ArgoBox
ArgoBox AI + Infrastructure + Labs
Ctrl+K
Features

Chat System

Multi-interface chat architecture — public widget, admin chat, Argonaut agent, demo chat, and user workbench with streaming, model selection, and conversation persistence

February 23, 2026

Chat System

ArgoBox has five chat interfaces sharing common streaming and persistence patterns, each targeting different users and capabilities.

Chat Interfaces

ChatWidget (Public)

Property Value
File src/components/chat/ChatWidget.astro
Endpoint POST /api/public/chat
Model google/gemini-2.0-flash-001 (free via OpenRouter)
Persistence sessionStorage — last 20 messages, cleared on tab close
Rate limit 15 requests/hour per IP

Floating panel in the bottom-right corner on all public pages. Features a welcome state with starter chips, RAG-powered answers using the top 5 matching chunks (min score 0.3), and live infrastructure data injection. Max 8 messages of history sent per request, max 1024 tokens per response.

Private Mode

The ChatWidget includes a private mode toggle (lock icon in header) that unlocks an authenticated tier:

Property Public Private
Endpoint /api/public/chat /api/public/chat-private
Model Gemini Flash (free) Claude Sonnet
Rate limit 15/hour 30/hour
RAG scope Public (blog only) Private (vaults, sessions, knowledge-base)
Biography Public facts only Full biographical context (SENSITIVE.md from KV)
Max history 8 messages 16 messages
Max tokens 1024 2048

Authentication flow:

  1. User clicks lock icon → password overlay appears
  2. Password validated against PRIVATE_PASSWORD env var via a probe request
  3. On success, password stored in sessionStorage (argobox-chat-private key)
  4. All subsequent messages route to /api/public/chat-private with password in payload
  5. Green border + "Private" indicator shows active state
  6. If server returns 401, auto-drops to public mode

Environment variables:

  • PRIVATE_PASSWORD — Plain text password (CF Pages production env)
  • KV key data:sensitive-bio — Contents of DANIEL-LAFORCE-SENSITIVE.md (loaded by assistant-prompt.ts in private scope)

Admin Chat

Property Value
File src/pages/admin/chat.astro
Endpoint GET/POST /api/admin/chat
Models Full catalog — Ollama, Anthropic, OpenAI, Google, Groq, NVIDIA NIM, OpenRouter
Persistence localStorage — up to 50 conversations

Full-featured interface with sidebar conversation list, model selector with categories (pricing tier, route type, context window), custom system prompt editor, and conversation search. Auto-titles conversations from the first message.

Sensitivity routing: Messages containing private topics (health, PII, legal, relationships, credentials) are automatically routed to groq/llama-3.3-70b-versatile for privacy-safe processing.

Argonaut Chat

Property Value
File src/pages/admin/argonaut/chat.astro
Endpoint POST /api/argonaut/chat
Backends Argonaut (default) + OpenClaw (toggle)
Persistence localStorage — up to 50 conversations

Advanced admin interface with two switchable backends:

  • Argonaut: Full agent with RAG, tool calling, and user profiles. Three-tier fallback: daemon proxy → local agent → cloud-only.
  • OpenClaw: Orchestration agent for automation tasks.

Extras: RAG mode toggle (Full: 85K chunks unsanitized vs Safe: 46K chunks blog-only), voice conversation mode with Whisper STT + ElevenLabs TTS, agentic tool indicators, and sources panel.

Registered tools (when daemon is available): Brave/DuckDuckGo search, Perplexity, file read/write/list, allowlisted shell commands, Resend email, Twilio voice/SMS, Google Calendar.

Demo Chat

Property Value
File src/pages/api/demo/chat.ts
Endpoint POST /api/demo/chat
Model google/gemini-2.0-flash-001 (free via OpenRouter)
Persistence None — ephemeral, cleared on session end
Rate limit 30 messages per demo sandbox

Demo-scoped chat for visitors exploring ArgoBox. Authenticated via demo session cookie (HMAC-signed), not admin auth. Uses the same RAG pipeline as the public chat but with demo-specific system prompt context.

Key restrictions vs. other interfaces:

  • RAG scope: Public/knowledge only — never searches private vaults or conversations
  • No tools: Pure chat, no agentic tool calling
  • No model selection: Fixed to budget model
  • Content moderation: Same blocklist as public chat (violence, explicit, prompt injection)
  • URL rewriting: The demo iframe's fetch interceptor rewrites /api/argonaut/chat and /api/public/chat to /api/demo/chat so the admin chat UI works transparently in demo mode

User Workbench Chat

Property Value
File src/pages/api/user/chat.ts
Endpoint POST /api/user/chat
Models Free models only
Persistence None (API-only)

Portal user access with role-based permissions (portal:workbench). Restricted to the free model list. Supports custom system prompts.

Provider Routing

The admin chat endpoint resolves providers by model prefix:

Prefix Provider Requirement
ollama/ Ollama (local) Ollama running on network
anthropic/ Anthropic Direct ANTHROPIC_API_KEY set
openai/ OpenAI Direct OPENAI_API_KEY set
google/ Google Gemini Direct GOOGLE_API_KEY set
groq/ Groq GROQ_API_KEY set
nvidia-nim/ NVIDIA NIM NVIDIA_NIM_API_KEY set
Everything else OpenRouter OPENROUTER_API_KEY set

Unified Chat Router

POST /api/admin/unified-chat classifies user intent with a two-tier system (regex patterns first, Groq LLM fallback) and routes to the appropriate backend:

Intent Backend Use Case
brain RAG + knowledge lookup Questions about the site or infrastructure
forge Tmux session dispatch Coding tasks
terminal forge-shell session Shell commands
openclaw Orchestration agent Agentic tool use
content Content-specific prompts Blog/journal writing
private Groq (privacy-safe) Sensitive topics

The mode parameter can override intent detection: code→forge, terminal→terminal, agent→openclaw.

Streaming Format

All endpoints use Server-Sent Events (SSE):

data: {"content": "text chunk"}
data: {"content": "more text"}
data: {"sources": [...]}
data: {"suggestions": [...]}
data: [DONE]

Client-side pattern:

  1. Fetch with AbortController for cancellation
  2. response.body.getReader() + TextDecoder for streaming
  3. Buffer lines, parse JSON from data: prefix
  4. Handle special events: [DONE], error, content, sources
  5. Update DOM incrementally with marked for markdown rendering
  6. Sanitize HTML output (strip scripts, iframes, event handlers)

Conversation Persistence

Interface Storage Key Limit
ChatWidget sessionStorage argobox-chat-history 20 messages
Admin Chat localStorage argobox-chat-conversations 50 conversations
Argonaut localStorage argonaut-conversations 50 conversations
Demo Chat None (ephemeral) — Cleared on session end

Each stored conversation includes: ID, title (auto-generated from first message), messages array, selected model, optional system prompt, and timestamps.

Key Files

File Purpose
src/components/chat/ChatWidget.astro Public floating widget
src/pages/admin/chat.astro Admin chat interface
src/pages/admin/argonaut/chat.astro Argonaut advanced interface
src/pages/api/public/chat.ts Public endpoint (free model + RAG)
src/pages/api/public/chat-private.ts Private endpoint (password auth + Claude Sonnet + full RAG)
src/pages/api/admin/chat.ts Admin endpoint (multi-provider)
src/pages/api/admin/personal-chat.ts Private vault endpoint (Groq-only)
src/pages/api/admin/unified-chat.ts Intent-routing endpoint
src/pages/api/argonaut/chat.ts Argonaut daemon proxy
src/pages/api/demo/chat.ts Demo chat endpoint (knowledge RAG + SSE)
src/pages/api/user/chat.ts Portal user endpoint
chataistreamingopenrouterargonautopenclawllm