Message flow

This page traces a single visitor message from the widget or API through Chatevo’s three-brain pipeline to a streamed reply.

End-to-end path

Visitor message (widget or API)
    → Auth + rate limits + deployment lookup
    → Intent Analyzer
    → Query Strategist (RAG / tools / hybrid / direct)
    → Response Synthesizer
    → Stream reply (+ citations, rich cards)

Step 1 — Ingress

The embeddable widget (widget.chatevo.ai/chat.js) or your Direct API client sends the message to api.chatevo.ai.
Chatevo validates the deployment, domain allowlist (widget), and plan quotas.
Conversation history loads from PostgreSQL for multi-turn context.

See Widget messages API and Deployments overview.

Step 2 — Intent Analyzer

The Intent Analyzer reads the latest message plus recent history and assigns intent — for example knowledge lookup, tool action, greeting, or handoff. On Starter+ plans, smart query understanding runs here. Free plans skip routing and always attempt RAG when knowledge bases are attached.

See Smart query understanding.

Step 3 — Query Strategist

Based on intent, the Query Strategist picks an execution path:

Path	When used
RAG	Answer should come from attached knowledge bases
Tools	Live data from API Request or MCP tools
Hybrid	Combine retrieved docs with a tool call (Standard+ full routing)
Direct	Small talk or questions answerable without retrieval

For RAG, the strategist may rewrite the query (full tier on Standard+) before hybrid search runs in Qdrant. For tools, it selects the tool and fills parameters from the conversation.

See Document retrieval architecture and Tools overview.

Step 4 — Response Synthesizer

The synthesizer receives:

Your assistant master prompt
Top retrieved chunks (with citation metadata)
Tool JSON responses (if any)
Model and RAG settings

It drafts the answer, cites sources where applicable, and attaches rich content only from verified tool payloads — never invented product data.

Step 5 — Stream to client

With stream: true, the widget receives SSE events as the pipeline progresses:

Event	Meaning
`step_message`	Progress text (“Searching knowledge base…”)
`tool_call` / `tool_result`	Tool execution lifecycle
`done`	Final message with optional `rich_content`
`error`	Failure details

See Widget SSE events.

Failure and fallback

If retrieval returns no chunks above the similarity threshold, or a tool call fails, the synthesizer follows your master prompt — typically acknowledging the gap and offering escalation rather than guessing.