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-indexer-observation.md

Generated Blob indexer observations

Read when Blob/ADLS generated schedule or version readback differs, including checkpoint recovery. Minimal/Standard/CU, lexical/vector use the same verifier. This is not schedule management, write resumption or a new planner.

Binding versus timing

| Observation | Decision | |---|---| | Wrong dataSourceName, targetIndexName or skillsetName | Hard binding failure; never a schedule warning | | Equivalent interval/start-time formats | INFO; continue all readiness checks | | Valid timing differs without an explicit approved constraint on that field | WARNING about possible recurring work/cost; continue verification | | Explicit approved interval/start-time differs | Block constraint acceptance, not evidence that ingestion failed | | Malformed/unknown schedule | Block; do not infer defaults | | ETag-only revision | V3 core-equal observation: INFO; legacy unexplained revision: block | | Any other indexer field changes | Hard drift failure, even alongside schedule changes |

The approved original wire's ingestionSchedule constrains a supplied interval and non-null startTime. An omitted/null start time is unconstrained. New typed planners require ingestion_schedule: null: no schedule override, not guaranteed one-shot processing or absence of recurring work. No monetary cap, no-run constraint or cost acceptance is inferred from absent fields. Those constraints have no supported planner representation; if required, keep their acceptance unresolved with the caller/owner, not silently verified. Unknown schedule properties block rather than becoming invented approval.

Missing, null and scheduled observations remain distinct; pinned defaults below are not substituted for a run's readback. If an authorized GET observes P1D, disclose daily cadence and possible recurring processing/cost. Otherwise do not invent that cadence. Neither observation is cost approval or ingestion proof. Schedules contain only interval and optional startTime; empty objects, null/missing intervals, unexpected properties and invalid types block.

Normalize bounded XSD day/time durations, 5 minutes through 24 hours inclusive: P1D, PT24H, PT1440M, PT86400S are equivalent. Calendar years/months, weeks, negatives and unsupported forms do not pass. Timestamp equivalence requires a valid explicit timezone and exact instant, preserving up to seven fractional digits (100 ns). Reject naive/unknown-offset timestamps, invalid dates, leap seconds, excess precision and offsets outside +/-14:00; no rounding.

Versioned private proof

New immutable checkpoints use 3.0 creation, 3.1 reuse, or 3.2 reuse with paired creation binding. Write/source acknowledgements remain 1.0. configuration.indexer retains the actual ETag and:

  • digest: complete returned child definition.
  • non_schedule_digest: every field except schedule and @odata.etag.
  • schedule_digest: validated canonical timing, including presence state.
  • schedule_raw_digest: actual schedule representation and presence.
  • V3 core_digest/field_digests: independent semantic configuration and safe field diagnostics.

All other fields, including unknown fields, mappings, parameters, identity and description, remain in the non-schedule hash. V3 other-child core/binding policy: readiness recheck. Source identity stays strict.

Exact snapshots pass. In V2, a changed ETag/full hash can pass only when an observed schedule change explains it, every other observed field matches and current approved timing constraints hold. V2 ETag-only/unexplained changes block. V3 core-equal ETag-only changes are INFO; all explicit timing constraints still apply. Equal full hashes with inconsistent projection/version evidence block. Always retain actual current hashes; never substitute the old ETag as readback. Comparisons do not modify receipts or lock resources and cannot prove unseen intervening revisions.

Legacy 1.0/1.1/1.2 checkpoints retain exact full-hash/ETag comparison. Legacy 1.2 still requires paired creation proof; 1.1 does not. A changed legacy snapshot lacks a schedule preimage/projection and stays blocked; neither a warning nor a schema-version edit upgrades that proof. Preserve it. GET current definitions and compare legitimately retained preimages if available. Separately planned fresh read-only reuse capture is a new observation, not recovered historical configuration, cutoff or ownership.

Lifecycle and safe continuation

Keep acknowledgement-before-GET unchanged. Checkpointed creation captures generated configuration before polling and rechecks after successful ingestion/source/Storage readback. Recheck observes configuration before and after inventory/status; recovery brackets its inventory with configuration observations before retaining a checkpoint. Both capture/recovery results still mean readiness unverified.

Use existing commands with unchanged private inputs and genuine receipts:

python helpers/blob_recheck.py --recover <approved.json> --receipt <acknowledgement.json> --receipt-dir <absolute-private-directory>
python helpers/blob_recheck.py --input <original.json> --receipt <checkpoint.json>

For a separate unowned reuse operation:

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

Follow recheck contracts for private files, paired creation proof when needed, original context/ownership/cutoff, and excluded reuse cycles. Recovery cannot erase known contradictory retained configuration.

indexer_diagnostics contains severity/code, a safe field name, static message and request ID. indexer_observations contains actual hashes and hashed ETag, not raw timing/resource content. WARNING messages also appear in warnings; INFO is not promoted to WARNING. First failures and acknowledged partial ownership remain intact. A qualified cycle observed before later drift is retained, but overall readiness returns unverified.

Warnings never bypass native cycle/error checks, Storage/ACL inventory or source bindings. Read-only recheck additionally requires nonempty, zero-skip completion and rejects future completion; this does not change the original monitor's counter semantics. A checkpoint is not the separate nonempty creation proof required for zero-work vector reuse. Snippets cannot substitute. Denied/missing reads and unknown evidence remain blockers with provenance. GET/ADLS metadata HEAD only; no schedule edit, run/reset, PUT replay, new source, cleanup, query or model calls. KB/retrieval completion and separate cleanup approval remain with their owners.

Authorities

Authorities: failure/conflict/uncertainty only. On uncertainty, consult indexer scheduling for documented duration bounds, and the Blob generated pipeline for generated resources/status. Neither proves a particular unread generated schedule or a null knowledge-source schedule default.

Supplied read-only implementation trace: AzureSearch commit 2e241af8939a0891836b78278df113dab31a5964, Source\Search\Product\SearchAgentCore\RestClient\SearchRestIndexingPipelineManager.cs: DefaultIndexingScheduleDays is 1 (line 64); extraction passes the nullable ingestion schedule (521-548) through indexer creation/payload construction (297-301, 1497-1504, 1540-1544). BuildScheduleJson (1798-1816) maps null to {"interval":"P1D","startTime":null}. This establishes that pinned implementation, not every deployment or the reported manual run. Null-to-daily stays WARNING when unconstrained, never semantic equivalence; explicit constraints still block.