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

knowledge-sources/create-file.md

Create a File knowledge source

When to use

Direct uploads; no customer Storage.

Do not use

Changing/scheduled corpora, lifecycle/DLP/permissions or overflow: Blob/ADLS. Unsupported content blocks. Never stage to Blob, split silently, alter files or upload outside the approved root.

Inputs and discovery order

Resolve prompt/session/workspace, then az account show for caller/tenant/default subscription. Explicit scope wins; never silently switch context. Missing CLI/sign-in blocks discovery, not proof of absence. Reads need no approval except content inspection: exact sample authorization. Silence is not approval. Recommend; after discovery batch compatible choices only for unresolved decisions. Reuse informed choices; separate sample-content consent from actual resource/permission/CU/source/KB approvals.

| Input | Why needed | Required? | Discovery order | Safe default | If missing or unanswered | Reconfirmation trigger | |---|---|---|---|---|---|---| | Scope/Search/access | Target | Output before upload | Exact Azure GET | Reuse/new | Choices | Scope/access | | Root/inventory | Data | Pre-upload | Selection/list | Never . | Choose root | Files | | Owner/name/purpose/cleanup | Owner | Pre-write | Caller | Azure user | Override | Owner | | Processing/models | Search | Yes | Content/choice | No extraction default | Assess/choose | Content/mode | | API/cost/network/consent | Security | Yes | GET | Preserve | Approve plan | Cost/access |

Decisions

Opt-in title/content; no Q&A required. No auto-sample/query or filename defaults.

Search operation; bootstrap. Resume after readback, never call Create KB. No routine policy reads. File-only creates no KB/synthesis model. Minimal lexical extraction needs no model; standard independently selects CU.

Honor explicit processing choices only when informed and compatible. Unknown content or mode/goal conflicts: read content-fit. MUST assess/clarify before planning: bounded authorized samples or questions about text-only versus multimodal content and answer dependence. Reuse answers; honor informed reduced outcomes. Unavailable/declined/unreadable inspection means ask, not assume text-only. No filename/MIME inference, silent default/downgrade or model calls. Explain tradeoffs; obtain concrete-plan/CU/data approvals. Vectors are independent. New unresolved search: recommend Hybrid; confirm tradeoffs/opt-out. Keyword-only: no embedding discovery. Lexical skips vector references. If CU/models: discovery.

Reject absolute/traversing/unreadable paths or escaping links/reparse points. Freeze path/MIME/bytes/mtime/SHA-256 in byte order; recheck pre-upload.

Prove formats, at most 200 files, per-file maximum: 50 MB for Free/Basic, 100 MB otherwise. Unknown/overflow: Blob.

  • minimal: supported non-image content with contentExtractionMode: minimal.
  • standard: contentExtractionMode: standard, 2026-08-01-preview, CU and approved cost/security/data movement; source embeddings are optional. Detected images with minimal are invalid; never downgrade.

GET exact source; only FIND lists candidates. Read all file pages each run. Reuse exact definition/inventory/index/markers. Duplicates/inaccessible/drift block; no suffixing/upload repair.

Proposed plan

Read helper contracts before planning. For standard planning, read CU choices; minimal skips it. CU auth: verified Search MI or explicit ARM/ENV; never expose secrets.

python helpers/file_source.py --plan request.json --execution-output <new-private-plan.json>

UTF-8:

{"schema_version":"1.0","endpoint":"https://svc.search.windows.net","name":"manuals","owner":"team","local_root":"C:\\docs","paths":["guide.txt"],"service_tier":"basic","extraction_mode":"minimal","vectorization":"none","rbac":{"assignments":[{"id":"<role>","principalId":"<principal>"}]},"network":{"posture":"<posture>","evidence":"<readback>"}}

Tier: free/basic/dedicated/serverless. Require RBAC and network.posture/network.evidence. paths: 1–200 unique relative POSIX paths; no globs; MIME hint only. Types/API. Vectors: vectorization: azureOpenAI; choices.

Exit 0: planned, not ready. Private output: retain unapproved input privately; show reference/counts, not file hashes. Review actions/mode/access/cost/retention/cleanup. Verify known/security requirements before consent; unapproved input blocks. Exact reuse: execution_required: false and mutation_approval_required: false: no approval or executor. Reread definition/ETag after all pages; ingestion isn't retrieval. No writes/installs/identity changes/cached consent.

Confirmation

Retain immutable plan/plan_fingerprint/cleanup_approved: false; review choices, not hashes. Requirements/inventory/endpoint/API/model/identity/network/role/definition/ETag drift voids consent. Rerun --plan, replace the saved execution_input and discard old consent; approve only new actual mutations. Never edit nested plans or recompute hashes manually.

Mutation

After approval:

python helpers/file_source.py --input <approved.json> --cleanup-receipt-dir <private-receipts> --upload-receipt-dir <separate-empty-dir>

Require validated private receipts pre-write; no inferred paths/GET ownership. reconcile-and-ingest validates source/ingestion agreement before writes. Synchronous POST: 180s; final inventory. 429: one bounded retry; create once. Timeout/409/5xx: readback only. Exit 2: blocked; 3: partial. Never invoke children alone. Create absent sources only; reuse never uploads. fileParameters.ingestionParameters binds contentExtractionMode; lexical omits embeddingModel; vectors bind observed azureOpenAI endpoint/deployment/model. Standard binds aiServices.uri; no verbalization/chatCompletionModel.

Block rather than downgrade to minimal. Reject networkAccessMode: one index, no indexer/schedule; never patch children.

List /knowledgesources('<source-name>')/files; upload only absent exact inventory entries by multipart POST; read stable file IDs. Never modify/stage files, use Search/Storage keys/SAS, or rename an ambiguous retry.

Verification

Verify definition/mode/dependencies, count/path/type/size/hash-to-file-ID, zero ingestion failures, generated index/child ownership, keyless Search access, network and unchanged inventory. Rerun discovery: stable IDs/zero writes. Vectors: staged verification; deployment/lexical results aren't vector proof.

Failure and partial completion

Report first failure/proved counts and progress. Resume; never recreate, replay uncertain files or suggest cleanup.

Cleanup

Receipts: retain pre-create. Separate cleanup plan/approval: delete run-owned source/children. Never delete or alter local files, reused resources or shared roles. operation: delete: owned digest/ETag; unclear ownership blocks.

Return contract

Return status/fingerprint/IDs/API/boundary/inventory, auth/RBAC/network/ingestion/idempotency/ownership/warnings/cleanup. Pass assessment/uncertainty/answer needs/mode through prerequisites to the KB owner; resume KB plus validated retrieval. Never return content/credentials.

References

Authorities: failure/conflict/uncertainty only.