Foundry Iq

MicrosoftOfficialCreated by Microsoft

Skill
Workflow
Data and AI
MIT
Medium risk
Has scripts

From the author

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.

— the description by Microsoft

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

Foundry IQ provides a structured framework for creating, diagnosing, and connecting knowledge bases to AI agents. It supports making local or Blob-based documents searchable, managing multi-source knowledge configurations, and retrieving information with citations. The skill emphasizes strict adherence to defined procedures for service discovery, resource creation, and troubleshooting to ensure reliable agent-knowledge integration.

What it does

Creates searchable knowledge bases from local or Azure Blob storage documents.
Connects existing knowledge bases to AI agents.
Diagnoses issues with knowledge base configurations or unsupported connectors.
Retrieves information from knowledge bases with source citations.
Manages the lifecycle of search services and knowledge sources.
Routes requests based on specific operational needs rather than source counts.

When to use it

When you need to make local or cloud-stored documents searchable by an AI agent.
When connecting an existing knowledge base to a new or current agent.
When troubleshooting or diagnosing a knowledge base configuration.
When you need to create a new search service or knowledge source.
When performing retrieval tasks that require citations from a knowledge base.

Advantages

Provides a clear, procedural approach to knowledge base management.
Ensures consistent handling of multi-source knowledge configurations.
Includes built-in diagnostic paths for troubleshooting.
Supports secure, citation-backed retrieval for agent responses.
Prevents unauthorized or unsupported configuration changes.

Usage examples

“Create a new knowledge base from my Azure Blob container.”: The agent will initiate the knowledge base creation procedure, validating the source and setting up the necessary search indexing.
“Connect my existing knowledge base to the current agent.”: The agent will follow the connection workflow to link the specified knowledge base to the agent's retrieval pipeline.
“Why is my multi-source knowledge base failing to index?”: The agent will trigger the diagnostic procedure to identify configuration drifts or unsupported connector issues.
“Retrieve information about project timelines from the knowledge base.”: The agent will query the knowledge base and provide an answer with citations from the source documents.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/microsoft/foundry-iq/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip foundry-iq.zip -d .agents/skills

Documentation by Microsoft

Foundry IQ

Read one procedure before questions/actions; invocation is not a read. Read the owning procedure before blocked/unsupported responses too. Before mutation plans/approval, successfully read its required pre-action references, selected branches only. Do not reload successful reads. On failure, try only permitted bounded exact-path reads with any supported reader; never bypass restrictions or search broadly. If still unavailable: blocked: reference-unavailable, name missing references and inability to plan. Never invent requirements/plans. Failures first.

Cleanup and receipt-backed execution go directly to their lifecycle/producer owner; do not reopen Search intake, provisioning or hardening.

Before expensive service discovery, reuse supplied resource/intent; otherwise ask one early USE EXISTING / FIND CANDIDATES / CREATE NEW choice. Only FIND enumerates; supplied identity uses exact/minimum scoped resolution. CREATE checks its proposed name, not existing-service inventories. Preserve these answers across source/KB/CU/model handoffs; selection is not write approval.

Searchable docs: KB + validated retrieval. Confirm intent once; KS-only must be explicit. Child success is not KB completion. Unclear/compound/mode/completion: read.

Route by requested operation, not existing KB source count. Connecting an existing KB with two or more sources, without changing the KB, uses Connect. Read-only operations use their owning procedure and actual helper constraints; this does not add retrieval modes or supported source kinds. Explicit unsupported provisioning or multi-source KB creation/reconfiguration uses Diagnose and stops before discovery, even when connecting an agent is also requested. This routing takes precedence over classic-search terms (e.g. hybrid search) in compound requests. Never end in a plain refusal: Diagnose gives the boundary, docs pointer and agent handoff; writes still need an approved concrete plan. Do not silently execute only the supported part of a compound request. If existing-KB connection versus KB creation/reconfiguration is unclear, read intent-routing and clarify that scope before Azure discovery. Never infer KB mutation.

Outcome	Read
Cleanup	Plan cleanup
Failure/drift	Diagnose
Unsupported connector provisioning / multi-source KB creation or reconfiguration	Diagnose
Connect	Connect
Read KB	Query
Search only	Search
File KS only	File
Blob/ADLS KS only	Blob
Searchable/KB	KB

Generic agents: microsoft-foundry. Reads: no approval; approve unchanged plans before writes. Hide hashes. No Search/Storage keys, scope widening, guessed identity/boundaries, drift repair or joint cleanup/creation approval. Acceptance != success.

Files

ATTRIBUTION.mdAdded by AutoGineer977 B
LICENSELicence file added by AutoGineer1.1 KiB
SKILL.md3.7 KiB
agents/connect-hosted-toolbox.md7.8 KiB
agents/connect-hosted.md7.8 KiB
agents/connect-prompt-sdk-fallback.md7.8 KiB
agents/connect.md7.8 KiB
agents/create-missing-agent.md7.8 KiB
helpers/_blob_observation.pyScript4.4 KiB
helpers/_blob_source_read.pyScript39.5 KiB
helpers/_bootstrap_io.pyScript30.1 KiB
helpers/_cleanup_dependencies.pyScript27.5 KiB
helpers/_cleanup_receipts.pyScript13.5 KiB
helpers/_common.pyScript36.2 KiB
helpers/_document_adapters.pyScript15.0 KiB
helpers/_document_limits.pyScript3.9 KiB
helpers/_embedding.pyScript7.0 KiB
helpers/_file_inventory.pyScript24.4 KiB
helpers/_file_source_config.pyScript15.6 KiB
helpers/_indexer_observation.pyScript7.1 KiB
helpers/_initial_prompt.pyScript12.5 KiB
helpers/_private_json.pyScript2.4 KiB
helpers/_progress.pyScript12.1 KiB
helpers/_prompt_read.pyScript17.7 KiB
helpers/_search_read.pyScript6.0 KiB
helpers/_source_readback.pyScript1.6 KiB
helpers/blob-contracts.md7.8 KiB
helpers/blob-cu-contracts.md7.8 KiB
helpers/blob_inventory.pyScript17.1 KiB
helpers/blob_recheck.pyScript64.6 KiB
helpers/blob_source.pyScript35.8 KiB
helpers/bootstrap-contracts.md7.8 KiB
helpers/bootstrap_azure.pyScript41.5 KiB
helpers/cleanup_plan.pyScript41.4 KiB
helpers/contracts.md7.8 KiB
helpers/cu_ingestion_auth.pyScript12.5 KiB
helpers/document-assessment.md7.8 KiB
helpers/document_assess.pyScript14.0 KiB
helpers/file-upload-recovery.md7.8 KiB
helpers/file_cu_canary.pyScript32.7 KiB
helpers/file_cu_mi.pyScript9.6 KiB
helpers/file_ingest.pyScript15.4 KiB
helpers/file_source.pyScript31.7 KiB
helpers/file_upload.pyScript53.7 KiB
helpers/hosted_connect.pyScript30.8 KiB
helpers/kb-contracts.md7.8 KiB
helpers/knowledge_base_retrieve.pyScript7.7 KiB
helpers/model-discovery-contracts.md7.8 KiB
helpers/model-discovery-scopes.md7.8 KiB
helpers/model_discovery.pyScript23.7 KiB
helpers/private-artifacts.md7.2 KiB
helpers/private_artifacts.pyScript4.8 KiB
helpers/prompt_cleanup.pyScript22.0 KiB
helpers/prompt_connect.pyScript62.3 KiB
helpers/requirements-assessment.txt103 B
helpers/retrieval-verification-contracts.md7.8 KiB
helpers/retrieval_verify.pyScript19.6 KiB
helpers/search-intake-contracts.md7.6 KiB
helpers/search_intake.pyScript18.4 KiB
helpers/search_reconcile.pyScript59.7 KiB
helpers/source_vector.pyScript33.2 KiB
helpers/throttle-recovery.md5.0 KiB
helpers/vector-contracts.md7.8 KiB
knowledge-bases/create.md8.1 KiB
knowledge-bases/retrieve.md7.8 KiB
knowledge-sources/create-azure-blob.md7.8 KiB
knowledge-sources/create-file.md7.8 KiB
lifecycle/cleanup.md7.8 KiB
references/abstention.md1.0 KiB
references/blob-binding-evidence.md7.8 KiB
references/blob-indexer-observation.md7.7 KiB
references/blob-ingestion-progress.md6.5 KiB
references/blob-readiness-recheck.md7.8 KiB
references/blob-vector-readiness.md2.6 KiB
references/bootstrap-azure.md7.8 KiB
references/content-fit.md7.8 KiB
references/cu-ingestion.md7.8 KiB
references/file-cu-canary.md7.8 KiB
references/intent-routing.md7.7 KiB
references/owner.md2.1 KiB
references/platform-contracts.md7.7 KiB
references/platform-interfaces.md7.8 KiB
references/resource-intake.md7.8 KiB
references/search-substrate.md7.8 KiB
references/standard-cu.md7.8 KiB
search-services/create.md7.8 KiB
troubleshooting/diagnose.md8.7 KiB

helpers/retrieval-verification-contracts.md

Retained retrieval verification

Owner: opt-in offline audit. Default audit_status: not-requested: no paired questions, bundle or helper run. Skipping the audit never means verified retrieval, nor blocks creation/ordinary queries. No hand-authored Q&A required. Reuse user questions or propose title/content and unrelated questions. Authorized content samples support drafts with source locators; structural assessment alone does not. Never invent ground truth from audited answers. No samples/questions: defer audit. Assistant assembles approved captures, not users. Sample, query and retention consent differ; retrieval can invoke models/cost/data. Expected answers are not inputs; missing opted-in evidence remains unverified.

Offline CLI

python "<skill-dir>/helpers/retrieval_verify.py" --input "<private-bundle>/audit.json"

No network/auth/models/writes/uploads/retries/cleanup or citation-URL following. Use retained responses, not custom parsers/dumps or new queries. Never change effort/output/models/permissions/limits.

Exit 0: review-required, not acceptance. Exit 2: blocked. overall_pass is always false. structural_status: passed/unverified/failed; failure dominates. Review semantics separately. REST isn't agent proof; offline timestamps don't prove live provenance.

Closed input

One source, two exact requests; paths resolve beneath the input's directory. Reject absolute/traversing/link paths; URI/network paths block before filesystem parsing on every platform. UTF-8 JSON: input/snapshot 1 MiB each, response 5 MiB each, nesting 48, arrays at most 1000 entries. Duplicate keys, malformed/truncated JSON, missing payloads and non-200 HTTP (including 206) cannot pass.

{
  "schema_version":"1.0","endpoint":"https://svc.search.windows.net",
  "api_version":"2026-08-01-preview","knowledge_base_name":"manuals-kb",
  "source_name":"manuals","source_kind":"file",
  "expected_source":{"path":"manuals/install.md","file_id":"file-123"},
  "before":"before.json","after":"after.json",
  "customer": {
    "query":"What are the installation steps?",
    "approved":true,"basis":"content",
    "request":{"intents":[{"type":"semantic","search":"What are the installation steps?"}]},
    "response":"customer.json","captured_at":"2026-09-14T12:01:00Z"
  },
  "unrelated": {
    "query":"What is the population of Neptune?",
    "approved":true,"basis":"unrelated",
    "request":{"intents":[{"type":"semantic","search":"What is the population of Neptune?"}]},
    "response":"unrelated.json","captured_at":"2026-09-14T12:02:00Z"
  }
}

Never refresh timestamps. Questions differ; basis: title/content versus unrelated records judgment, not semantic classification. File path: exact fileName. For Blob/ADLS use source_kind: azureBlob, path as exact native blobUrl, and etag instead of file_id.

Local snapshot: captured_at, endpoint, api_version, knowledge_base, knowledge_source, original. Definitions are complete retained GET bodies with @odata.etag; KB knowledgeSources must be exactly [{"name":"<source_name>"}]. original: uniquely selected successful File-list record across all retained pages, with fileName, fileId, lastUpdatedAt, errorMessage: null. For Blob, use { "url":"<original blob URL>", "etag":"<original HEAD ETag>" }, not an indexed chunk. Bind to authorized inventory; no file listing/Blob HEAD/content inspection or verification of the operator's unique selection.

Definitions/originals must match exactly, including ETags/models. Order: before ≤ both responses ≤ after ≤ current time, with before at most one hour old. Missing/unsupported/old evidence blocks; retain failure, request evidence, no ingestion replay. Equality isn't ingestion-version proof or atomic Storage.

Responses: knowledge_base_retrieve.py envelopes (http_status, response_body). Optional producer metadata: status must be response-received/blocked; operation must be retrieve. Other values block. blocked needs a first_blocker object and always fails; no blocker on success. Keep exact errors/request IDs private. Output: codes/counts/pointers/hashes and first error/warning locations, never messages/documents/paths/questions/secrets/IDs. Caps: 40 findings/20 diagnostics; total counts and failure severity stay accurate.

Wire and identity limits

GA 2026-04-01: Blob minimal/extractive; GET has no mode/effort fields. Preview 2026-08-01-preview: File/Blob/ADLS, minimal/low/medium, extractive/synthesis; snapshot selects mode/effort/model. Minimal uses intents above; low/medium uses {"messages":[{"role":"user","content":[{"type":"text","text":"<query>"}]}]}. Overrides block. Synthesis/low/medium need KB models, not invocation proof.

REST response[].content[].text is JSON-encoded extract rows with ref_id in extractive mode; synthesis uses native [ref_id:<id>] markers. All text blocks are inspected. Used IDs must resolve to unique references[].id, then integer activitySource to the selected source activity. activity[].error fails; activity[].warning leaves completeness unverified, including token/score drops. Absent/null diagnostics are empty. Warnings must be strings (empty allowed); errors must be objects, even {} fails. Wrong types block before truthiness.

Native File references expose docName, not File ID. Blob references expose blobUrl, not ETag. Thus path-only native citations correctly return original-identity-not-in-native-citation, not a pass. docKey, citationUrl, reference IDs and chunk sourceData.id are never original identity/version proof. sourceData is schema-dependent and can be null.

expected_source.identity_field is only an operator-declared value comparison. declared_identity_value_matches counts equality, not original-source proof. original_identity stays unverified; the native File ID/ETag gap remains even with matching arbitrary data. Missing values are unverified; mismatches fail. No attestation upgrades a mapping. No stronger binding contract is implemented. Don't invent fields, relabel chunks, fill missing data or retrofit indexes.

Unrelated extractive needs serialized [] with no references, not missing blocks; this proves only structural absence. Synthesis without citations/references does not prove semantic abstention: review all claims/exact-output requirements. Unused references aren't answer citations. No keyword/phrase matching.

Filename search limitation — explain before ingestion

File/Blob expose no supported switch for searchable generated filenames. File user metadata isn't searchable/filterable; listing search/prefix isn't KB full-text retrieval. Citation/filter fields need actual searchable: true. Filename words can still match document content. No generated-index patches, reuploads or extra metadata/schema/models. Customer-managed searchable fields need separate design/cost/data approval, not an option silently added to these executors.

Authorities: failure/conflict/uncertainty only. Checked 2026-09-14:

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Needs Python

The author says: Azure

Security

What our scanner noticed· Medium risk

References to credential files
Instructions that use the network
Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

Microsoft

Publisher

Microsoft (Official)

Source

microsoft/skills · .github/plugins/foundry-iq-skills/skills/foundry-iq · commit 0696600

Licence

MIT License (redistribution allowed)

LICENSE (added by AutoGineer from .github/plugins/foundry-iq-skills/LICENSE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by Microsoft or Microsoft.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download foundry-iq.zip