gatherBlogData():
- Fetches real prices from price_observations (last 30 days) per product
- Filters transceivers by speed extracted from topic keywords
- Enriches every product with verified_prices array + has_verified_price flag
- Joins DB products with vector search results (DB first — they have real prices)
contextData injection (blog.ts):
- [PRODUCT] lines: exact standard_name, form_factor, speed, reach, connector, dBm specs, Watts
- [VERIFIED PRICE] lines: real EUR/USD price, vendor, observed date, source URL
- [NO VERIFIED PRICE IN DB]: explicit tag — LLM must not invent a number
- [NO PRODUCT DATA AVAILABLE]: fallback when DB returns nothing
fo-blog-pipeline.ts system prompt:
- DATA INTEGRITY RULES block: prices/part numbers/vendors ONLY from context
- Never approximate with ~€350 or 'typically $200-600' for specific products
- Power specs only from [PRODUCT] data or REFERENCE VALUES
STEP4 context instructions:
- Explicit rules on how to use [VERIFIED PRICE] vs [NO VERIFIED PRICE]
- Invented data = HARD FAIL in QA
STEP9 QA — 3 new hard fail checks (30, 31, 32):
- Check 30: invented prices → remove or replace with flexoptix.net reference
- Check 31: invented part numbers → remove, use class name instead
- Check 32: invented vendor names → remove if not in known list
- Fix SR4/DR4 fiber count: both use 8 fibers (4TX+4RX), difference is MMF vs SMF
- Fix power per port: 400G=10-15W/port, 800G=15-25W/port (not 1kW/port)
- Fix pricing context: always distinguish OEM ($1-5K) vs compatible ($200-600)
- Add HARD RULES 15-21: fiber count, power, pricing, no markdown headers,
max 6 sections, no repeated topics, flow over format
- Add QA CALIBRATION FAILS 16-21: same rules enforced at QA step
- Add fiber/power reference tables with correct values
- Strip markdown (##/###/####/**) from all output — plain text only
- Add Style B gold example (10/10 validated prose article)
- Update STEP5 reality injection with correct SR4→DR4 description
- Update STEP8 kill-AI-tone to strip markdown headers + merge duplicates
- CALIBRATION_GOLD_STANDARD now covers two validated styles: A (structured) and B (prose)
- Style B: no headers, no bullets, 1-3 sentence paragraphs, reframe ending
- STEP8_KILL_AI_TONE: prose conversion option for over-structured articles
- STEP4_MASTER_DRAFT: explicit style choice instruction (A vs B based on angle)
- Gold standard includes exact prose rhythm patterns from 10/10 human-reviewed article
- Wrong patterns expanded: symmetric sections, checklist endings, transition clichés
- POST /api/blog/:id/regenerate — re-runs full 10-step LLM pipeline on existing draft
- Regenerate button visible when quality_issues present or status=review
- SEO keywords now displayed as clickable #hashtags (copy-to-clipboard)
- fo-blog-pipeline: added PoE misuse, DR4 mislabeling, ZR/DR4 conflation as hard QA fails
- fo-blog-pipeline: 14 hard rules in system prompt (was 10)
- fo-blog-pipeline: CALIBRATION_GOLD_STANDARD + withCalibration() from 10/10 human review
- System prompt now includes gold standard example on every pipeline run
System prompt: 10 HARD RULES (non-negotiable, article fails QA without them)
- Mandatory WHAT BREAKS IN PRODUCTION section (2+ specific failures with symptoms/cause/fix)
- Mandatory HIDDEN COSTS section (cleaning, troubleshooting time, cabling redesign, training)
- Mandatory WHEN NOT TO USE section for every recommendation
- Absolute statement rule: NEVER without conditions/context
- Cabling reality: MPO polarity, SR4→DR4 migration, cleaning requirements
- Brutal hook requirement: not 'If you're still...' but 'You're about to sign a PO. Stop.'
- Minimum 2500 words (was 2000)
Step 5 (Reality Injection): Now checks for ALL mandatory sections and adds if missing
Step 9 (QA Check): Hard fail checks — article is NOT publishable without production failures + hidden costs
Feedback source: Human expert review scoring 7.5/10, targeting 9.5+
