AI-native knowledge base & Evernote/Obsidian alternative with instant sync & native MCP. Free on Cloudflare or self-hosted via Docker.
HTML ai-agentcloudflarecloudflare-d1cloudflare-r2cloudflare-workersdockerelectronevernoteevernote-alternativeknowledge-basemcpnote-takingobsidian-alternativeoffline-firstpersonal-knowledge-basepwaself-hostedsqlitetiptapweb-clipper
import { isStepCount, tool, ToolLoopAgent, type LanguageModel } from "ai";
import { z } from "zod";
import { infographicRequestsTimeline, type InfographicAgentEvent, type InfographicAgentRequest } from "@edgeever/shared";
import { throwIfAborted } from "./abort-signal";
export const runInfographicAgent = async (args: {
input: InfographicAgentRequest;
model: LanguageModel;
signal: AbortSignal;
onEvent: (event: InfographicAgentEvent) => void;
}) => {
const { input, onEvent } = args;
let proposed = false;
let asked = false;
let fallbackResponse = "";
let response = "";
const candidates = new Set(input.candidates);
const timelineRequested = infographicRequestsTimeline(input.prompt);
const agent = new ToolLoopAgent({
model: args.model,
instructions: `You edit one AntV infographic through a conversation. The current infographic content is the source of truth; previous assistant replies may describe changes that were not applied. Understand the user's meaning and previous turns. Keep the current template for content-only changes, including replacing one comparison subject. Change the template when the data relationship changes, even if the user does not name a diagram type. A request for one company's development history needs chronological events in a sequence-timeline template, not the old two-company comparison. Choose only from the allowed template IDs. Prefer a comparison template with a VS divider for two subjects and matched aspects. Preserve information the user did not ask to change. data.title is the note title: change it when the subject changes, and keep it when the user only edits details. Ask a short clarification only if you cannot make a useful edit. Call one accepted tool before your final concise response, in the user's language. Do not output JSON in your final response.\nCurrent template: ${input.currentTemplate ?? "none"}.\nCurrent infographic content: ${input.currentContent || "none"}.\nAllowed templates: ${input.candidates.join(", ")}.\nData shapes: chart={title,values:[{label,value:number}]}; comparison={title,compares:[{label,children:[{label,desc}]}]}; quadrant={title,compares:[four items]}; hierarchy={title,root:{label,children:[...]}}; relation={title,nodes:[{id,label}],relations:[{from,to}]}; sequence={title,sequences:[{label,desc}]}; list={title,lists:[{label,desc}]}. Binary comparisons need exactly two compares with matching children. Use concise labels and complete content.`,
tools: {
propose_infographic: tool({
description: "Propose a complete, editable infographic version. Choose an allowed template and its matching data shape.",
inputSchema: z.object({
template: z.string(),
data: z.record(z.string(), z.unknown()),
explanation: z.string().max(500),
}),
execute: async ({ template, data, explanation }) => {
if (proposed || asked || !candidates.has(template)) return { accepted: false, reason: "Choose one allowed template and call one tool only." };
if (timelineRequested && !template.startsWith("sequence-timeline-")) return { accepted: false, reason: "The user requested one subject's development history. Choose an allowed sequence-timeline template and chronological events." };
proposed = true;
fallbackResponse = explanation;
onEvent({ type: "proposal", template, data, explanation });
return { accepted: true, template, instruction: "Briefly tell the user what changed." };
},
}),
ask_clarification: tool({
description: "Ask a focused question when the infographic cannot be edited correctly without the user's answer.",
inputSchema: z.object({ question: z.string().trim().min(1).max(300) }),
execute: async ({ question }) => {
if (proposed || asked) return { accepted: false, reason: "Call one tool only." };
if (timelineRequested && input.candidates.some((template) => template.startsWith("sequence-timeline-"))) return { accepted: false, reason: "The timeline request is clear. Propose a chronological infographic instead of asking for clarification." };
asked = true;
fallbackResponse = question;
onEvent({ type: "question", question });
return { accepted: true, instruction: "Repeat the question briefly to the user." };
},
}),
},
prepareStep: () => !proposed && !asked
? { toolChoice: "required" as const }
: { activeTools: [] as [], toolChoice: "none" as const },
stopWhen: isStepCount(4),
maxOutputTokens: 3000,
maxRetries: 0,
});
const messages = [
...input.history.flatMap((turn) => [
{ role: "user" as const, content: turn.prompt },
{ role: "assistant" as const, content: turn.response },
]),
{ role: "user" as const, content: input.prompt },
];
const result = await agent.stream({ messages, abortSignal: args.signal });
for await (const part of result.fullStream) {
throwIfAborted(args.signal);
if (part.type === "error") throw part.error;
if (part.type !== "text-delta" || !part.text || !proposed || asked) continue;
response += part.text;
if (response.length > 4000) throw new Error("The infographic agent response was too long.");
onEvent({ type: "text-delta", text: part.text });
}
if (!proposed && !asked) throw new Error("The infographic agent did not propose a change or question.");
if ((asked || !response.trim()) && fallbackResponse) onEvent({ type: "text-delta", text: fallbackResponse });
onEvent({ type: "finish" });
};