Meet Wavo — your AI assistant in Wevion

Wavo is the AI assistant built into Wevion. Access at /chat or /chat/:sessionId. Streams answers word-by-word. Reads your team's data only (campaigns, metrics, rules, audiences) via tools. Takes actions with explicit approval for high-risk operations. Multi-platform read (Meta, Google, TikTok, Taboola, Snapchat); write currently Meta-only. Persona file: apps/backend/src/services/campaign-by-context/agent/prompts/wavo-persona.ts (canonical).

Who is this for

Everyone in Wevion. Mediabuyers, admins, owners, creatives, viewers — Wavo works for all roles, scoped by your RBAC.

What Wavo does

Three categories:

Answers questions about your data

• "What's my ROAS for Campaign X last 7 days?"

• "Which adset has the worst CPA on Meta?"

• "Compare Google vs Meta CPC for prospecting last month."

Wavo invokes the relevant tool (get_performance_data, compare_tracker_meta, etc.), fetches live data, and answers.

Takes actions with approval

• "Pause adset Y."

• "Create a rule that kills losers."

• "Increase budget on Campaign Z by 20%."

High-risk actions (delete, duplicate, budget changes) require explicit approval via a confirmation card. Lower-risk actions (drafts, queries) run immediately. See ai-105 actions.

Generates creative + content

• "Draft 3 ad copy variants for product X."

• "Generate an image of [description] for a Reels ad."

• "Remove the background from this image."

Wavo invokes Creative Hub tools (generate_copy, generate_image, generate_video, etc. — see ch-111).

Where Wavo lives

• Page: /chat (latest session) or /chat/:sessionId (specific session)

• Sidebar: Chat / Wavo entry (depending on UI version)

• Embedded: some pages show a Wavo sidebar for in-context questions

How conversations are organized

Three-level structure (see ai-106 projects + sessions):

Project
  ├── Session 1 (e.g. "Q1 review")
  │   ├── Message 1
  │   ├── Message 2
  │   └── ...
  ├── Session 2 (e.g. "ClientA campaigns")
  └── ...

• Project: groups sessions with shared instructions, color, extracted memory facts

• Session: single conversation, persistable + shareable + favoritable

• Message: streaming UIMessage with tool calls, follow-ups, feedback

Models + providers

Wavo supports 10 models across 5 providers (apps/backend/src/services/chat/chat-models.ts):

Default is Claude Sonnet 4.5. Switch via /settings → AI preferences or per-session in the model picker.

Faster models (Haiku, Flash, GPT-4.1 mini) are cheaper but less capable on complex multi-step reasoning. Heavier models (Opus, Sonnet, GPT-4.1, Gemini Pro) better for analysis + multi-tool chains.

Persona

Wavo is:

• Confident professional — direct answers, not hedging

• Data-driven analyst — cites the tool that produced the number

• Creative for copy / ideation — looser when drafting ad text or brainstorming

• First-person — "I'll check the data" not "the assistant will check"

Wavo will ask back when info is missing — budget, target country, product, objective. Don't try to make Wavo guess; it'll ask.

What Wavo can't do

• Speak back via voice — transcribes your voice input (Whisper) but doesn't generate audio output

• See data outside your team — strictly team-scoped (RBAC enforced)

• Bypass approval for HIGH-risk actions (delete rule, duplicate entity, change budget)

• Expose secrets — output sanitization redacts Meta tokens, JWTs, AWS keys, Stripe keys, Google API keys, DB strings before display

• Take action without your permission — even non-approval tools respect RBAC

See ai-102 how Wavo works for limits + caps.

Quick start

1. Open /chat

2. (Optional) Pick or create a project for context grouping

3. Type a question: "What's my top campaign by ROAS this week?"

4. Hit Enter — Wavo streams the answer

5. Follow up with refinements

Related

• How Wavo works — orchestration mechanics

• Asking questions — best practices

• Projects + sessions — organizing chats