[P2-Performance] Upgrade embedding models for resume-JD similarity scoring

[Ztrimus]

Problem

Current metrics in metrics.py use:

• text-embedding-ada-002 (OpenAI) — old model, replaced by text-embedding-3
• models/text-embedding-004 (Gemini) — older model
• TF-IDF cosine similarity — basic, misses semantic relationships

Solution

Upgrade to modern embedding models:

Model	Price/1M tokens	Quality
text-embedding-3-large (OpenAI)	$0.13	Best commercial
text-embedding-3-small (OpenAI)	$0.02	Good balance
models/text-embedding-004 (Gemini)	$0.00	Free tier
BGE-M3 (open-source)	Free	Self-hosted

Also add reranking

For a two-stage pipeline:

1. Embed with cheap model for initial scoring
2. Rerank top matches with LLM or Cohere reranker for precision

Metrics to compute

• Keyword coverage: % of Tier 1 keywords present (not just any keyword match)
• Semantic similarity: embedding cosine between resume sections and JD requirements
• Overall match score: weighted combination

Files

• zlm/utils/metrics.py (upgrade models, add reranking)
• zlm/variables.py (add embedding model config)