Peer Settings

This guide maps the Peer settings UI to each field and explains what they do. It follows the tabs as seen in the app: Basic, Prompt, Search, Features, and Advanced. It also explains how Peers use Memory.

Basic

• Name: Display name for your Peer.
• Model: Base model for text reasoning and responses.
• Short Description: Summary shown in lists and overviews.
• Language: Default language your Peer uses in conversations.
• Role: High-level role context for the Peer (who it is).
• Goal: Outcome-focused objective (what it should optimize for).
• Peer ID: Read-only identifier (useful for API and integrations).

Prompt

• Initial Prompts: Up to 4 short instruction lines added early to the system context.
• Temperature: Creativity slider (0–100). Higher = more diverse output.
• Additional Prompt: Longer guiding instructions appended to system context.
• Content Type: Preferred response style preset from workspace preferences.

Search & Knowledge Base

• Knowledgebase Language: Preferred language of indexed KB content.
• Text Search Weight: 0–1 weight for keyword scoring.
• Vector Search Weight: 0–1 weight for semantic scoring.
• Search Threshold: Minimum relevance score to include results.
• Search Limit: Max number of results to retrieve.

Features

• Enable RAG: Allow retrieval from knowledge sources during answers.
• Include Datasources as Tools: Expose each connected datasource as an individual agent tool for targeted retrieval, instead of searching all sources together.
• Disable Tools: Prevent tool actions (Flows, HTTP, etc.) during reasoning.
• Disable Document: Hide/disable document upload in the chat UI.
• Canvas: Enable rich canvas for creating/editing documents and code. Configure which artifact types are allowed (documents, slides, diagrams, code, canvas).
• Media Model: Optional model for image or media understanding tasks.
• Web Search: Enable built-in web search capability for the Peer.
• Voice: Enable real-time voice conversation mode (speech-to-text → LLM → text-to-speech pipeline).

Advanced

Reasoning Level

• Reasoning: Set the model's reasoning depth. Low is fastest and cheapest; High uses extended chain-of-thought reasoning for complex problems. Requires a model that supports reasoning mode (e.g., o3, o4-mini, Claude 3.7 Sonnet).

Topic Detection & Off-Topic Handling (Beta)

• Topic Detection: Enable automatic topic analysis from recent messages.
• Continue on Off-Topic: If off-topic, still answer normally (on), or switch to special handling (off).
• Off-Topic Response Mode: Static (fixed message) or AI (dynamic explanation and redirect).
• Off-Topic Message: Message used when mode=Static.

Response Controls

• Response Length: Short (1–2 sentences), Medium (1–2 paragraphs), Long (multi-paragraph).
• Response Format: Conversational, Structured (headings/lists), or Technical.

Tool Usage & Context

• Tool Usage Intensity: Low / Medium / High profile for max tool calls.
• Include Recent Tool Logs in Context: Inject recent tool runs into system prompt.
• Tool Log Context Limit: How many recent logs to include (1–25).

Planning & History

• Planning: Enable multi-step planning and coordinated tool orchestration before executing actions.
• Messages Count: How many recent messages to include from history.

Flow-Based Peer

• Flow-Based: When enabled, the Peer delegates every conversation turn to a specific Flow instead of using the standard agent pipeline.
• Flow: Select the Flow that will handle messages. The selected Flow is triggered on every incoming message.

Use Flow-Based mode to build Peers with fully custom conversation logic, branching, multi-step processing, or integrations that the standard agent cannot handle.

Semantic Cache

• Cache Enabled: Enable semantic caching. When a user sends a message semantically similar to a previously cached query, the cached response is returned instantly without invoking the LLM — saving time and credits.
• Cache Settings: Configure the similarity threshold (how close a new query must be to a cached one to be served from cache) and the vector index used for cache lookup.

Session Management

• Session End: Enable automatic session termination based on conditions.
• Session End Conditions: Define conditions that signal the end of a conversation (e.g., user says "goodbye", task is marked complete). When a condition is matched, the session closes automatically.

Guardrails

• Guardrails: Attach content safety policies to the Peer. Each guardrail inspects either the user's message or the AI's response and can block, warn, or mask content.

Three policy types are available:

Type	What It Does
PII Detection	Identifies and optionally masks sensitive personal data (emails, phone numbers, credit cards, IBANs, national IDs, passports, etc.)
Prompt Injection Shield	Detects and blocks prompt injection attacks in user messages
Moderation	Flags content that violates hate speech, violence, or other moderation categories

Each guardrail specifies:

• Target: Apply to user message, AI response, or both.
• Action: Block (reject the message), Warn (add a warning), or Mask (replace detected content).

---

Memory

Peers can store and retrieve user-specific information to personalize conversations.

User Memory

• Enable User Memory: Toggle to allow the Peer to automatically generate and store memories from conversations. Memories are extracted asynchronously after each session.
• User Memory Prompt: Custom instructions that guide how the Peer decides what to remember about users.
• Shared Memory: When enabled, memories are shared across all users interacting with this Peer (useful for team-shared context). When disabled (default), memories are scoped per user.

Memory entries can also be manually managed through Flow steps:

• Save: Use the "Save User Memory" Flow step to persist a key/value with a description.
• Retrieve: Use "Search User Memory" to look up previously saved items.
• Delete: Use "Delete User Memory" to remove records when they are no longer needed.

Quick links:

• Save User Memory: ../flow/steps/save-user-memory
• Search User Memory: ../flow/steps/search-user-memory
• Delete User Memory: ../flow/steps/delete-user-memory

Notes

• Enabling User Memory will automatically extract and persist relevant information from conversations after each session.
• Respect workspace permissions and data policies when storing personal data.
• "Messages Count" in Advanced controls recent chat history, not long-term memory.

Common patterns

• Greet users by name after saving profile details once.
• Store preferences (language, tone) and reuse them across sessions.
• Keep task or ticket IDs to track ongoing requests.

Tips

• Keep Initial Prompts brief; move long guidance to Additional Prompt.
• Balance Text vs Vector weights based on your content quality and query types.
• Start with Tool Usage = Low and increase as you build confidence.
• Use Static Off-Topic messages for consistent brand voice; AI mode for flexibility.