Response Generation controls how the Enterprise Search agent formulates and delivers its final answer. This screen mirrors the AI Agent editor's Response Generation settings exactly.

Overview

Response Generation sets up how your agent formulates and delivers responses, including communication style and output format. From here you choose the output format, the model that writes the answer, the language behavior, and the tone, style, and format of the prose, plus options for showing the agent's reasoning, using graph-based retrieval, and watermarking generated images. The section saves automatically.

Output format

Use Output Format to choose how the agent returns its answer:

• Text — a normal written reply (the default).

• Object — structured data. Choosing Object opens a schema builder where you define the fields the agent must return, so the output is machine-readable and consistent.

Pick Text for conversational answers and Object when another system consumes the agent's output and needs a fixed shape. The Answer Generation Model picker beneath it sets which model writes the response.

In the schema builder, each field has a type — String, Number, Integer, Boolean, Object, or Array — and an Array field's items can themselves be edited recursively, including nested objects, so you can describe multi-level structured output. You can generate a starting schema from your instructions, or paste one in via the JSON import option. Clear All needs a second confirming click before it wipes the schema, since that can't be undone.

Response language

Use Language Type to choose how the agent picks a language:

• User Locale — responses match the user's locale settings (the default).

• Same as User Query — responses use the same language as the query.

• Fixed Response Language — the chosen language is used for every response. Selecting this reveals a Select Language picker.

Use User Locale or Same as User Query for multilingual audiences; use Fixed when every reply must be in one specific language regardless of who's asking.

Tone, style, and format

Match these to your audience — for example Friendly + Conversational + Paragraph for end-user support, or Neutral + Technical + Bullet points for an internal IT agent.

Chain of Thought, Graph-Based Retrieval, and Watermarking

• Chain of Thought — shows the agent's step-by-step reasoning process (on by default). Turn it off to hide the reasoning and return only the answer.

• Graph-Based Retrieval — uses graph traversal and relationship analysis to find contextually relevant answers (off by default). Useful when answers depend on relationships between facts, not just matching text.

• Watermarking — adds subtle identifiers to agent-generated images for verification (off by default). When on, you choose a Watermarking Technique (Frequency Based) and optional Watermarking Text.