nothumanallowed 13.2.72 → 13.2.73

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nothumanallowed",
-  "version": "13.2.72",
+  "version": "13.2.73",
   "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

@@ -2852,35 +2852,73 @@ export async function cmdUI(args) {
 
           // ── Fetch REAL data for each agent type ──────────────────────
           if (agent === 'DocumentReaderAgent') {
-            // Extract text from attached PDF and return it as the step output.
-            // This becomes context for all subsequent steps (WebSearchAgent etc.)
+            // 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...] ');
+            let rawText = '';
             if (stepPdfBase64) {
               try {
                 const b64 = stepPdfBase64.includes(',') ? stepPdfBase64.split(',')[1] : stepPdfBase64;
                 const pdfBuffer = Buffer.from(b64, 'base64');
-                const extracted = extractTextFromPdf(pdfBuffer);
-                if (extracted && extracted.length > 20) {
-                  toolData = `## Document: ${stepPdfName || 'attached'}\n\n${extracted.slice(0, 20000)}`;
-                } else {
-                  // Fallback: ask vision model to describe/OCR the document
-                  sendToken('[No text found — using vision OCR...] ');
+                rawText = extractTextFromPdf(pdfBuffer) || '';
+                if (!rawText || rawText.length < 20) {
+                  // Scanned PDF — use vision OCR
+                  sendToken('[No text layer — using vision OCR...] ');
                   try {
-                    const visionText = await callLLMVision(config, stepPdfBase64, 'application/pdf',
-                      `Extract ALL text, technical specifications, model numbers, part codes, product names, manufacturer details, dimensions, ratings, and any other data from this document. List every detail exactly as printed.`);
-                    toolData = `## Document (OCR): ${stepPdfName || 'attached'}\n\n${visionText}`;
-                  } catch (ve) {
-                    toolData = `Could not extract text from document: ${ve.message}`;
-                  }
+                    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) { toolData = `Document read failed: ${e.message}`; }
-            } else {
-              toolData = 'No document attached.';
+              } catch (e) { rawText = ''; }
+            }
+            if (!rawText) {
+              sendToken('Could not extract text from the attached document.');
+              clearInterval(keepalive);
+              sendEvent({ done: true, usage: { input: 0, output: 0 } });
+              res.end();
+              logRequest(method, pathname, 200, Date.now() - start);
+              return;
+            }
+            // Ask LLM to structure the raw extracted text into readable markdown
+            sendToken('[Structuring document content...] ');
+            const LANG_MAP_DOC = {en:'English',it:'Italian',es:'Spanish',fr:'French',de:'German',pt:'Portuguese',zh:'Chinese',ja:'Japanese',ar:'Arabic',hi:'Hindi',ru:'Russian',nl:'Dutch',pl:'Polish',tr:'Turkish',ko:'Korean',sv:'Swedish',da:'Danish',fi:'Finnish',no:'Norwegian',cs:'Czech'};
+            const docLang = LANG_MAP_DOC[(config?.language||'it').toLowerCase().slice(0,2)] || 'Italian';
+            const docSys = `You are a technical document analyst. Extract and structure the content of this document into clear, readable markdown. Respond in ${docLang}.
+Rules:
+- List ALL technical specifications with their exact values (codes, voltages, pressures, temperatures, dimensions, flow rates, etc.)
+- Use markdown headers (##), bullet points (-), and tables where appropriate
+- Do NOT invent, interpret, or add anything not present in the raw text
+- Include all product/part codes exactly as written
+- Keep all numeric values with their units`;
+            const docUser = `Here is the raw text extracted from "${stepPdfName || 'document.pdf'}". Structure it into clean, readable markdown:\n\n${rawText.slice(0, 18000)}`;
+            let structuredOutput = '';
+            let inThink = false;
+            try {
+              await withTimeout(
+                callLLMStream(config, docSys, docUser,
+                  (token) => {
+                    // Strip <think> blocks
+                    let buf = token;
+                    if (inThink) {
+                      const ci = buf.indexOf('</think>');
+                      if (ci >= 0) { buf = buf.slice(ci + 8); inThink = false; }
+                      else return;
+                    }
+                    const oi = buf.indexOf('<think>');
+                    if (oi >= 0) { buf = buf.slice(0, oi); inThink = true; }
+                    if (buf) { structuredOutput += buf; sendToken(buf); }
+                  },
+                  { max_tokens: 3000 }
+                ),
+                90000
+              );
+            } catch (e) {
+              // LLM failed — fall back to raw text
+              structuredOutput = `## ${stepPdfName || 'Document'}\n\n${rawText.slice(0, 8000)}`;
+              sendToken(structuredOutput);
             }
-            // Stream the extracted content as the step output directly — no LLM rewrite needed
-            sendToken(toolData);
             clearInterval(keepalive);
-            sendEvent({ done: true, usage: { input: 0, output: Math.ceil(toolData.length / 4) } });
+            sendEvent({ done: true, usage: { input: Math.ceil(rawText.length / 4), output: Math.ceil(structuredOutput.length / 4) } });
             res.end();
             logRequest(method, pathname, 200, Date.now() - start);
             return;

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 = '13.2.72';
+export const VERSION = '13.2.73';
 export const BASE_URL = 'https://nothumanallowed.com/cli';
 export const API_BASE = 'https://nothumanallowed.com/api/v1';