Architectures
Four retrieval architectures compared under a frozen Mistral-7B-Instruct-v0.2 generator (greedy decoding, chat template applied).
| ID | Architecture | Type | Role | Key Parameters |
|---|---|---|---|---|
| (a) | BM25 | Non-neural baseline | Reference floor | k1=1.5, b=0.75 |
| (b) | BGE-M3 | Dense retriever (CLS pooling) | Primary dense baseline | lr=1e-5, batch=32, epochs=3, 1024-subword chunks |
| (c) | Hybrid BM25+BGE-M3+CrossEncoder | Lexical + Dense + Reranker | Expected strongest | RRF top-50 → rerank → top-10, max_length=1024, batch_size=4 |
| (d) | Legal-BERT Bi-Encoder | Domain-specific Transformer | Optional domain-reference | lr=2e-5, warmup=10%, batch=32, epochs=3, 512-subword chunks |
API Placeholder
GET /api/architectures — returns architecture configs and training status (pending)