What's New

3-Stage Keyword Extraction Pipeline (Stage A/B/C)

• Stage A: Hard-filter boilerplate words + Chinese→English mapping
• Stage B: Intent anchor detection (90+ anchors across 15+ domains)
• Stage C: Synonym expansion (80+ keys) generating 3–8 anchored queries

Full Chinese/Mixed-Language Support

• 150+ term CHINESE_TECH_MAP covering tech, SaaS, medical, legal, education, manufacturing, agriculture, aerospace, religion, art, gaming, government, and more
• Sorted by key length to prevent compound match collisions
• Never returns raw Chinese — fallback strips unmapped characters
• 98%+ pass rate across 99 Taiwanese Chinese test cases

Quality Metrics

• 93 pytest tests (up from 73)
• 54-idea golden evaluation set: 100% anchor hit rate, 4% junk ratio
• Zero Chinese character leakage in query output

Bug Fixes

• Fixed synonym expansion duplicate word bug ("redis redis", "mcp server server")
• Fixed 追蹤 mapping: general tracking instead of infra-specific tracing
• Fixed fallback returning raw Chinese text as search queries

New Files

• scoring/synonyms.py — SYNONYMS dict + INTENT_ANCHORS set
• tests/golden_ideas.json — 54-idea evaluation set (EN + ZH)
• tests/eval_keywords.py — keyword quality evaluation script
• tests/test_tw_chinese.py — 46 TW Chinese test cases
• tests/test_tw_niche.py — 53 niche domain Chinese test cases

Full Changelog: https://github.com/mnemox-ai/idea-reality-mcp/blob/main/CHANGELOG.md