BlazAI — the agent

Type or speak what you want. BlazAI classifies the intent, extracts the parameters, routes to the best provider, and hands back a one-click confirmation card. It executes; it does not merely advise.

Home: ai.blazpay.com · Chat routes: / and /c/:id · Cost: free

What it is

A conversational front door to everything in the platform. 27 intents covering trading, safety, automation, discovery and dApp building. Replies stream token-by-token over SSE, so it reads like a messaging app rather than a form submission.

The interaction contract is deliberately narrow: the agent never silently spends money. Anything with financial consequence resolves to a card the user signs.

The interaction model

Three repos, three responsibilities:

Repo	Does	Does not
bz-agent	Classify intent, extract parameters	Touch a chain, execute anything
bz-backend	Persist chat, execute the intent, produce { text, data, action }	Decide what the user meant
defi-dex	Stream, render the right card, drive the wallet	Interpret free text

The model stack

The intent engine is Blazpay's own: a LangGraph state machine in bz-agent with one prompt template and one Zod schema per intent, a Redis checkpointer for multi-turn state, and a PGVector knowledge base behind the fallback path. That graph — not any single model — is what turns a sentence into a structured, validated intent.

Underneath it, models are addressed by tier, never by name. A node asks for a tier; which model serves it is an env var.

Tier	Role	Used for
extraction	Fast, deterministic, cheap. temperature: 0	Every intent node — parameter extraction
conversation	Richer prose	The conversation node — DeFi answers
complex	Most capable	dApp creation, contract audit

A separate JSON-mode drafting model backs Autopilot policy generation, and a self-hosted Rasa NLU service plus a wit.ai integration back the legacy swap-ai chat path.

Every node's output is validated by a Zod schema, so a malformed model response fails structurally rather than propagating.

Why the tier indirection matters

Swapping a provider is an env change, not a code change — LLM_MODEL_EXTRACTION, LLM_MODEL_CONVERSATION, LLM_MODEL_COMPLEX. Nothing in the graph names a model. See agent graph.

Why temperature: 0 everywhere

Intent classification and parameter extraction are not creative tasks. The same sentence should produce the same intent every time, because users retry and support needs to reproduce. Creativity lives only in the conversation node's prose.

Capability discovery without a nav bar

The AI site has no navigation. Users find capabilities three ways:

1. Suggestion chips — GET /v1/blaz/suggestions and GET /v1/blaz/home-chips.
2. The ⌘K command palette — src/component/ai/CommandPalette.tsx fuzzy-searches ten agent actions plus past chats. Every action resolves to a prompt seeded into the composer via ?q=, never a route, never auto-sent. Openable anywhere via the blazai:palette-open window event.
3. Just asking. The CONVERSATION intent answers "what can you do".

The ?q= effect must key on location.search

The palette navigates while the chat is already mounted, so a mount-only effect never fires — leaving ?q= in the URL with an empty composer. Strip the param with react-router navigate(path, { replace: true }), not history.replaceState: replaceState leaves the router's location stale, so picking the same action twice registers as the same location and silently does nothing.

Voice

Voice input uses the Web Speech API and auto-submits after a pause. A user can speak "swap half my eth to usdc" and get the same card. Voice confirmations and hands-free navigation are roadmap, not shipped.

Session and history

Concern	Detail
Session model	bz-backend/v1/models/blazSession.js — { wallet, name, pending, activeDappProjectId }
Message model	v1/models/blazMessage.js — { session, senderType, text, action, data }
Identity	Keyed by wallet address. There is no cookie or bearer session
Chat list	GET /v1/blaz/sessions
One chat	GET /v1/blaz/chat/:chatId

Because history is keyed by wallet, a returning user with no wallet connected sees an empty sidebar — which is exactly why the AI shell keeps a persistent WalletControl pill despite having no other chrome. See domains.

Disconnect must clear the chat

enabled: !!address only stops the chat-list query; React Query keeps serving the last data, so the previous wallet's history stayed visible after disconnect — someone else's data on a shared machine. The fix clears via queryClient.removeQueries({ queryKey: ['address'] }) and resets local state, gated on a connected→disconnected transition so guests are not reset on mount.

Streaming protocol

POST /v1/blaz/message/stream emits three event types:

Event	Payload
meta	The resolved { action, data } — arrives first so the UI can start rendering the card
token	Successive text chunks
done	Stream complete

POST /v1/blaz/message is the non-streaming equivalent, returning the whole { text, data, action } at once.

Client: defi-dex/src/apis/ai.api.ts → sendAIMessageStream.

Accuracy

The measured non-fallback rate is 99.9%+ — roughly one unresolved reply in 18,000. That number is a proxy for accuracy, not a direct measure: it counts messages that reached a real intent rather than the fallback path, so it does not catch a message classified as the wrong intent.

Source: PROOF_POINTS in defi-dex/src/data/chainSeo.mjs, all-time snapshot.

Training-data loop

bz-backend/v1/routes/training.routes.js plus models/trainingDataModel.js provide a review pipeline: conversations get labelled, an operator approves or rejects, and approved examples export for prompt refinement.

Endpoint	Purpose
GET /v1/training	List candidates
GET /v1/training/stats	Counts by status
GET /v1/training/export	Export approved examples
PATCH /v1/training/:id	Approve / reject one
POST /v1/training/bulk	Bulk update

Under the hood

Concern	Where
Intent list	bz-agent/src/utils/constant.ts → AVAILABLE_INTENTS
Entry prompt	bz-agent/src/template/entry.ts → getInitSystemTemplate(intents)
Per-intent templates	bz-agent/src/template/*.ts
Nodes	bz-agent/src/agent/node.ts (+ bzTokenNode, createDappNode, goOnchainNode, modifyDappNode)
Graph	bz-agent/src/agent/graph.ts
Compilation + checkpointer	bz-agent/src/agent/index.ts, src/utils/redisCheckPointer.ts
Vector store	bz-agent/src/agent/vector.ts — PGVector on PostgreSQL
Dispatcher	bz-backend/v1/services/agent.service.js → getAgentResponse
Controller	bz-backend/v1/controllers/blaz.controller.js
Renderer	defi-dex/src/component/defi/swapAI/BotMessage.tsx

Deep dives: agent graph, chat request path, action renderer.

Known limits

• No perpetuals intent. The agent cannot open or manage a perp position.
• LIMIT_ORDER classifies but cannot execute — the spot limit product is not shipped. The user gets a text reply, which reads as a failure. See limit orders.
• The DCA amount semantic mismatch — the agent template says per-cycle, the app deposits it as total. See CSIP.
• connect_wallet is EVM-only and only appears via a suggestion chip or the five-message gate. Solana and TRON connect through the pill.
• No message-count quota exists. The only per-wallet limit in the platform is the audit quota (SystemSetting.audit_free_limit_per_wallet, default 3). If a chat quota is needed, the pattern to copy is that one.
• Voice is input-only.