credit-risk-retriever-bge-small-india

A 33M-parameter sentence-transformer fine-tuned for dense retrieval over Indian consumer-credit regulation: CICRA 2005, RBI Master Directions (Credit Information Companies 2025, Credit Information Reporting 2025, IRACP 2025, Cards Issuance & Conduct 2025, NBFC Scale Based Regulation, Internal Ombudsman, Digital Lending 2025) and allied statutes (SARFAESI, DPDP 2023, Consumer Protection Act 2019).

Base model: BAAI/bge-small-en-v1.5.

Intended use

Given a compliance/regulatory question in English, retrieve the most relevant passages from Indian consumer-credit regulatory text. Best paired with a hybrid pipeline (BM25 + dense + reranking) but effective standalone.

Out of scope: non-English queries, jurisdictions other than India, general consumer chat, and questions outside credit regulation.

Training data

520 triplets built from an 11-document official corpus:

• Anchor: a compliance-analyst question synthesized per passage by an LLM (2 questions per chunk; short, varied phrasing).
• Positive: the source passage itself.
• Hard negative: the top-ranked neighbour retrieved by the deployed hybrid index that is not the positive - legally adjacent text, which forces the encoder to learn fine-grained distinctions between neighbouring sections.

The generation and training pipeline deliberately excludes all hand-curated evaluation questions, so held-out numbers below are uncontaminated.

Training procedure

Setting	Value
Loss	MultipleNegativesRankingLoss (InfoNCE), triplet form
Learning rate	2e-5, cosine decay, 10% warmup
Batch size	64
Epochs	4
Max sequence length	320 tokens
Seed	42

Training loss: 3.39 -> 1.55 over 32 optimizer steps.

Evaluation

Held-out set of 78 gold-labelled questions (each anchored to a specific corpus document). A hit at rank k means a chunk from the gold document appears in the top-k. Stock baseline measured in the same run, same seed, same encoding code path.

Model	Recall@1	Recall@5	Recall@10	MRR@5
bge-small-en-v1.5 (stock)	0.680	0.910	0.962	0.764
this model	0.731	0.910	0.974	0.807

Recall@1 improves by +7.5% relative over stock and reaches parity with BM25 lexical retrieval on the same corpus (BM25: Recall@1 0.731 / Recall@10 0.923 / MRR@5 0.791), while dense retrieval extends its lead at k>=5 (Recall@10 +5.1 pts over BM25).

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("vivekkopthsd/credit-risk-retriever-bge-small-india")
model.max_seq_length = 320   # keep consistent with training

query = "Within how many days must a CIC complaint be resolved?"
docs = [
    "Complaints shall be resolved within thirty calendar days ...",
    "A credit card closure request must be honoured within seven working days ...",
]

q_emb = model.encode(["query: " + query], normalize_embeddings=True)
d_emb = model.encode(["passage: " + d for d in docs], normalize_embeddings=True)
scores = q_emb @ d_emb.T   # higher = more relevant

Prefix conventions from training: query: for questions, passage: for documents. Set max_seq_length = 320 at load time.

Limitations

• English only; Hindi-language regulations were not part of training.
• Domain-specific: tuned on Indian credit regulation; expect degraded quality on unrelated domains (general web text performed worse than stock in analogous runs).
• Trained on ~520 pairs - gains are documented at these depths rather than claimed as universally better than MS MARCO pretraining.