AI chat

The AI chat is the conversational layer in front of every Vizzor primitive. Predictor, Scanner, Whale Terminal, Forensics, Pre-news, Cross-venue — the chat can call any of them as a tool and stitch the results into a single reply.

Plain-English in, structured output back. You can say "Long ETH from 2,100 with TP1 at 2,150, TP2 at 2,200, SL at 2,080" and the chat arms three alerts, opens the position record, and replies with a confirmation. You can say "Audit 0xABCD on Base" and the chat runs the Scanner + the contract auditor and renders the verdict inline.

Multi-LLM, switchable per query

The chat is provider-agnostic. Four backends are supported, each with its own strengths:

Switching is a chat command: /llm anthropic · /llm openai · /llm gemini · /llm ollama. Per-query overrides via --llm= work on the CLI; the REST endpoint accepts ?provider=.

When no API keys are configured anywhere, Ollama is the automatic fallback. Vizzor ships with the integration baked in — install Ollama locally, pull a model, and the chat works fully offline against everything except the chains' RPCs.

Tool-use, native

The chat doesn't pretend to predict prices itself — it calls the Predictor. The same goes for every other Vizzor primitive:

Tool calls happen server-side — the LLM emits structured tool invocations, the server executes them, and the LLM gets the structured result back to weave into the reply. The user never sees the tool envelope; they see a natural-sounding reply backed by real data.

Plain-English trade plan automation

The chat's standout move is reading a trade plan from natural language and arming the alerts in one round-trip:

1you  >  Long ETH from 2,100. TP1 at 2,150, TP2 at 2,200, SL at 2,080.2 3vizzor >4✓ Position recorded: ETH long · entry 2,100.005✓ TP1 alert armed at 2,150.00  (+2.38%)6✓ TP2 alert armed at 2,200.00  (+4.76%)7✓ SL  alert armed at 2,080.00  (−0.95%)8 9Predictor's read on this setup:10  Tier:        tracked11  Direction:   long  ·  4h horizon12  Probability: 0.61  (calibrated)13  Notes:       on-chain whale flow confirms; funding-z neutral14               manipulation guard: clean15 16I'll DM you when any alert fires. Type /pending to see active trades.

The same flow works for shorts, range trades, scaled entries, and trailing stops. The chat infers the structure from natural-language cues — "with a trailing 1% stop", "scale in over 4h", "only if BTC holds above 60k" — and emits the matching set_trade_plan_alerts payload.

Conversation persistence

Every chat thread is durably stored in the conversations table (REST surface). Threads survive process restarts and can be reopened on the API or rendered on the Web Dashboard.

The conversation context is bounded: prompt-stitching uses the last N exchanges (default 20) plus a system-side summary of older context to keep token cost predictable. Long threads naturally compress; nothing is lost, but old context is summarized rather than re-paid for at every turn.

Where the chat surfaces

The Telegram and Discord chats double as the subscription surfaces — every command (/predict, /scan, etc.) is a structured shortcut for what the chat would do via tool calls. Power users mix the two: typed commands for fast lookups, free-text chat for compound queries.

Streaming + cost control

REST responses stream via SSE so a long reasoning chain renders progressively. Per-user quotas are enforced server-side via the billing tier:

• Free — limited LLM calls per day (default 0 — free is read-only on chat features)
• Trial — 10 calls / day during the 7-day window
• Pro — 200 calls / day
• Elite — unlimited

Per-article LLM operations (like the catalyst classifier) use prompt caching so repeated reads cost nothing within the cache window — the same article re-classified within 24h doesn't re-incur Claude / GPT cost.

Adjacent reading

• Predictor — the tool the chat calls most
• Configuration — provider switching, API keys, Ollama integration
• REST API — POST /v1/chat reference with SSE streaming