| name | azure-agentic-retrieval |
| description | This skill should be used when implementing private grounded retrieval on Azure: Azure AI Search, agentic retrieval, knowledge sources, File Search, SharePoint grounding, citations, access-controlled indexes, and retrieval traces. |
Azure Agentic Retrieval
Agentic retrieval is the Microsoft-native path for giving agents private, high-signal context without stuffing entire corpora into prompts. Use Foundry IQ when the agent needs a reusable managed enterprise knowledge layer; use Azure AI Search and Foundry retrieval tools when the implementation needs direct index control, custom pipelines, or lower-level retrieval tuning.
When to Activate
Activate this skill when:
- Building RAG or grounded agents on Azure AI Search.
- Deciding between File Search, Azure AI Search, SharePoint grounding, Fabric Data Agent, or custom retrieval.
- Designing knowledge sources, indexes, chunks, citations, and retrieval traces.
- Preserving document permissions while grounding responses.
- Comparing simple vector search with hybrid, semantic, or agentic retrieval.
Do not activate this skill for adjacent work owned by other skills:
- Generic memory architecture and retrieval-shape decisions: core
memory-systems.
- Token budgeting, masking, and context partitioning independent of Azure Search: core
context-optimization.
- Identity and OBO design for retrieval:
azure-identity-for-agents.
- Reusable Foundry IQ knowledge bases and enterprise knowledge APIs:
foundry-iq-knowledge-layer.
- Structured data in Fabric semantic models or lakehouses:
fabric-data-agent.
Core Concepts
Retrieval is a context filter. The objective is not to retrieve the most text; it is to retrieve the smallest set of permission-valid evidence that can answer the question with citations.
Use the shallowest Azure retrieval path that satisfies the data shape:
| Data shape | Default |
|---|
| Uploaded documents for one agent | Foundry File Search |
| Reusable managed enterprise knowledge base | Foundry IQ |
| Enterprise search index | Azure AI Search |
| Microsoft 365 documents with user permissions | SharePoint grounding |
| Analytical data and semantic models | Fabric Data Agent |
| Mixed private sources with custom pipeline | Azure AI Search knowledge sources and custom indexing |
Practical Guidance
- Start with the user's answer contract: citation requirements, freshness, permission semantics, and latency tolerance.
- Choose the retrieval surface based on data ownership and access control before choosing chunking.
- Use hybrid retrieval when exact terms and semantic meaning both matter.
- Keep chunks evidence-sized, with stable source identifiers and citation metadata.
- Return retrieval activity in traces so failures can be debugged without replaying the whole conversation.
- Evaluate groundedness, relevance, and citation accuracy separately.
Examples
Example: internal policy assistant
Private HR policies in SharePoint: use SharePoint grounding with OBO.
Published product docs in Blob: index into Azure AI Search.
Final answer: cite both sources separately and do not merge permission domains.
Example: retrieval escalation
Start: File Search for a small uploaded corpus.
Scale: Azure AI Search when multiple agents share the corpus or need custom indexing.
Escalate: agentic retrieval when query planning over multiple knowledge sources improves answer quality.
Guidelines
- Design retrieval around evidence and permissions before model prompts.
- Prefer private grounding sources over public web grounding for regulated workloads.
- Store source IDs, timestamps, and access scope with every retrieved chunk.
- Evaluate retrieval separately from generation.
- Surface "not enough evidence" as a valid answer.
Gotchas
- Search relevance is not answer groundedness: A high-ranked chunk can still be misused. Evaluate the generated answer against retrieved evidence.
- Permission filters are part of retrieval quality: A result the user cannot access is a retrieval failure, even if semantically relevant.
- Public web grounding changes the compliance story: Prefer Azure AI Search, SharePoint, Fabric, or customer-controlled sources when data boundaries matter.
Integration
- core memory-systems - Owns the generic memory and retrieval architecture.
- core context-optimization - Owns token efficiency and selective context loading.
- foundry-iq-knowledge-layer - Owns reusable managed enterprise knowledge bases.
- azure-identity-for-agents - Owns access semantics for private sources.
- responsible-ai-guardrails - Evaluates groundedness and injection risks.
References
Skill Metadata
Created: 2026-05-24
Last Updated: 2026-05-24
Author: Agent Skills for Context Engineering Contributors
Version: azure-0.1.0
