This commit improves keyword generation by providing hierarchical context for each element and fixing the LLM response format parsing.
Changes:
1. lib/MissingKeywords.js:
- Add buildHierarchicalContext() to generate compact contextual info for each element
- Display hierarchy in prompt (e.g., "H2 existants: 'Titre1', 'Titre2'")
- For Txt elements: show associated MC keyword + parent title
- For FAQ elements: count existing FAQs
- Fix LLM response format by providing 3 concrete examples from actual list
- Add explicit warning to use exact tag names [Titre_H2_3], [Txt_H2_6]
- Improve getElementContext() to better retrieve hierarchical elements
2. lib/selective-enhancement/SelectiveUtils.js:
- Fix createTypedPrompt() to use specific keyword from resolvedContent
- Remove fallback to csvData.mc0 (log error if no specific keyword)
3. lib/pipeline/PipelineExecutor.js:
- Integrate generateMissingSheetVariables() as "Étape 0" before extraction
Prompt format now:
1. [Titre_H2_3] (titre) — H2 existants: "Titre1", "Titre2"
2. [Txt_H2_6] (texte) — MC: "Plaque dibond" | Parent: "Guide dibond"
3. [Faq_q_1] (question) — 3 FAQ existantes
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
StillHammer
3751ab047b