How we make AI draft a notice reply without inventing case law

Ask a general-purpose chatbot to draft a reply to a DRC-01 and it will happily cite 'CCE v. Somewhere Industries, 2019' — a case that does not exist. It reads plausibly. It has a year, a forum, a ratio. And it is completely made up. For a CA, one such citation in a filed reply is a professional-conduct problem, not a typo.

So the single technical rule we will not compromise on is simple: every citation shown to you resolves to a real, verified row in our law corpus, or it does not ship.

1. A real corpus, kept current

We maintain a structured corpus of Indian tax law — CGST/IGST Acts and Rules, GST circulars and notifications, and Income-Tax circulars — synced directly from CBIC and the Income-Tax department. New documents are pulled on a schedule, de-duplicated by content hash, and stored with their source URL and issue date.

Crucially, a crawled document is not trusted automatically. It stays unverified until a curator approves it. Only verified rows are eligible to appear in a draft.

2. Retrieval before generation

When you upload a notice, we don't just hand the PDF to a model and hope. We first retrieve the specific sections, rules, circulars and case law relevant to the allegations — using hybrid search (keyword + semantic embeddings tuned for legal text). The model drafts against those retrieved authorities, not against its training memory.

3. A citation validator on the way out

Before a draft is shown, every citation it contains is checked back against the corpus. Anything that doesn't resolve is flagged. You can require that exports are blocked entirely if any citation is unverified — so a hallucinated cite physically cannot make it into your DOCX or PDF.

The result

You still review and sign — AI drafts, you decide. But you start from a reply that cites the law correctly, with the source of every authority one click away. That is the difference between a tool you can put in front of a client and a chat box you keep second-guessing.