nothumanallowed 13.2.76 → 13.2.78

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "13.2.76",
+  "version": "13.2.78",
   "description": "NotHumanAllowed — 38 AI agents, 80 tools, Studio (visual agentic workflows). Email, calendar, browser automation, screen capture, canvas, cron/heartbeat, Alexandria E2E messaging, GitHub, Notion, Slack, voice chat, free AI (Liara), 28 languages. Zero-dependency CLI.",
   "type": "module",
   "bin": {

package/src/commands/ui.mjs

@@ -2852,24 +2852,29 @@ export async function cmdUI(args) {
 
           // ── Fetch REAL data for each agent type ──────────────────────
           if (agent === 'DocumentReaderAgent') {
-            // Extract text from attached PDF, then ask the LLM to structure it cleanly.
-            // The structured output becomes context for all subsequent steps.
-            sendToken('[Reading attached document...] ');
+            // Always use vision for PDF reading — text extraction loses table structure,
+            // column alignment, and layout-dependent data for the vast majority of
+            // technical PDFs (datasheets, catalogs, forms, scanned docs).
+            // Vision reads exactly what a human sees on the page.
+            sendToken('[Reading document with vision...] ');
             let rawText = '';
             if (stepPdfBase64) {
               try {
                 const b64 = stepPdfBase64.includes(',') ? stepPdfBase64.split(',')[1] : stepPdfBase64;
-                const pdfBuffer = Buffer.from(b64, 'base64');
-                rawText = extractTextFromPdf(pdfBuffer) || '';
-                if (!rawText || rawText.length < 20) {
-                  // Scanned PDF — use vision OCR
-                  sendToken('[No text layer — using vision OCR...] ');
-                  try {
-                    rawText = await callLLMVision(config, stepPdfBase64, 'application/pdf',
-                      'Extract ALL text from this document exactly as printed, preserving all numbers, codes, and values.');
-                  } catch (ve) { rawText = ''; }
-                }
-              } catch (e) { rawText = ''; }
+                rawText = await callLLMVision(
+                  config,
+                  'You are a technical document analyst. Extract ALL content exactly as it appears on the page.',
+                  'Extract ALL content from this document: all text, tables (preserve rows and columns with exact values), product codes, numbers, units, notes, headers. Do not summarize — transcribe everything visible.',
+                  { base64: b64, mimeType: 'application/pdf' }
+                );
+              } catch (ve) {
+                // Vision failed — fall back to text extraction
+                sendToken('[Vision unavailable — falling back to text extraction...] ');
+                try {
+                  const b64 = stepPdfBase64.includes(',') ? stepPdfBase64.split(',')[1] : stepPdfBase64;
+                  rawText = extractTextFromPdf(Buffer.from(b64, 'base64')) || '';
+                } catch (e) { rawText = ''; }
+              }
             }
             if (!rawText) {
               sendToken('Could not extract text from the attached document.');

package/src/constants.mjs

@@ -5,7 +5,7 @@ import { fileURLToPath } from 'url';
 const __filename = fileURLToPath(import.meta.url);
 const __dirname = path.dirname(__filename);
 
-export const VERSION = '13.2.76';
+export const VERSION = '13.2.78';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';