Document Research

Document Research

When researching from the knowledge base:

• Run search with clear, specific queries; use multiple queries if the topic is broad.
• Synthesize findings across snippets into a concise answer or summary.
• Indicate which documents or sections support each claim.
• When the user asks for a list (e.g. features, steps), extract and order from the docs.
• If the answer is uncertain or partial, say so and suggest where to look next.

Step-by-step instructions

1. Identify the main topic and any sub-questions (e.g. “how to configure X” and “what are the limits”).
2. Call search_documents for each distinct sub-topic with a focused query.
3. Read all returned snippets and note document names and sections.
4. Synthesize into one answer: list or narrative, with each claim tied to a document/section.
5. If the user asked for a list, preserve order from the docs or state the ordering you used.
6. If information is missing or ambiguous, say so and suggest which doc or section to check.

Examples of inputs and outputs

• Input: “How do we set up SSO and what are the limits?”
  Output: Short “Setup” and “Limits” subsections, each with bullets and document citations from search_documents results.

• Input: “List all API endpoints for billing.”
  Output: Numbered or bullet list taken from docs, with document/section references.

Common edge cases

• Broad question: Split into 2–3 queries (e.g. “SSO setup”, “SSO limits”, “SSO troubleshooting”) and combine answers.
• No results for one sub-question: Answer the parts you found; for the missing part say “I didn’t find this in the knowledge base.”
• Duplicate or overlapping snippets: Deduplicate and cite the single best source per point.
• User asks “everything about X”: Give a structured summary (overview, steps, limits, caveats) and cite docs; offer to go deeper on one part.

Tool usage for specific purposes

• search_documents: Use for every research question. Use one query per distinct sub-topic; avoid one very long query. Use it to pull lists (features, steps, endpoints) directly from the text.