GoldenMatch

Phase 1 of the v1.0.0 Roadmap

In-Context LLM Clustering

Instead of asking the LLM "is A the same as B?" one pair at a time, GoldenMatch now sends blocks of 50-100 borderline records in a single prompt and asks the LLM to cluster them directly. The LLM sees all candidates at once and can make better group decisions.

llm_scorer:
  enabled: true
  mode: cluster          # new! (default: "pairwise" for legacy behavior)
  cluster_max_size: 100
  cluster_min_size: 5
  budget:
    max_cost_usd: 0.50

How it works:

1. Traditional pipeline scores all pairs as usual
2. Borderline pairs (0.75-0.95) are grouped into connected components
3. Each component is sent to the LLM as a single prompt
4. LLM returns cluster assignments with confidence scores
5. Results merge back into the pipeline seamlessly

Smart degradation: If a block is too large, it splits by removing weakest edges. If the LLM call fails, it falls back to pairwise scoring. If the budget runs out, it stops gracefully.

Uncertainty Scores

Every LLM cluster now carries a confidence score (0.0-1.0) from the LLM. Low-confidence clusters (< 0.7) are auto-flagged for human review.

Stats

• 875 tests passing (20 new, 0 regressions)
• CI green on Python 3.11/3.12/3.13

Install / Upgrade

pip install --upgrade goldenmatch

What's Next

• v0.6.0: Privacy-preserving record linkage (multi-party SMC)
• v1.0.0: API freeze, production-stable release