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/search-intake-contracts.md

Reuse-first Search intake

Use before service enumeration/provisioning, not for existing KB retrieval or agent connection (their owners keep their actual read-only checks). No supplied service/intent: ask USE EXISTING / FIND CANDIDATES / CREATE NEW once before expensive discovery. Supplied IDs/names/endpoints imply USE EXISTING unless creation was explicit. Carry resolved intent and selections through File/Blob/KB/CU/model handoffs; never reopen informed answers. A supplied candidate completes FIND as USE EXISTING, without another inventory.

This checker is Search-only. For CU/model dependencies use purpose-specific discovery: CU uses AIServices; embedding/chat use OpenAI or AIServices (Foundry). Carry separate selections; USE/FIND/CREATE is workflow state, not a model-helper input field. Call model discovery for supplied selectors or FIND; CREATE goes to its native owner, never an empty selector. No generic AI-host bootstrap/quota/RBAC automation. CU, vectors and KB own API, auth/MI, RBAC, network and actual processing checks. File Standard defaults to system MI (v1.2), implemented but not live-verified. Explicit legacy ARM-key v1.1 and ENV-key v1.0 remain; no fallback. File auth checks exact CU/role/identity/network. Blob CU uses Search system MI; neither mode selects embeddings or KB chat.

Native helper

python helpers/search_intake.py --select intake.json

Closed UTF-8 input; only schema_version is required:

{"schema_version":"1.0","intent":null,"service":null,"subscription_id":null,
"resource_group":null,"source_region":null,"page":0,"operation":null}

intent: null, use-existing, find-candidates, create-new. service: exact Search ARM ID, service name or clean HTTPS Search root. Names: 2-60 lowercase ASCII letters/digits/dashes; first, second and last characters alphanumeric, no consecutive dashes. Thus ab-c is valid, a-b is not. Full ARM IDs are matched case-insensitively, then their service names validated. GUID subscription; group names follow ARM rules. IDs supply scope; conflicts block. source_region: observed source region, only a latency/data-movement preference, never inferred from names or used as compatibility proof. No default location or subscription scanning. Explicit subscription wins; otherwise az account show supplies the stated default. No context switch.

| Intent | Reads/result | |---|---| | Missing | intent-required, three choices; zero CLI/auth calls | | USE EXISTING without locator | service-required; ask for locator, no inventory | | Supplied ID or name+group | Exact ARM GET only, plus selected account context | | Name/endpoint without group | One az resource list --name <name> --resource-type Microsoft.Search/searchServices --subscription <sub>, then exact GET if unique | | FIND CANDIDATES | Known group: az search service list --resource-group <group> --subscription <sub>; otherwise type-filtered az resource list in selected subscription | | CREATE NEW | Require name/group, exact collision GET; creation-plan-required only on exact ResourceNotFound. No service inventory or implicit reuse |

Lists use --query "[].{id:id,name:name,location:location}"; native CLI consumes pages in that scope. Keep the complete bounded inventory internally; return at most five rows, total/remaining/unassessed counts and next_page. page is zero-based, below 40; each page refreshes the scoped inventory. All candidates remain unassessed, not recommended or unique from a top-five view. Same observed source region first; retain cross-region rows. Explicit choice wins. Use the row's selection_input for exact readback, not another list. No model calls, cross-subscription fanout, arbitrary URL fetches or writes. More than 200 rows, duplicates, scope escape, malformed/denied reads block without partial choices or absence claims. Shared native reader: 20s/command, 120s total, one MiB post-capture limit, not a hard capture-memory bound.

Operation-specific reuse

Initially operation: null selects ARM metadata, not compatibility/readiness. Once the actual mode is chosen, rerun the same selected ID with one operation:

{"kind":"file-source","api_version":"2026-08-01-preview","extraction_mode":"minimal","vectorization":"none"}

Closed variants:

| kind | Other required fields | |---|---| | service | api_version | | file-source / blob-source | api_version, extraction_mode: minimal/standard, vectorization: none/azureOpenAI | | knowledge-base | api_version, reasoning_effort: minimal/low/medium, output_mode: extractiveData/answerSynthesis |

APIs: 2026-04-01 or 2026-08-01-preview; File, Standard extraction and nonminimal/synthesis KB require preview. Downstream source/KB validators retain their additional API, source-kind, tier/size, regional capability and data-boundary constraints; these variants are not a new capability matrix.

Existing keys may remain enabled when ARM shows aadOrApiKey and Entra read works. No Search/Storage keys, key retrieval, secret input, identity/firewall/network/tag changes, or hardening to fit creation defaults. Require running/succeeded, exact endpoint/ID and recognized observed SKU/region. Public firewall restrictions are preserved; actual caller read must succeed. Private/perimeter access is unsupported here: report search-network-unsupported, not a global compliance failure or instruction to open networking.

Search system MI is required for Blob (Storage/CU), source embeddings and KB chat/nonminimal reasoning, not lexical File or minimal/extractive KB alone. file-source does not encode CU auth: File's owner verifies MI or explicit legacy keys before its plan; this Search read is not File CU readiness. Verify required system principal/tenant; never accept a missing MI by request. UAMI selection is unsupported; an unused UAMI is not caller authentication.

For a selected operation, scoped az rest --method get --resource https://search.azure.com reads <endpoint>/servicestats?api-version=<selected-api> using CLI Entra auth, then exact ARM reread rejects access/identity drift. No billable write/model probes. Statistics GET proves only that metadata read, not source/KB write RBAC, feature-region support, Search-to-Storage/CU/model access, ingestion or retrieval. Verify these separate edges with the owning procedure before its plan/approval; no success-shaped fallback or permission relaxation on denial.

Handoff

selected: observed ID/endpoint/SKU/tier/location and qualified read evidence. After unique name/endpoint resolution, scope.resource_group and selection_input.resource_group carry the observed group in the selected subscription. Unresolved/ambiguous names and FIND retain their original scope; candidate rows carry each group's exact handoff without narrowing that inventory. operation_required means mode unresolved. Reuse selection_input; no receipt, approval envelope, ownership claim or creation-profile matching. Fresh selection/readback is not future consent. Creation delegates to unchanged bootstrap choices: secure new-service defaults, required region/quota/cost checks, exact collision reread and separate approval. Never run strict legacy bootstrap reuse merely to validate an existing operation.

Authorities

Authorities: failure/conflict/uncertainty only. Service naming rules, Entra and keys together, statistics GET.