Skip to content
LeFlux

AI conversation

The agent is powered by a fast frontier LLM (Gemini 2.5 Flash by default; configurable on self-hosted installs). Visitor messages flow through:

  1. Frontend — visitor types in the chat input, hits Enter.
  2. WebSocket emit — message + current page context posted to the server.
  3. Prompt assembly — server builds a per-turn prompt: visitor message + page context + Sitemap + Cross-page knowledge + conversation history + a per-iteration state trace.
  4. LLM call — server streams the response. Each text chunk relays back to the widget for typewriter-style rendering.
  5. Action dispatch — when the LLM returns an action (click, navigate, sequence), the widget executes it and reports the result back.
  6. Iterate — server feeds the new context + result into the next LLM call. Up to 30 iterations per task.

Streaming

Responses stream chunk-by-chunk so visitors see the agent “writing” in real time. Latency to first token is typically under 800ms; full responses average 1.5–3s for short answers.

You can’t disable streaming today — it’s the only mode.

Conversation history

Each session keeps a sliding window of the last 50 messages (visitor + assistant + system traces). On every LLM call the prompt includes:

  • The first 2 messages (oldest, anchors context)
  • The last 50 (most recent)

Long conversations don’t blow the context window. The trim function keeps the prompt bounded.

Context resets

A “task” is a single visitor intent. The server resets per-task state (progress tracker, last-action ID, scroll counts) when:

  • Visitor types a new message that doesn’t look like an iterative continuation.
  • The current task hits task_complete or abort_workflow.
  • Iteration cap (30) is reached.
  • Loop guard fires (same click 3x, same nav 2x, scroll-only 3x).

History persists across tasks — the agent remembers earlier exchanges within the session — but task-scoped state resets so the next intent starts clean.

Multi-step tasks

A single visitor message can trigger multiple actions. Example: “fill out the contact form for me with my info”:

Iteration 1: action_plan → scroll to contact section
Iteration 2: ask_user → "What's your name?"
            (visitor types: Ahmed)
Iteration 3: ask_user → "Email?"
            (visitor: ahmed@example.com)
Iteration 4: ask_user → "Your message?"
            (visitor: Interested in pricing.)
Iteration 5: action_plan → type name, type email, type message, click submit
Iteration 6: task_complete → "Done! Your message was sent."

The visitor sees this as one continuous interaction. The agent decides when to ask, when to execute, when to verify.

Cancellation

Visitors can cancel a multi-step task at any point by typing “cancel” / “stop” / “ruko” / equivalent in any language. The intent classifier picks this up and routes to abort. The agent confirms (“Okay, cancelled”) and the session goes idle.

Memory across sessions

LeFlux doesn’t persist long-term per-visitor memory by default. Each session starts fresh.

If a visitor identifies themselves via a form (name, email, phone), those values land in user-data-store (localStorage, 30-day TTL) and the agent will auto-fill the same fields on subsequent forms. This is opt-in per visitor — controlled by your privacy policy.