State-commissioned AI legal search platform built at NAIC under Azerbaijan’s AI Strategy 2025–2028, officially launched September 2025.
What It Does#
Operates across three simultaneous search spaces:
- Lexical — BM25 for exact-match and rare-term recall
- Semantic — SBERT/DPR fine-tuned on 500k+ legal documents
- Structural — legal-hierarchy-aware ranking (chapter, article, clause)
Pipeline#
- Query normalization
- Azerbaijani spell correction
- NER (quantized DistilBERT)
- Intent classification
- Hybrid retrieval (BM25 + dense)
- Cross-encoder + LRT re-ranking
- Result fusion and ranking
All stages run under a 200ms end-to-end latency budget.
Key Numbers#
| Metric | Value |
|---|---|
| Model size reduction | 39% (ONNX INT8) |
| End-to-end latency | < 200ms |
| Accuracy (MRR@10 / NDCG@5) | 85% |
| Corpus size | 500k+ documents |