file or pending count. Item errors remain failures; unchanged observations don't prove a stall.

Output

Exit 0: source readiness only. Historical provenance: original_run; fresh IDs: read_only_evidence.request_ids. writes_performed: [], no follow-up ownership. Retrieval/KB remain unverified. original_run.historical_created discloses retained creation separately. creation_acknowledgement v1: state/type/name/http_status/request_id, observed before receipt IO. <operation_id>.blob-result.json retains full original result/IDs/integrity for historical reporting, not authority. original_run.original_failure.watch retains v1 pause context; missing legacy results stay unavailable. Persistence failure is terminal. generated_diagnostics covers all four children with sanitized field/severity/code; indexer_diagnostics retains its prior timing/binding contract. datasource_binding_observations: v1 records (schema_version, status verified/ unverified, request_id, plain, sanitized). Read fields: etag_digest, definition_digest, core_digest; unavailable sanitized read is null. No checkpoint replacement/cleanup proof.

--compact: closed blob-operation-summary 1.1 adds versioned readiness.progress; without progress retain 1.0. Fields: status/outcome, current writes/acknowledgement, historical creations, generated identities, readiness/watch, first retry, current/historical failure code/status/ID (watch pauses aren't failures), diagnostics, unique request-ID count, private evidence/checkpoint leaf files and next decision. Full native JSON is retained privately first; without the flag output stays compatible. No hashes/raw skills in summaries. Retention failure emits full native output and nonzero exit, preserving the original error.

Exit 2: blocked, not reversed creation. Preserve resources/evidence/first failure.

--no-progress disables stderr; libraries default silent. Local integrity isn't service/lock/initiator proof. No reconstructed cutoffs or replacement zero-work vector creation proof. Return to KB/retrieval/agent owners. Cleanup needs separate approval and strict full-definition/ETag ownership; semantic exceptions never authorize deletion.

Authorities: failure/conflict/uncertainty only. Indexer status.