Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.dzap.io/llms.txt

Use this file to discover all available pages before exploring further.

​
The loop

1. User prompt + metadata (wallet, chain) arrive
   └─ Vercel AI SDK generateText() with tools[]
2. Model produces a plan
   └─ Either: textual answer, OR tool calls
3. Tool calls execute in sequence
   └─ Each tool wraps a DZap API or external service
4. Tool results feed back into model
   └─ Model decides: another tool? answer? clarification?
5. Final text + structured tool calls return
   └─ result.finalText, result.toolCalls

​
Reasoning trace

For a query like “Swap 100 USDC for WETH on Arbitrum”: 
[1] BalanceTool: confirm USDC balance ≥ 100
    → 100 USDC available
[2] TokenAddressTool: resolve symbols
    → USDC: 0xaf88..., WETH: 0x82aF...
[3] ZapCallDataGeneratorTool: build the trade
    → quote: 100 USDC → 0.041 WETH, fee 0.05%, slippage 1%
[4] (with confirmation) PerformZapTool: execute
    → tx hash: 0xabc...
[5] Final answer with tx link
You see the structured trace in result.toolCalls; the user sees result.finalText.

​
Tool selection

Tools have descriptions. The model picks based on description match. To bias selection:
  • Make user intent explicit — “show top pools” reaches for PoolTool; “what’s my balance” reaches for BalanceTool.
  • Use metadata — accountInfo in metadata anchors BalanceTool and DefiPositionsTool automatically.
  • Custom system prompt — pass systemPrompt at SDK init to nudge selection (e.g. “always check balance before swap”).

​
Errors

When a tool fails, ZapBot retries with backoff (default 2 retries). On exhaustion, it surfaces the error to the model — which usually picks a different tool or asks the user for clarification. You can inspect retry behavior in result.toolCalls[i].error.

​
Streaming

sdk.ask() streams events when onStep is provided: 
await sdk.ask({
  userQuery: "...",
  metadata: { ... },
  onStep: (event) => {
    console.log(event.type, event);  // 'tool_call' | 'tool_result' | 'text_delta'
  },
});
Use this to render real-time progress in a UI.

​
Memory

ZapBot reads ChatHistory at the start of each turn and writes the new turn back. For multi-turn conversations: 
const sdk = new DZapSDK();
await sdk.initialize({ /* ... */ });

await sdk.ask({ userQuery: "What's the price of ETH?", sessionId: 'user-123' });
await sdk.ask({ userQuery: "Same question for BTC.", sessionId: 'user-123' });
// Second call sees the first turn's history.
Reset by passing a new sessionId or clearing memory explicitly.