Two-pronged fix for repeated article topics:
1. Feed last 30 days of published article titles into the LLM analysis
prompt with explicit "avoid these" instruction. Zero cost — just
extra context in an existing API call.
2. Replace keyword-overlap dedup (60% word match) with embedding-based
cosine similarity via sentence-transformers/all-MiniLM-L12-v2 on
OpenRouter. Catches rephrasings that keyword matching misses
(e.g. "BPC-157 dosing" vs "how much BPC-157 to take"). Falls back
to keyword matching if the embedding API is unavailable.
Cost: ~$0.00001 per generation run ($0.02/M tokens, ~500 tokens/run).
New files:
- class-semantic-dedup.php — Embedding cosine similarity + keyword fallback
- class-open-router-embedding-provider.php — OpenRouter /embeddings client
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Move system prompt, user prompt, and performance context builders into
PRAutoBlogger_Analysis_Prompts. Analyzer retains orchestration, data
fetch, parsing, and storage. Backward-compatible private proxies kept.
Both files well under 300-line limit.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
