fix(blog): anti-repetition engine — 6 angle types, forbidden structures, existing article context injection

packages/api/src/llm/fo-blog-pipeline.ts

@@ -12,6 +12,8 @@
  *   8. Kill AI Tone (remove all AI fingerprints)
  *   9. QA Check (technical accuracy + weak section fixes)
  *   10. Quality Score (1-10 ratings + improvement suggestions)
+ *   + APM (Auto-Precision Mode — final word-level cut)
+ *   + Viral Signal (Social Masterfile — AVC, carry line, auto-kill, LinkedIn)
  *
  * Dedicated FO_Blog_LLM:
  *   - Model: qwen2.5:14b on .213 (or override via FO_BLOG_MODEL env)
@@ -303,14 +305,23 @@ CONTENT MODULES (use 2-3 per article):
 
 export const STEP1_TOPIC_EXPANSION = `You are a senior network engineer.
 
-Given the topic below, expand it into:
+Given the topic below, expand it into scenarios and angles from SIX DIFFERENT PERSPECTIVES.
-- 5 real-world scenarios where this topic becomes a problem
+Do NOT default to "physical layer failure" or "lab-to-production gap" — those are overused.
-- 5 common wrong assumptions engineers make about this
+
-- 5 operational risks nobody talks about
+For each perspective, generate 2-3 concrete, specific observations:
+
+1. ECONOMIC PERSPECTIVE — TCO, hidden costs, budget allocation, vendor pricing, ROI reality
+2. OPERATIONAL PERSPECTIVE — procurement workflows, validation processes, team skills, SLAs
+3. MARKET/TIMING PERSPECTIVE — when to buy, when to wait, what's mature vs. hype
+4. TECHNICAL DEPTH PERSPECTIVE — specific failure modes at the protocol level, not just physical layer
+5. POLITICAL PERSPECTIVE — vendor lock-in, procurement decisions, OEM vs. compatible debates
+6. MIGRATION PERSPECTIVE — step-by-step realities when upgrading from previous generation
 
 Topic: {{TOPIC}}
 
-Keep it practical, not theoretical. Think about what actually goes wrong in production.`;
+{{EXISTING_ANGLES}}
+
+Be concrete, not generic. Think: what would a senior engineer with budget responsibility know that a junior engineer wouldn't?`;
 
 // ═══════════════════════════════════════════════════════
 // STEP 2: ANGLE SELECTION
@@ -318,16 +329,46 @@ Keep it practical, not theoretical. Think about what actually goes wrong in prod
 
 export const STEP2_ANGLE_SELECTION = `Based on the expanded scenarios below, select ONE strong angle for a technical blog post.
 
-The angle must be:
+════════════════════════════════════════════════════════
-- Practical and decision-driven (helps the reader DO something)
+ANGLE DIVERSITY — MANDATORY
-- Involves real trade-offs (not a clear-cut answer)
+════════════════════════════════════════════════════════
-- Relevant for real deployments (not academic)
-- Controversial enough to generate discussion
 
-Then define:
+{{FORBIDDEN_ANGLES}}
-- Target audience (e.g., DC leaf-spine engineer, ISP architect, enterprise campus)
+
-- Core decision question the article answers
+ANGLE TYPES TO CHOOSE FROM (pick the one that fits the topic best — rotate through these):
-- The one thing the reader should DO after reading
+
+TYPE A — ECONOMIC: "What this actually costs" (TCO, hidden spend, budget reality)
+  Example: "Your $350 optic just cost you $18,000 in engineering time — here's the math"
+
+TYPE B — DECISION FRAMEWORK: "How to decide" (buy now vs wait, OEM vs compatible, which spec)
+  Example: "The 3 questions that determine whether 400G ZR is right for your deployment"
+
+TYPE C — MARKET REALITY: "What's hype vs production-ready right now" (timing, maturity, supply chain)
+  Example: "800G: which parts of the standard are actually shippable today"
+
+TYPE D — OPERATIONAL PLAYBOOK: "Step-by-step process" (how to do something, not what goes wrong)
+  Example: "The 6-step validation checklist before you deploy any 400G transceiver"
+
+TYPE E — VENDOR POLITICS: "The uncomfortable truth about vendor dynamics"
+  Example: "Why OEM compatibility lists exist — and why they're not what you think"
+
+TYPE F — MIGRATION REALITY: "What the upgrade path actually looks like, not what the datasheet says"
+  Example: "12 months into our 100G→400G migration: what we got wrong in month 1"
+
+════════════════════════════════════════════════════════
+FORBIDDEN ANGLE STRUCTURES (these are overused — auto-reject if you start here):
+- "Lab worked fine → production failed → physical layer was the cause" → BANNED
+- "Compatible optics get blamed → investigation → connector was dirty" → BANNED
+- "400G exposes assumptions that 100G hid" → BANNED (used too many times)
+- Any structure where the resolution is "clean your connectors" → BANNED
+════════════════════════════════════════════════════════
+
+Select the angle type, then define:
+- ANGLE TYPE: (A/B/C/D/E/F)
+- ANGLE SUMMARY: one sentence describing the specific angle
+- TARGET AUDIENCE: (e.g., DC leaf-spine engineer with budget, ISP procurement lead, enterprise campus architect)
+- CORE QUESTION: the specific decision or insight the article answers
+- READER ACTION: the one thing the reader does differently after reading
 
 Expanded scenarios:
 {{SCENARIOS}}`;
@@ -342,13 +383,21 @@ NOT a section list. NOT a structure. A flow plan — the sequence of ideas as th
 
 FORMAT: Write the outline as 3-4 narrative beats. Each beat = one core idea and how it connects to the next. No bullet points. No section headers.
 
-The outline should describe:
+════════════════════════════════════════════════════════
-- Opening situation: what moment the reader is in
+BANNED OUTLINE STRUCTURES — DO NOT USE THESE:
-- Core tension: what assumption they have that is wrong
+- Beat 1: "Lab works fine" → Beat 2: "Production fails" → Beat 3: "Physical layer was the cause"
-- Production reality: 1-2 specific things that fail (described as moments, not scenarios)
+- Any arc where the climax is "dirty connector" or "polarity mismatch"
-- Consequence/resolution: what actually matters at the end
+- Any arc where the resolution is "validate your setup" as a generic close
+- Opening with an engineer working late in a DC finding a failing link
+════════════════════════════════════════════════════════
 
-Keep the outline focused on 3-4 ideas MAX. If you can't write it in 3-4 beats, it's too broad.
+INSTEAD — match the outline structure to the angle type:
+- ECONOMIC angle → open with a cost moment, close with a calculation the reader can use
+- DECISION angle → open with the choice the reader is about to make, close with clear criteria
+- MARKET angle → open with what the market says vs. what the data shows, close with timing advice
+- OPERATIONAL angle → open with a process gap, close with a concrete improved process
+- POLITICAL angle → open with the vendor dynamic, close with what independence actually costs/saves
+- MIGRATION angle → open with the planning assumption, close with what month 6 actually looked like
 
 Angle: {{ANGLE}}
 Target audience: {{AUDIENCE}}

packages/api/src/routes/blog.ts

@@ -16,7 +16,7 @@ import { pool } from "../db/client";
 const pipelineProgress = new Map<string, { step: number; total: number; label: string; pct: number }>();
 
 function setProgress(draftId: string, step: number, label: string): void {
-  const pct = Math.round((step / 17) * 92) + 2; // 2%..94% during run, 100% on complete
+  const pct = Math.round((step / 18) * 92) + 2; // 2%..94% during run, 100% on complete
   pipelineProgress.set(draftId, { step, total: 17, label, pct });
 }
 
@@ -31,6 +31,7 @@ import {
   ANTI_GENERIC_INTRO_PROMPT,
   QUALITY_CONTROL_PROMPT,
   PROCUREMENT_LAYER_PROMPT,
+  VIRAL_SIGNAL_PROMPT,
   buildTopicPrompt,
 } from "../llm/blog-prompts";
 
@@ -1019,7 +1020,7 @@ async function runLlmPipeline(
 
   const LLM_OPTS = { temperature: 0.7, maxTokens: 8192, timeoutMs: 480000 };
   const LLM_REFINE = { temperature: 0.4, maxTokens: 6144, timeoutMs: 480000 };
-  const TOTAL_STEPS = 17; // 16-step pipeline + APM final cut
+  const TOTAL_STEPS = 18; // 16-step pipeline + APM + Viral Signal + LinkedIn
   let stepsCompleted = 0;
 
   try {
@@ -1097,11 +1098,35 @@ async function runLlmPipeline(
     // Get blog type config
     const blogType = BLOG_TYPES[selectedTopic as keyof typeof BLOG_TYPES] || BLOG_TYPES.tutorial;
 
+    // Load existing articles to prevent angle repetition
+    let existingAnglesContext = "";
+    let forbiddenAnglesContext = "";
+    try {
+      const existingResult = await pool.query(
+        `SELECT title, draft_content FROM blog_drafts
+         WHERE status IN ('published', 'review', 'ready') AND draft_content IS NOT NULL
+         ORDER BY created_at DESC LIMIT 10`
+      );
+      if (existingResult.rows.length > 0) {
+        const summaries = existingResult.rows.map((r: { title: string; draft_content: string }) => {
+          // Extract first 150 chars of content as summary
+          const preview = (r.draft_content || "").replace(/^#[^\n]*\n/, "").trim().slice(0, 150);
+          return `- "${r.title}": ${preview}...`;
+        }).join("\n");
+
+        existingAnglesContext = `\n\nALREADY PUBLISHED ARTICLES (do NOT repeat their angles or structure):\n${summaries}\n\nFor this new article, choose a COMPLETELY DIFFERENT perspective and angle than any of the above.`;
+
+        forbiddenAnglesContext = `ALREADY WRITTEN ANGLES (forbidden — do not repeat these):\n${existingResult.rows.map((r: { title: string }) => `- "${r.title}"`).join("\n")}\n\nThe new article MUST have a structurally different angle — different story type, different reader takeaway, different perspective lens.\n`;
+      }
+    } catch { /* fine if no articles yet */ }
+
     // ═══ STEP 1: Topic Expansion ═══
     console.log("  Step 1/10: Topic Expansion...");
     setProgress(draftId, 1, "Step 1/10: Topic Expansion");
     const step1 = await generate(systemPrompt,
-      STEP1_TOPIC_EXPANSION.replace("{{TOPIC}}", title),
+      STEP1_TOPIC_EXPANSION
+        .replace("{{TOPIC}}", title)
+        .replace("{{EXISTING_ANGLES}}", existingAnglesContext),
       LLM_OPTS
     );
     stepsCompleted = 1;
@@ -1110,7 +1135,9 @@ async function runLlmPipeline(
     console.log("  Step 2/10: Angle Selection...");
     setProgress(draftId, 2, "Step 2/10: Angle Selection");
     const step2 = await generate(systemPrompt,
-      STEP2_ANGLE_SELECTION.replace("{{SCENARIOS}}", step1.text),
+      STEP2_ANGLE_SELECTION
+        .replace("{{FORBIDDEN_ANGLES}}", forbiddenAnglesContext)
+        .replace("{{SCENARIOS}}", step1.text),
       LLM_REFINE
     );
     stepsCompleted = 2;
@@ -1259,8 +1286,8 @@ async function runLlmPipeline(
     stepsCompleted = 15;
 
     // ═══ STEP APM: Auto-Precision Mode (Final Cut — last filter before publish) ═══
-    console.log("  Step 16/17: Auto-Precision Mode (final cut — if a word can go, it must go)...");
+    console.log("  Step 16/18: Auto-Precision Mode (final cut — if a word can go, it must go)...");
-    setProgress(draftId, 16, "Step 16/17: Auto-Precision Mode");
+    setProgress(draftId, 16, "Step 16/18: Auto-Precision Mode");
     const stepAPM = await generate(systemPrompt,
       STEP_APM.replace("{{ARTICLE}}", step9.text),
       LLM_REFINE
@@ -1271,34 +1298,72 @@ async function runLlmPipeline(
     const pctAPM = Math.round((1 - wordsAPM / wordsBeforeAPM) * 100);
     console.log(`    APM: ${wordsBeforeAPM} → ${wordsAPM} words (−${pctAPM}%) — precision cut done`);
 
-    // ═══ LinkedIn Post ═══
+    // ═══ STEP 17: Viral Signal — FLEXOPTIX Social Masterfile transformation ═══
-    console.log("  Step 17/17: LinkedIn Post (max 2,800 chars)...");
+    // Applies AVC (Auto-Viral-Check), ASS (Auto-Signal-Score), carry line enforcement,
-    setProgress(draftId, 17, "Step 17/17: LinkedIn Post");
+    // auto-kill phrase filter, and generates LinkedIn post in one pass.
-    let linkedinPost: string | null = null;
+    console.log("  Step 17/18: Viral Signal (Social Masterfile transformation)...");
-    let linkedinCharCount: number | null = null;
+    setProgress(draftId, 17, "Step 17/18: Viral Signal");
+    let viralArticle = stepAPM.text;
+    let viralLinkedinPost: string | null = null;
     try {
-      const stepLinkedIn = await generate(systemPrompt,
+      const stepViral = await generate(systemPrompt,
-        STEP_LINKEDIN_POST.replace("{{ARTICLE}}", stepAPM.text),
+        VIRAL_SIGNAL_PROMPT + "\n\nArticle:\n" + stepAPM.text,
-        { temperature: 0.6, maxTokens: 1024, timeoutMs: 120000 }
+        { temperature: 0.5, maxTokens: 8192, timeoutMs: 480000 }
       );
-      linkedinPost = stepLinkedIn.text.trim();
+      const viralOutput = stepViral.text.trim();
-      linkedinCharCount = linkedinPost.length;
+      // Parse output: article + ---LINKEDIN--- + linkedin post
-      // Enforce hard limit — truncate at last sentence before 2800 if too long
+      const linkedinSep = viralOutput.indexOf("---LINKEDIN---");
-      if (linkedinCharCount > 2800) {
+      if (linkedinSep !== -1) {
-        linkedinPost = linkedinPost.slice(0, 2800).replace(/[^.!?]*$/, "").trim();
+        viralArticle = viralOutput.slice(0, linkedinSep).trim();
-        linkedinCharCount = linkedinPost.length;
+        viralLinkedinPost = viralOutput.slice(linkedinSep + "---LINKEDIN---".length).trim();
-        console.log(`    LinkedIn post truncated to ${linkedinCharCount} chars`);
+        console.log(`    Viral Signal: article ${viralArticle.split(/\s+/).length} words + LinkedIn ${viralLinkedinPost.length} chars`);
       } else {
-        console.log(`    LinkedIn post: ${linkedinCharCount} chars`);
+        // No separator — treat entire output as article
+        viralArticle = viralOutput;
+        console.log(`    Viral Signal: article ${viralArticle.split(/\s+/).length} words (no LinkedIn section)`);
+      }
+      // Validate viral output isn't too short (LLM may have over-cut)
+      if (viralArticle.split(/\s+/).length < 400) {
+        console.log("    ⚠ Viral Signal output too short — falling back to APM output");
+        viralArticle = stepAPM.text;
       }
     } catch {
-      console.log("    LinkedIn post generation skipped");
+      console.log("    Viral Signal skipped (error) — using APM output");
     }
     stepsCompleted = 17;
 
-    // Extract only the article from APM output (APM returns clean article only)
+    // ═══ STEP 18: LinkedIn Post ═══
-    // Fall back to step9.text if APM output looks too short or empty
+    // Use Viral Signal LinkedIn if available, otherwise generate via STEP_LINKEDIN_POST
-    let finalArticleText = stepAPM.text.trim().length > 200 ? stepAPM.text : step9.text;
+    console.log("  Step 18/18: LinkedIn Post (max 2,800 chars)...");
+    setProgress(draftId, 18, "Step 18/18: LinkedIn Post");
+    let linkedinPost: string | null = viralLinkedinPost;
+    let linkedinCharCount: number | null = viralLinkedinPost ? viralLinkedinPost.length : null;
+    if (!linkedinPost) {
+      // Fallback: dedicated LinkedIn post generator
+      try {
+        const stepLinkedIn = await generate(systemPrompt,
+          STEP_LINKEDIN_POST.replace("{{ARTICLE}}", viralArticle),
+          { temperature: 0.6, maxTokens: 1024, timeoutMs: 120000 }
+        );
+        linkedinPost = stepLinkedIn.text.trim();
+        linkedinCharCount = linkedinPost.length;
+      } catch {
+        console.log("    LinkedIn post generation skipped");
+      }
+    }
+    // Enforce hard limit — truncate at last sentence before 2800 if too long
+    if (linkedinPost && linkedinPost.length > 2800) {
+      linkedinPost = linkedinPost.slice(0, 2800).replace(/[^.!?]*$/, "").trim();
+      linkedinCharCount = linkedinPost.length;
+      console.log(`    LinkedIn post truncated to ${linkedinCharCount} chars`);
+    } else if (linkedinPost) {
+      console.log(`    LinkedIn post: ${linkedinCharCount} chars`);
+    }
+    stepsCompleted = 18;
+
+    // Extract article from Viral Signal output (or APM fallback)
+    // Fall back to step9.text if output looks too short or empty
+    let finalArticleText = viralArticle.trim().length > 200 ? viralArticle : step9.text;
     const articleMarkers = [
       "### COMPLETE FIXED ARTICLE",
       "## COMPLETE FIXED ARTICLE",
@@ -1345,8 +1410,8 @@ async function runLlmPipeline(
     await pool.query(
       `UPDATE blog_drafts
        SET draft_content = $1, word_count = $2,
-           generated_by = 'fo-blog-engine-v6',
+           generated_by = 'fo-blog-engine-v7',
-           pipeline_version = 'v6-precision-mode',
+           pipeline_version = 'v7-viral-signal',
            pipeline_steps_completed = $3,
            auto_qa_score = $4,
            outline = $5,
@@ -1390,13 +1455,13 @@ async function runLlmPipeline(
     }
 
     clearProgress(draftId);
-    console.log(`Blog FO Pipeline: ${draftId} complete — ${wordCount} words, ${stepsCompleted}/17 steps, QA: ${(autoQaScore as any)?.overall || "N/A"}/10, LinkedIn: ${linkedinCharCount ?? "n/a"} chars`);
+    console.log(`Blog FO Pipeline: ${draftId} complete — ${wordCount} words, ${stepsCompleted}/18 steps, QA: ${(autoQaScore as any)?.overall || "N/A"}/10, LinkedIn: ${linkedinCharCount ?? "n/a"} chars`);
   } catch (llmErr) {
     clearProgress(draftId);
-    console.warn(`Blog FO Pipeline failed at step ${stepsCompleted + 1}/16 for ${draftId}: ${(llmErr as Error).message}`);
+    console.warn(`Blog FO Pipeline failed at step ${stepsCompleted + 1}/18 for ${draftId}: ${(llmErr as Error).message}`);
     // Update with partial progress
     await pool.query(
-      `UPDATE blog_drafts SET pipeline_steps_completed = $1, pipeline_version = 'v5-narrative-control',
+      `UPDATE blog_drafts SET pipeline_steps_completed = $1, pipeline_version = 'v7-viral-signal',
        outline = $2, updated_at = NOW() WHERE id = $3::uuid`,
       [stepsCompleted, JSON.stringify({ error: (llmErr as Error).message, steps_completed: stepsCompleted }), draftId]
     ).catch(() => {});
@@ -1562,7 +1627,7 @@ blogRouter.post("/llm/reset-queue", (_req: Request, res: Response) => {
 blogRouter.get("/:id/progress", (req: Request, res: Response) => {
   const p = pipelineProgress.get(String(req.params.id));
   if (!p) {
-    res.json({ success: true, running: false, step: 0, total: 10, label: "Idle", pct: 0 });
+    res.json({ success: true, running: false, step: 0, total: 18, label: "Idle", pct: 0 });
     return;
   }
   res.json({ success: true, running: true, ...p });