| name | rag-search |
| description | Azure AI Search knowledge base |
| enabled | true |
| index_name | knowledge-base |
RAG Search Skill
Instructions
When the user asks about internal policies, documentation, procedures, or knowledge base content, use the rag_search tool to retrieve relevant content from the Azure AI Search knowledge base.
IMPORTANT: Always pass index_name: "knowledge-base" when calling rag_search. This is the dedicated search index for the generic knowledge base.
1. When to Use
Use this skill when the user's question relates to:
- Internal documentation, policies, or procedures
- Product or service information stored in the knowledge base
- Reference material, FAQs, or how-to guides
- Any question that should be grounded in authoritative internal content
Do NOT use for: real-time web data (→ web_search), calculations (→ code_interpreter).
2. Craft the Search Query
- Be specific with domain terms rather than generic phrases
- Include the key concept and context (e.g., "onboarding process new employee" rather than just "onboarding")
- If the first query returns low-relevance results, rephrase with synonyms or more specific terms
3. Interpret Search Results
Results contain: title, content, source, page, relevance score.
- Synthesize multiple results into a coherent answer rather than dumping raw excerpts
- Cite sources with document title and page number (e.g., "Per the Employee Handbook, p. 12...")
- Quote exact wording when precision matters (policies, procedures, requirements)
- Flag gaps — if the retrieved content doesn't fully answer the question, say so explicitly
- Note confidence — if results have low relevance scores, mention that the information may not be complete
4. Multi-Step Workflows
Combine RAG search with other skills for richer answers:
- Knowledge + summary:
rag_search → document_summary to condense retrieved content
- Knowledge + analysis:
rag_search → code_interpreter to analyze data referenced in docs
- Knowledge + web verification:
rag_search → web_search to check if internal info is still current
- Knowledge + email:
rag_search → email_draft to compose a response grounded in policy
5. Response Guidelines
- Present information clearly and accurately — do not paraphrase in ways that change the meaning
- Use bullet points or tables for structured information
- Always include the source document and page reference
- If multiple documents are relevant, present findings from each and note any differences
- If no relevant results are found, say so clearly and suggest the user refine their query
6. Error Handling
If the search returns an error or no results:
- Report the issue transparently to the user
- Suggest alternative search terms
- Offer to try
web_search as a fallback if appropriate