A small language model built deep, not wide.
Not a general-purpose LLM with a compliance prompt bolted on. ReQL's Small Language Model starts deeply grounded in BFSI regulation — first corpus: India — with every answer traceable back to the circular, clause, and date it came from. Built to extend the same approach across industries and geographies over time.
SLM
Why we're building this.
A world where no financial institution — however small, however remote — is one missed circular away from a compliance failure. Where regulatory knowledge isn't locked inside a handful of expensive consultants and legal teams, but available instantly, accurately, and in plain language to whoever needs it.
To build the regulatory intelligence layer for financial services — starting with Indian BFSI — using a language model that understands how RBI, SEBI, MCA, and IRDAI regulation is written, cross-referenced, and amended, and that never answers a compliance question without showing its source.
Narrow domain. Deep coverage. Every answer cited.
General-purpose LLMs know a little about everything. A domain SLM is built the opposite way — and for regulated work, that matters more than raw parameter count.
What the model actually reads.
Every clause is chunked, embedded, and stored with its citation metadata intact — regulator, circular number, section, and effective date.
RBI · MCA · SEBI 2024 Public statute and regulator counts.
Small, deep, and grounded — not big and general.
A general-purpose LLM knows a little about everything. Our SLM is built the opposite way: narrow domain, deep coverage, and a hard requirement that every claim traces to a source document.
Corpus ingestion
Circulars, master directions, and amendments from RBI, SEBI, MCA, ICAI, CBDT, CBIC, FEMA, Ind-AS, and IRDAI, structured from 2021 to present.
Structured indexing
Every clause is chunked and embedded with its citation metadata intact — regulator, circular number, section, and effective date.
Domain fine-tuning
The model is tuned specifically on regulatory Q&A, gap analysis, and clause comparison — the actual tasks compliance teams perform daily.
Citation-grounded evaluation
Every output is scored not just on correctness, but on whether it cites the right source — answers without a valid citation are treated as failures.
Every answer either finds a source, or admits it can't.
A model tuned on a fixed corpus doesn't need to guess. On our internal regulatory-Q&A eval set, the SLM either returns a valid citation or explicitly declines.
Internal eval Measured on ReQL's regulatory-Q&A eval set; representative, not audited.
Training is the start. It keeps learning after that.
The corpus doesn't go stale and the model doesn't stay static — both update on a continuous loop, fed by new regulation and by how compliance teams actually use it.
New circular published
A new circular, amendment, or master direction is picked up as soon as the regulator releases it.
Ingested and cross-linked
It's chunked, cited, and cross-referenced against every existing clause it amends or supersedes.
Reviewed by compliance officers
Where an answer is corrected or flagged by a real user, that correction feeds back into evaluation.
Model re-evaluated
Accuracy and citation quality are re-scored against the updated corpus before anything ships to users.
Why a narrow, grounded model beats a bigger one here.
No hallucinated compliance advice
A general LLM will confidently invent a plausible-sounding rule. A model grounded in a fixed corpus either finds the source or says it can't find one — there's no in-between.
Keeps up with amendments, not just headlines
Regulation moves in circulars and master direction updates, not press releases. A domain-specific model is built to track that granularity continuously.
Self-hostable, because compliance data is sensitive
A smaller model can run inside your own perimeter — no policy document or client contract needs to leave your infrastructure to get an answer.
Deep before wide
Every design decision — from corpus selection to evaluation criteria — starts from how RBI, SEBI, and IRDAI actually publish and structure regulation. Depth first is what makes expansion into new industries and geographies reliable later, not a rewrite.
Deep in BFSI today. Built to go wider from here.
The corpus, the citation engine, and the evaluation approach are all built to generalize — the roadmap is to extend the same grounded, cited approach to new industries and new geographies, not to rebuild it from scratch each time.
Now — BFSI (India first)
RBI, SEBI, MCA, ICAI, CBDT, CBIC, FEMA, Ind-AS, and IRDAI, covering banks, NBFCs, fintechs, and insurers.
Next — more industries
Extending the same corpus-and-citation approach to sectors beyond BFSI as the model architecture generalizes.
Then — new geographies
Applying the same training methodology to other regulatory regimes, starting where demand is clearest.
Vision — cross-industry, cross-border
One grounded, citation-first regulatory intelligence layer, wherever compliance work happens.
See the SLM answer a real regulatory question.
Book a demo and bring a real circular or policy document — we'll run it live.