Skip to main content

Additional features: Reflection

When to use​

Enable Reflection for checks an LLM can verify without fetching new context—then auto-fix the draft if needed. Ideal for:

  • Format & structure: validates JSON/YAML/Markdown sections, headings present, table columns, code fences.
  • Counts & ranges: list has ≥ N items, word/character limits, numeric fields within bounds.
  • Language/style: reading level, tone, banned phrases, inclusive language, tense/voice consistency.
  • Required fields: keys present, non-empty values, citation stubs included, filename/ID included.

Not for fact-checking against external sources. If you need retrieval or tool data, use KB/tools first, then reflect on form and criteria only.

What it does​

  • After the agent drafts a reply, it runs a self-review pass (or several) using a Reflection prompt.
  • If issues are found, the agent rewrites the reply and repeats until it passes or hits the Maximum Reflection Runs limit.
  • This happens before delivery to the user, improving correctness, completeness, tone, and formatting.

Trade-off: more reflection runs = higher latency and token usage.

Enable in the UI​

  1. Open your agent → Edit.
  2. In Additional features, click Add and toggle Reflection ON.
  3. Click Configuration to set:
    • Maximum Reflection Runs (slider).
    • Custom Reflection Prompt (textarea, with Reset to Default).
  • 0: Off (fastest).
  • 1–2: Good default for most agents (balanced).
  • 3–5: High-stakes or long-form outputs (PRDs, policy answers with citations).
  • >5: Rare; expect noticeable latency.

A strong Reflection prompt (starter)​

Use or adapt this checklist:

Review the draft and fix issues without changing meaning. Check: (1) accuracy vs. provided context/KB; (2) completeness and direct answer; (3) clarity, structure, and tone; (4) contradictions or leaps in logic; (5) required fields/format (JSON/Markdown). If unsure, propose the briefest clarifying question.

If citations are used, ensure they support each claim. Output only the improved final answer.

Variants by task​

  • Policy/KB Q&A: “Do not invent facts. If the KB doesn’t cover it, say so and suggest next steps.”
  • Structured JSON: “Validate against this schema; if invalid, correct and re-emit valid JSON only.”
  • Code: “Ensure runnable, add minimal comments; avoid unsafe operations; include example usage.”
  • Tickets/Forms: “Verify all required fields present and consistent; surface missing fields as a short checklist.”

Best practices​

  • Keep the Reflection prompt short and specific to your domain.
  • If you use Artifacts or tool citations, include them in the checklist (“preserve artifact structure”, “verify tool-sourced data is labeled”).
  • For strict formats, ask reflection to validate and re-emit in the exact format.
  • If responses become verbose, include “avoid repetition; remove filler language.”

Troubleshooting​

  • Too slow → lower Maximum Reflection Runs or tighten the prompt.
  • Still incorrect → emphasize “do not invent” and “defer to KB/citations”; ensure the base answer actually had access to the right KB/tools.
  • Breaks JSON/format → add explicit schema/format validation to the Reflection prompt.
  • Over-editing tone → remove style-heavy checks; focus on accuracy/completeness.