@fuma-content/studio 1.0.0 → 1.0.2

package/src/app/api/ai/command/prompts.ts

@@ -1,313 +0,0 @@
-import type { ChatMessage } from "@/components/editor/use-chat";
-import type { SlateEditor } from "platejs";
-
-import { getMarkdown } from "@platejs/ai";
-import dedent from "dedent";
-
-import {
-  addSelection,
-  buildStructuredPrompt,
-  formatTextFromMessages,
-  getMarkdownWithSelection,
-  isMultiBlocks,
-} from "./utils";
-
-export function getChooseToolPrompt({ messages }: { messages: ChatMessage[] }) {
-  return buildStructuredPrompt({
-    examples: [
-      // GENERATE
-      'User: "Write a paragraph about AI ethics" → Good: "generate" | Bad: "edit"',
-      'User: "Create a short poem about spring" → Good: "generate" | Bad: "comment"',
-
-      // EDIT
-      'User: "Please fix grammar." → Good: "edit" | Bad: "generate"',
-      'User: "Improving writing style." → Good: "edit" | Bad: "generate"',
-      'User: "Making it more concise." → Good: "edit" | Bad: "generate"',
-      'User: "Translate this paragraph into French" → Good: "edit" | Bad: "generate"',
-
-      // COMMENT
-      'User: "Can you review this text and give me feedback?" → Good: "comment" | Bad: "edit"',
-      'User: "Add inline comments to this code to explain what it does" → Good: "comment" | Bad: "generate"',
-    ],
-    history: formatTextFromMessages(messages),
-    rules: dedent`
-      - Default is "generate". Any open question, idea request, or creation request → "generate".
-      - Only return "edit" if the user provides original text (or a selection of text) AND asks to change, rephrase, translate, or shorten it.
-      - Only return "comment" if the user explicitly asks for comments, feedback, annotations, or review. Do not infer "comment" implicitly.
-      - Return only one enum value with no explanation.
-    `,
-    task: `You are a strict classifier. Classify the user's last request as "generate", "edit", or "comment".`,
-  });
-}
-
-export function getCommentPrompt(
-  editor: SlateEditor,
-  {
-    messages,
-  }: {
-    messages: ChatMessage[];
-  },
-) {
-  const selectingMarkdown = getMarkdown(editor, {
-    type: "blockWithBlockId",
-  });
-
-  return buildStructuredPrompt({
-    backgroundData: selectingMarkdown,
-    examples: [
-      // 1) Basic single-block comment
-      `User: Review this paragraph.
-
-    backgroundData:
-  <block id="1">AI systems are transforming modern workplaces by automating routine tasks.</block>
-
-  Output:
-  [
-    {
-      "blockId": "1",
-      "content": "AI systems are transforming modern workplaces",
-      "comments": "Clarify what types of systems or provide examples."
-    }
-  ]`,
-
-      // 2) Multiple comments within one long block
-      `User: Add comments for this section.
-
-  backgroundData:
-  <block id="2">AI models can automate customer support. However, they may misinterpret user intent if training data is biased.</block>
-
-  Output:
-  [
-    {
-      "blockId": "2",
-      "content": "AI models can automate customer support.",
-      "comments": "Consider mentioning limitations or scope of automation."
-    },
-    {
-      "blockId": "2",
-      "content": "they may misinterpret user intent if training data is biased",
-      "comments": "Good point—expand on how bias can be detected or reduced."
-    }
-  ]`,
-
-      // 3) Multi-block comment (span across two related paragraphs)
-      `User: Provide comments.
-
-  backgroundData:
-  <block id="3">This policy aims to regulate AI-generated media.</block>
-  <block id="4">Developers must disclose when content is synthetically produced.</block>
-
-  Output:
-  [
-    {
-      "blockId": "3",
-      "content": "This policy aims to regulate AI-generated media.\\n\\nDevelopers must disclose when content is synthetically produced.",
-      "comments": "You could combine these ideas into a single, clearer statement on transparency."
-    }
-  ]`,
-
-      // 4) With <Selection> – user highlighted part of a sentence
-      `User: Give feedback on this highlighted phrase.
-
-  backgroundData:
-  <block id="5">AI can <Selection>replace human creativity</Selection> in design tasks.</block>
-
-  Output:
-  [
-    {
-      "blockId": "5",
-      "content": "replace human creativity",
-      "comments": "Overstated claim—suggest using 'assist' instead of 'replace'."
-    }
-  ]`,
-
-      // 5) With long <Selection> → multiple comments
-      `User: Review the highlighted section.
-
-  backgroundData:
-  <block id="6">
-  <Selection>
-  AI tools are valuable for summarizing information and generating drafts.
-  Still, human review remains essential to ensure accuracy and ethical use.
-  </Selection>
-  </block>
-
-  Output:
-  [
-    {
-      "blockId": "6",
-      "content": "AI tools are valuable for summarizing information and generating drafts.",
-      "comments": "Solid statement—consider adding specific examples of tools."
-    },
-    {
-      "blockId": "6",
-      "content": "human review remains essential to ensure accuracy and ethical use",
-      "comments": "Good caution—explain briefly why ethics require human oversight."
-    }
-  ]`,
-    ],
-    history: formatTextFromMessages(messages),
-    rules: dedent`
-      - IMPORTANT: If a comment spans multiple blocks, use the id of the **first** block.
-      - The **content** field must be the original content inside the block tag. The returned content must not include the block tags, but should retain other MDX tags.
-      - IMPORTANT: The **content** field must be flexible:
-        - It can cover one full block, only part of a block, or multiple blocks.
-        - If multiple blocks are included, separate them with two \\n\\n.
-        - Do NOT default to using the entire block—use the smallest relevant span instead.
-      - At least one comment must be provided.
-      - If a <Selection> exists, Your comments should come from the <Selection>, and if the <Selection> is too long, there should be more than one comment.
-    `,
-    task: dedent`
-      You are a document review assistant.
-      You will receive an MDX document wrapped in <block id="..."> content </block> tags.
-      <Selection> is the text highlighted by the user.
-
-      Your task:
-      - Read the content of all blocks and provide comments.
-      - For each comment, generate a JSON object:
-        - blockId: the id of the block being commented on.
-        - content: the original document fragment that needs commenting.
-        - comments: a brief comment or explanation for that fragment.
-    `,
-  });
-}
-
-export function getGeneratePrompt(editor: SlateEditor, { messages }: { messages: ChatMessage[] }) {
-  if (!isMultiBlocks(editor)) {
-    addSelection(editor);
-  }
-
-  const selectingMarkdown = getMarkdownWithSelection(editor);
-
-  return buildStructuredPrompt({
-    backgroundData: selectingMarkdown,
-    examples: [
-      // 1) Summarize content
-      "User: Summarize the following text.\nBackground data:\nArtificial intelligence has transformed multiple industries, from healthcare to finance, improving efficiency and enabling data-driven decisions.\nOutput:\nAI improves efficiency and decision-making across many industries.",
-
-      // 2) Generate key takeaways
-      "User: List three key takeaways from this text.\nBackground data:\nRemote work increases flexibility but also requires better communication and time management.\nOutput:\n- Remote work enhances flexibility.\n- Communication becomes critical.\n- Time management determines success.",
-
-      // 3) Generate a title
-      "User: Generate a short, catchy title for this section.\nBackground data:\nThis section explains how machine learning models are trained using large datasets to recognize patterns.\nOutput:\nTraining Machines to Recognize Patterns",
-
-      // 4) Generate action items
-      "User: Generate actionable next steps based on the paragraph.\nBackground data:\nThe report suggests improving documentation and conducting user interviews before the next release.\nOutput:\n- Update all technical documentation.\n- Schedule user interviews before the next release.",
-
-      // 5) Generate a comparison table
-      "User: Generate a comparison table of the tools mentioned.\nBackground data:\nTool A: free, simple UI\nTool B: paid, advanced analytics\nOutput:\n| Tool  | Pricing | Features         |\n|-------|----------|-----------------|\n| A     | Free     | Simple UI        |\n| B     | Paid     | Advanced analytics |",
-
-      // 6) Generate a summary table of statistics
-      "User: Create a summary table of the following statistics.\nBackground data:\nSales Q1: 1200 units\nSales Q2: 1500 units\nSales Q3: 900 units\nOutput:\n| Quarter | Sales (units) |\n|----------|---------------|\n| Q1       | 1200          |\n| Q2       | 1500          |\n| Q3       | 900           |",
-
-      // 7) Generate a question list
-      "User: Generate three reflection questions based on the paragraph.\nBackground data:\nThe article discusses the role of creativity in problem-solving and how diverse perspectives enhance innovation.\nOutput:\n1. How can creativity be encouraged in structured environments?\n2. What role does diversity play in innovative teams?\n3. How can leaders balance creativity and efficiency?",
-
-      // 8) Explain a concept (selected phrase)
-      'User: Explain the meaning of the selected phrase.\nBackground data:\nDeep learning relies on neural networks to automatically extract patterns from data, a process called <Selection>feature learning</Selection>.\nOutput:\n"Feature learning" means automatically discovering useful representations or characteristics from raw data without manual intervention.',
-    ],
-    history: formatTextFromMessages(messages),
-    rules: dedent`
-      - <Selection> is the text highlighted by the user.
-      - backgroundData represents the user's current Markdown context.
-      - You may only use backgroundData and <Selection> as input; never ask for more data.
-      - CRITICAL: DO NOT remove or alter custom MDX tags such as <u>, <callout>, <kbd>, <toc>, <sub>, <sup>, <mark>, <del>, <date>, <span>, <column>, <column_group>, <file>, <audio>, <video> unless explicitly requested.
-      - CRITICAL: when writing Markdown or MDX, do NOT wrap output in code fences.
-      - Preserve indentation and line breaks when editing within columns or structured layouts.
-    `,
-    task: dedent`
-      You are an advanced content generation assistant.
-      Generate content based on the user's instructions, using the background data as context.
-      If the instruction requests creation or transformation (e.g., summarize, translate, rewrite, create a table), directly produce the final result using only the provided background data.
-      Do not ask the user for additional content.
-    `,
-  });
-}
-
-export function getEditPrompt(
-  editor: SlateEditor,
-  { isSelecting, messages }: { isSelecting: boolean; messages: ChatMessage[] },
-) {
-  if (!isSelecting) throw new Error("Edit tool is only available when selecting");
-  if (isMultiBlocks(editor)) {
-    const selectingMarkdown = getMarkdownWithSelection(editor);
-
-    return buildStructuredPrompt({
-      backgroundData: selectingMarkdown,
-      examples: [
-        // 1) Fix grammar
-        "User: Fix grammar.\nbackgroundData: # User Guide\nThis guide explain how to install the app.\nOutput:\n# User Guide\nThis guide explains how to install the application.",
-
-        // 2) Make the tone more formal and professional
-        "User: Make the tone more formal and professional.\nbackgroundData: ## Intro\nHey, here's how you can set things up quickly.\nOutput:\n## Introduction\nThis section describes the setup procedure in a clear and professional manner.",
-
-        // 3) Make it more concise without losing meaning
-        "User: Make it more concise without losing meaning.\nbackgroundData: The purpose of this document is to provide an overview that explains, in detail, all the steps required to complete the installation.\nOutput:\nThis document provides a detailed overview of the installation steps.",
-      ],
-      history: formatTextFromMessages(messages),
-      outputFormatting: "markdown",
-      rules: dedent`
-        - Do not Write <backgroundData> tags in your response.
-        - <backgroundData> represents the full blocks of text the user has selected and wants to modify or ask about.
-        - Your response should be a direct replacement for the entire <backgroundData>.
-        - Maintain the overall structure and formatting of the background data, unless explicitly instructed otherwise.
-        - CRITICAL: Provide only the content to replace <backgroundData>. Do not add additional blocks or change the block structure unless specifically requested.
-      `,
-      task: `The following <backgroundData> is user-provided Markdown content that needs improvement. Modify it according to the user's instruction.
-      Unless explicitly stated otherwise, your output should be a seamless replacement of the original content.`,
-    });
-  }
-
-  addSelection(editor);
-
-  const selectingMarkdown = getMarkdownWithSelection(editor);
-  const endIndex = selectingMarkdown.indexOf("<Selection>");
-  const prefilledResponse = selectingMarkdown.slice(0, endIndex);
-
-  return buildStructuredPrompt({
-    backgroundData: selectingMarkdown,
-    examples: [
-      // 1) Improve word choice
-      "User: Improve word choice.\nbackgroundData: This is a <Selection>nice</Selection> person.\nOutput: great",
-
-      // 2) Fix grammar
-      "User: Fix grammar.\nbackgroundData: He <Selection>go</Selection> to school every day.\nOutput: goes",
-
-      // 3) Make tone more polite
-      "User: Make tone more polite.\nbackgroundData: <Selection>Give me</Selection> the report.\nOutput: Please provide",
-
-      // 4) Make tone more confident
-      "User: Make tone more confident.\nbackgroundData: I <Selection>think</Selection> this might work.\nOutput: believe",
-
-      // 5) Simplify language
-      "User: Simplify the language.\nbackgroundData: The results were <Selection>exceedingly</Selection> positive.\nOutput: very",
-
-      // 6) Translate into French
-      "User: Translate into French.\nbackgroundData: <Selection>Hello</Selection>\nOutput: Bonjour",
-
-      // 7) Expand description
-      "User: Expand the description.\nbackgroundData: The view was <Selection>beautiful</Selection>.\nOutput: breathtaking and full of vibrant colors",
-
-      // 8) Make it sound more natural
-      "User: Make it sound more natural.\nbackgroundData: She <Selection>did a party</Selection> yesterday.\nOutput: had a party",
-    ],
-    history: formatTextFromMessages(messages),
-    outputFormatting: "markdown",
-    prefilledResponse,
-    rules: dedent`
-      - <Selection> contains the text segment selected by the user and allowed to be modified.
-      - Your response will be directly concatenated with the prefilledResponse, so please make sure the result is smooth and coherent.
-      - You may only edit the content inside <Selection> and must not reference or retain any external context.
-      - The output must be text that can directly replace <Selection>.
-      - Do not include the <Selection> tags or any surrounding text in the output.
-      - Ensure the replacement is grammatically correct and reads naturally.
-      - If the input is invalid or cannot be improved, return it unchanged.
-    `,
-    task: dedent`
-      The following background data is user-provided text that contains one or more <Selection> tags marking the editable parts.
-      You must only modify the text inside <Selection>.
-      Your output should be a direct replacement for the selected text, without including any tags or surrounding content.
-      Ensure the replacement is grammatically correct and fits naturally when substituted back into the original text.
-    `,
-  });
-}

package/src/app/api/ai/command/route.ts

@@ -1,203 +0,0 @@
-import type { ChatMessage, ToolName } from "@/components/editor/use-chat";
-import type { NextRequest } from "next/server";
-
-import { createGateway } from "@ai-sdk/gateway";
-import {
-  type LanguageModel,
-  type UIMessageStreamWriter,
-  createUIMessageStream,
-  createUIMessageStreamResponse,
-  Output,
-  generateText,
-  streamText,
-  tool,
-} from "ai";
-import { NextResponse } from "next/server";
-import { type SlateEditor, createSlateEditor, nanoid } from "platejs";
-import { z } from "zod";
-
-import { BaseEditorKit } from "@/components/editor/editor-base-kit";
-import { markdownJoinerTransform } from "@/lib/ai/markdown-joiner-transform";
-
-import { getChooseToolPrompt, getCommentPrompt, getEditPrompt, getGeneratePrompt } from "./prompts";
-
-export async function POST(req: NextRequest) {
-  const { apiKey: key, ctx, messages: messagesRaw = [], model } = await req.json();
-
-  const { children, selection, toolName: toolNameParam } = ctx;
-
-  const editor = createSlateEditor({
-    plugins: BaseEditorKit,
-    selection,
-    value: children,
-  });
-
-  const apiKey = key || process.env.AI_GATEWAY_API_KEY;
-
-  if (!apiKey) {
-    return NextResponse.json({ error: "Missing AI Gateway API key." }, { status: 401 });
-  }
-
-  const isSelecting = editor.api.isExpanded();
-
-  const gatewayProvider = createGateway({
-    apiKey,
-  });
-
-  try {
-    const stream = createUIMessageStream<ChatMessage>({
-      execute: async ({ writer }) => {
-        let toolName = toolNameParam;
-
-        if (!toolName) {
-          const { output } = await generateText({
-            model: gatewayProvider(model || "google/gemini-2.5-flash"),
-            output: Output.choice({
-              options: isSelecting
-                ? (["generate", "edit", "comment"] as const)
-                : (["generate", "comment"] as const),
-            }),
-            prompt: getChooseToolPrompt(messagesRaw),
-          });
-
-          writer.write({
-            data: output as ToolName,
-            type: "data-toolName",
-          });
-
-          toolName = output;
-        }
-
-        const stream = streamText({
-          experimental_transform: markdownJoinerTransform(),
-          model: gatewayProvider(model || "openai/gpt-4o-mini"),
-          // Not used
-          prompt: "",
-          tools: {
-            comment: getCommentTool(editor, {
-              messagesRaw,
-              model: gatewayProvider(model || "google/gemini-2.5-flash"),
-              writer,
-            }),
-          },
-          prepareStep: async (step) => {
-            if (toolName === "comment") {
-              return {
-                ...step,
-                toolChoice: { toolName: "comment", type: "tool" },
-              };
-            }
-
-            if (toolName === "edit") {
-              const editPrompt = getEditPrompt(editor, {
-                isSelecting,
-                messages: messagesRaw,
-              });
-
-              return {
-                ...step,
-                activeTools: [],
-                messages: [
-                  {
-                    content: editPrompt,
-                    role: "user",
-                  },
-                ],
-              };
-            }
-
-            if (toolName === "generate") {
-              const generatePrompt = getGeneratePrompt(editor, {
-                messages: messagesRaw,
-              });
-
-              return {
-                ...step,
-                activeTools: [],
-                messages: [
-                  {
-                    content: generatePrompt,
-                    role: "user",
-                  },
-                ],
-                model: gatewayProvider(model || "openai/gpt-4o-mini"),
-              };
-            }
-          },
-        });
-
-        writer.merge(stream.toUIMessageStream({ sendFinish: false }));
-      },
-    });
-
-    return createUIMessageStreamResponse({ stream });
-  } catch {
-    return NextResponse.json({ error: "Failed to process AI request" }, { status: 500 });
-  }
-}
-
-const getCommentTool = (
-  editor: SlateEditor,
-  {
-    messagesRaw,
-    model,
-    writer,
-  }: {
-    messagesRaw: ChatMessage[];
-    model: LanguageModel;
-    writer: UIMessageStreamWriter<ChatMessage>;
-  },
-) =>
-  tool({
-    description: "Comment on the content",
-    inputSchema: z.object({}),
-    execute: async () => {
-      const { partialOutputStream } = streamText({
-        model,
-        prompt: getCommentPrompt(editor, {
-          messages: messagesRaw,
-        }),
-        output: Output.array({
-          element: z
-            .object({
-              blockId: z
-                .string()
-                .describe(
-                  "The id of the starting block. If the comment spans multiple blocks, use the id of the first block.",
-                ),
-              comment: z.string().describe("A brief comment or explanation for this fragment."),
-              content: z
-                .string()
-                .describe(
-                  String.raw`The original document fragment to be commented on.It can be the entire block, a small part within a block, or span multiple blocks. If spanning multiple blocks, separate them with two \n\n.`,
-                ),
-            })
-            .describe("A single comment"),
-        }),
-      });
-
-      for await (const comments of partialOutputStream) {
-        for (const comment of comments) {
-          const commentDataId = nanoid();
-
-          writer.write({
-            id: commentDataId,
-            data: {
-              comment,
-              status: "streaming",
-            },
-            type: "data-comment",
-          });
-        }
-      }
-
-      writer.write({
-        id: nanoid(),
-        data: {
-          comment: null,
-          status: "finished",
-        },
-        type: "data-comment",
-      });
-    },
-  });