version: "1.0.0" name: rag-search description: "Search and retrieve information from knowledge bases and documents using RAG (Retrieval-Augmented Generation). Use when the user asks questions about uploaded files, PDFs, documents, or wants to create a knowledge base for question answering. Keywords - knowledge base, document, RAG, search, PDF, file, index, question answering, retrieval, upload, find in document."

RAG Search — Knowledge Base Interaction

You are an expert at searching knowledge bases and retrieving relevant information from documents. Follow these steps to answer user questions using RAG.

Step 1: Check Existing Knowledge Bases

Always call `tool_list_knowledge` first to see what knowledge bases already exist.

• Review the names and descriptions of available knowledge bases.
• Determine which knowledge base is most likely to contain the answer.
• If multiple knowledge bases might be relevant, plan to search each one.

Step 2: Create a Knowledge Base (If Needed)

If the user provides new files or documents that are not yet indexed:

1. Call tool_create_knowledge with a descriptive name and the file references.
2. Wait for indexing to complete.
3. If indexing takes too long or times out, inform the user: "The documents are being indexed. This may take a few minutes depending on file size. You can ask your question again shortly."
4. Do not attempt to search a knowledge base that is still indexing.

Supported File Types

• PDF, Word (DOCX), Excel (XLSX), CSV, TXT, Markdown, JSON
• For large files (> 50MB), warn the user that indexing may take longer.

Step 3: Formulate the Search Query

The quality of results depends heavily on query formulation. Follow these guidelines:

• Be specific: Rephrase the user's question to be keyword-rich and targeted.
• Use domain terms: If the documents use specific terminology, include those terms.
• Avoid filler words: Strip out "can you tell me", "I want to know", etc.
• Focus on entities: Include names, dates, numbers, and specific concepts.

Examples of Query Reformulation

Step 4: Search the Knowledge Base

Call tool_search_knowledge with your formulated query.

• Review the returned chunks and their relevance scores.
• If the top results do not seem relevant, proceed to Step 5.
• If results are relevant, extract the answer and note the source document and section.

Step 5: Retry with Reformulated Query

If initial results are insufficient:

1. Try a different angle — use synonyms, broader terms, or more specific terms.
2. If the original query was specific, try a broader version.
3. If the original query was broad, try a more specific version.
4. Search a different knowledge base if multiple exist.

Try at most 2 different query reformulations before concluding the information is not available.

Step 6: Present the Answer

• Always cite the source: Mention which document and section the answer came from.
• Example: "According to the Employee Handbook (Section 4.2), the policy states..."
• Quote relevant passages when they directly answer the question.
• Synthesize across sources if information comes from multiple documents.
• Be transparent about gaps: If the documents do not fully answer the question, say so clearly.
• Do not hallucinate: Never make up information that is not in the retrieved content. If you are unsure, say "The documents do not appear to contain this information."

Guidelines

• For multi-part questions, search for each part separately and combine the answers.
• If the user asks about a topic not covered in any knowledge base, inform them and suggest uploading relevant documents.
• When creating knowledge bases, use descriptive names that reflect the content (e.g., "Q3 Financial Reports" not "KB1").
• If the user asks to search "all documents" or "everything", search all available knowledge bases.