Query Drafts with Claude

script

// AI-Powered Note Query for Drafts
// Two-pass approach: Filter relevant notes, then answer with context

// Get available Claude models
const models = AnthropicAI.knownModels();

// === HELPER FUNCTIONS ===

function generateDraftSummaries(drafts) {
  return drafts.map((d, index) => {
    const firstLine = d.content.split('\n')[0] || "(Untitled)";
    const preview = d.content.substring(0, 150).replace(/\n/g, ' ');
    const tags = d.tags.length > 0 ? ` [Tags: ${d.tags.join(', ')}]` : '';
    
    return {
      index: index,
      uuid: d.uuid,
      summary: `[${index}] ${firstLine}\n${preview}${tags}`
    };
  });
}

function filterRelevantDrafts(ai, model, question, summaries) {
  const summaryText = summaries.map(s => s.summary).join('\n\n');
  
  const filterPrompt = `You are helping filter notes for relevance to a user's question.

USER'S QUESTION: "${question}"

AVAILABLE NOTES:
${summaryText}

Task: Analyze which notes are relevant to answering the user's question. Return a JSON array of note indices that are relevant. Be generous in your selection - include notes that might contain useful context or related information.

IMPORTANT: 
- Return ONLY a valid JSON array of numbers, nothing else
- If no notes are relevant, return an empty array: []
- Example valid responses: [0, 3, 7] or [] or [1, 2, 3, 4, 5]

Response:`;

  try {
    const response = ai.quickPrompt(filterPrompt, {"model": model, "temperature": 0});
    
    // Clean up response - remove any markdown code blocks
    let cleaned = response.trim();
    cleaned = cleaned.replace(/```json\n?/g, '').replace(/```\n?/g, '');
    
    const relevantIndices = JSON.parse(cleaned);
    return relevantIndices;
  } catch (e) {
    console.log("Error parsing filter response: " + e.message);
    console.log("Raw response: " + response);
    // If parsing fails, return empty array (will answer without notes)
    return [];
  }
}

function answerWithContext(ai, model, question, relevantDrafts) {
  let prompt;
  
  if (relevantDrafts.length === 0) {
    prompt = `The user asked: "${question}"

No relevant notes were found in their collection.

Please answer this question using your general knowledge. Start your response by saying "No relevant notes found in your collection." and then provide a helpful answer.`;
  } else {
    const contextText = relevantDrafts.map((d, i) => {
      const title = d.content.split('\n')[0] || "(Untitled)";
      return `--- Note ${i + 1}: ${title} ---\n${d.content}`;
    }).join('\n\n');
    
    prompt = `You are answering a question using the user's personal notes as context.

USER'S QUESTION: "${question}"

RELEVANT NOTES FROM USER'S COLLECTION:
${contextText}

Task: Answer the user's question based primarily on the information in their notes. If the notes don't fully answer the question, you may augment with your general knowledge, but clearly distinguish between what's from their notes vs. your knowledge.

Be conversational and helpful. Do not explicitly list which notes you're referencing unless it adds value to the answer.

Answer:`;
  }
  
  return ai.quickPrompt(prompt, {"model": model});
}

// === MAIN SCRIPT ===

let f = () => {
  // Prompt user for question, model, and output preference
  let p = new Prompt();
  p.title = "AI Note Query";
  p.message = "Ask a question about your notes";
  
  p.addTextView("question", "Your Question", "");
  p.addLabel("label1", "Claude will search your notes and answer your question");
  
  p.addSelect("model", "Model", models, [models[0]], false);
  
  p.addSelect("output", "Output", 
    ["New Draft", "Append to Current"], 
    ["New Draft"], 
    false
  );
  
  p.addButton("Search & Answer");
  
  if (!p.show()) {
    return false;
  }
  
  const question = p.fieldValues["question"];
  const selectedModel = p.fieldValues["model"][0];
  const outputMode = p.fieldValues["output"][0];
  
  if (!question || question.trim().length === 0) {
    alert("Please enter a question.");
    return false;
  }
  
  // Get all drafts
  const allDrafts = Draft.query("", "all", [], [], "modified", false, false);
  
  console.log(`Found ${allDrafts.length} total drafts`);
  
  // Generate summaries
  const summaries = generateDraftSummaries(allDrafts);
  
  // Initialize AI
  const ai = new AnthropicAI();
  
  // PASS 1: Filter relevant drafts
  console.log("Pass 1: Filtering relevant drafts...");
  const relevantIndices = filterRelevantDrafts(ai, selectedModel, question, summaries);
  
  console.log(`Found ${relevantIndices.length} relevant drafts`);
  
  // Get full content of relevant drafts
  const relevantDrafts = relevantIndices.map(idx => allDrafts[idx]);
  
  // PASS 2: Answer the question
  console.log("Pass 2: Generating answer...");
  const answer = answerWithContext(ai, selectedModel, question, relevantDrafts);
  
  // Check if we got an answer
  if (!answer || answer.length === 0) {
    alert("No response from Claude. Please check your API credentials.");
    return false;
  }
  
  // Format the result
  const timestamp = new Date().toLocaleString();
  
  // Add sources section with wiki-style links if we found relevant drafts
  let sourcesSection = "";
  if (relevantDrafts.length > 0) {
    const sourceLinks = relevantDrafts.map(d => {
      const title = d.content.split('\n')[0].replace(/^#+\s*/, '') || "(Untitled)";
      return `- [[${title}]]`;
    }).join('\n');
    sourcesSection = `\n\n## Sources\n\n${sourceLinks}`;
  }
  
  const result = `# Q: ${question}

**Model:** ${selectedModel}
**Date:** ${timestamp}
**Notes Found:** ${relevantIndices.length}

---

${answer}${sourcesSection}`;
  
  // Output based on user preference
  if (outputMode === "New Draft") {
    const newDraft = Draft.create();
    newDraft.content = result;
    newDraft.addTag("project/temporary/ai-query");
    newDraft.update();
    
    // Load the new draft in the editor
    editor.load(newDraft);
    
    console.log("Created new draft with answer");
  } else {
    // Append to current draft
    draft.content = draft.content + "\n\n" + result;
    draft.update();
    
    console.log("Appended answer to current draft");
  }
  
  // Show success message
  app.displaySuccessMessage("Query complete!");
  
  return true;
};

// Run the function
if (!f()) {
  context.fail();
}