Skill-Extraction v0.2.2 (ESCO + KSA)

Release v0.2.2 · Skill-Extraction Refactor (“ESCO + KSA”)

⚡PR⚡#101

✨ Highlights

1. Taxonomy-aware skill extraction
   • Integrates a FAISS index of the ESCO skills taxonomy.
   • Skill_Extractor.get_top_esco_skills() now returns {Skill, index, score} enabling deterministic Skill Tag values (ESCO.<index>).

2. KSA enrichment with vLLM
   • New helper get_ksa_details() generates Knowledge Required and Task Abilities lists for each skill.
   • Automatically invoked when a GPU/vLLM backend is available.

3. Unified output schema
   The extractor returns a tidy DataFrame with seven columns:
   Research ID, Description, Raw Skill, Knowledge Required, Task Abilities, Skill Tag, Correlation Coefficient.

---

🔧 Detailed Changes

Area	Description
utils.py	get_top_esco_skills() enhanced to include ESCO index and similarity score.
llm_methods.py	Added get_ksa_details() plus supporting imports.
skill_extractor.py	• Ensured self.index is always defined. • build_faiss_index_esco() / load_faiss_index_esco() now instance methods storing the index under laiser/input. • New taxonomy-first pipeline inserted at the top of extractor(); legacy alignment kept for fallback.

---

⚠️ Deprecated / To Be Removed

• align_skills() and align_KSAs() will be dropped in v0.3 once consumers migrate to the new output format.

---

🚧 Known Issues / Roadmap

1. JSON parsing in get_ksa_details() needs additional resilience checks.
2. LLM calls are still executed per skill; batching will come in v0.3.
3. Duplicate import json lines remain in llm_methods.py.
4. Consider CPU-only fallback for KSA generation.
5. Persistence of the ESCO vector index should move to a cloud vector DB.
6. vLLM isn't supported on MPS/MacOS as of now.

---

⬆️ Upgrade Notes

pip install -U laiser==0.2.2 

No changes to input parameters are required, but downstream code should read the new seven-column schema.

---

Next up

0.2 → 0.3

• adding batching and dropping deprecated APIs; increment patch for bug fixes.