In beta · May 2026 Case study ~7 min read

A tax-research AI Indian practitioners can actually trust.

CBIC RAG is a quote-grounded research and notice-response assistant for the full Indian indirect-tax corpus — GST, customs, central excise, service tax, and the constantly accreting layer of notifications and circulars on top. It runs on hardware we own. It cites the section behind every claim. And when it doesn't know, it says so.

Refuses on out-of-scope questions every time. Cites the source section every time it doesn't.
Request beta access Book a 30-min call
/ Status   In beta with chartered-accountancy firms in Amritsar and Delhi.
Editorial illustration: an open Indian tax statute with a single sentence highlighted; a thin orange line traces from the highlighted line to a small architectural block diagram, depicting quote-grounded answers as system architecture.
Quote-grounded answers visualized as architecture.

The practitioner's problem

Ask any working CA how they research a tax question and the answer is a workflow held together with bookmarks, memory, and patience. The CBIC website's keyword search returns document titles, not the relevant section. PDFs get opened in a dozen tabs. The right clause for a scenario phrased as "our client received an adverse advance ruling, what's the appeal window" lives on page 147 of an act, three sub-clauses deep, and finding it is the work.

The new generation of AI assistants doesn't fix this. Most of them wrap a frontier model with a "you are a tax assistant" preamble and answer fluently — sometimes citing real provisions, sometimes inventing section numbers that don't exist. A practitioner can't issue billable advice on output like that. Verifying every cited section yourself takes longer than the original lookup.

There's a quieter problem underneath: the questions practitioners ask are usually about specific clients. A tool that pipes those queries to a US model vendor isn't usable for compliance-sensitive work. The data sovereignty question isn't theoretical — it's the reason most firms have not adopted any AI tool yet.

A practitioner asks a tax question; the system returns a paragraph that reads like a legal brief, with each claim citing the section, document and page it came from.
The user-facing outcome — a cited paragraph, not a search result.

What CBIC RAG does

Section-aware ingestion pipeline feeding an on-prem embedding pool and a vector store; queries flow through retrieval, reranking and a synthesis LLM that cites or refuses.
Section-aware ingestion. Quote-grounded answers.

Two workflows, one underlying system.

The first is research. A practitioner types a tax question in plain English. The system finds the passages from CBIC source documents that actually answer it, composes an answer grounded in those passages, and shows the cited sections — section reference, document title, page range — so the practitioner can verify before advising.

The second is notice response. A large slice of an Indian tax practitioner's billable time goes to drafting responses to notices issued by tax departments — show-cause notices, demand notices, scrutiny intimations. The practitioner pastes the notice text. The system retrieves the cited provisions, the relevant supporting statute, and the surrounding circulars and FAQs, and helps draft a response with every claim grounded back to its source. The same retrieval substrate as research, applied to a workflow that pays per case.

Both workflows speak the same language to the practitioner: an answer, the sections it came from, links to the source PDFs. Or, when the question can't be grounded in the corpus, a clean refusal.

Four behaviours that make it different

A question enters retrieval; a groundedness judge decides yes or no. Yes leads to a cited answer; no leads to a clean refusal.
Refusal is engineered, not accidental.

Most tax-AI tools answer. CBIC RAG does four specific things on top of answering. They sound small. They reshape how a practitioner can use the system.

It cites every claim back to a section.

Every sentence in an answer is traceable to a specific passage from a specific CBIC document. Click the citation, the source PDF opens at the cited page. Not a summary, not a paraphrase — the actual statutory text the answer was composed from.

Example: ask about the appeal window for an adverse advance ruling, and the answer comes back with the section reference, the document name, and the exact paragraph used. You verify in seconds, not minutes.

It refuses cleanly when the question is out of scope.

Income tax is governed by a different board entirely. Ask CBIC RAG an income-tax question and it tells you so, instead of hallucinating an answer that sounds confident. The same is true for everything outside the indirect-tax corpus — companies-act questions, RBI matters, general legal trivia.

Example: ask "how do I file ITR-3" and the system declines, naming the reason. Most generic AI assistants will cheerfully invent an answer.

It runs on local hardware.

The retriever, the answer model, the reranker — none of them call out to a vendor API. A practitioner's queries don't leave the network the system runs on. That's the data-sovereignty story; it's also the cost-predictability story. A firm that adopts CBIC RAG isn't exposed to a model-provider price increase or a deprecation cycle.

Example: a question about a specific client's pending demand notice never reaches a US server. It's answered on a rig in our basement, by open-weight models, with no per-token meter running.

It doesn't compute numbers.

If the question asks for a tax computation on user-provided figures, the system returns the applicable rule, rate, and method — but does not perform the arithmetic. Language models are unreliable at math, and giving a wrong final number to a practitioner is a liability we won't transmit. The practitioner does the calculation; the system shows them the rule.

Example: "compute GST on a hotel bill of fifty thousand including a service charge" returns the relevant rate notification and the valuation rules — not a final rupee figure.

The corpus, in scope

CBIC RAG covers the full body of indirect-tax law that the Central Board of Indirect Taxes and Customs publishes: GST, customs, central excise, service tax, allied acts, and the entire layer of notifications, circulars, instructions, orders, and FAQs that sits on top of the primary statutes. Bilingual documents are indexed with their Hindi twin available as a citation, even though the system itself reasons in English — we don't claim bilingual reasoning we haven't built.

The architecture is corpus-shaped, not domain-shaped. The same retrieval substrate is intended to extend, over time, to other Indian legal corpora that share the structural properties — income tax, the Companies Act, RBI and SEBI circulars, labour law, case law. CBIC is the first comprehensive corpus we ship, not the last.

What "in beta" means here

Beta is not a marketing tense for us. It is a literal one.

Today, real chartered-accountancy firms in Amritsar and Delhi are running CBIC RAG against real research and notice-response questions. Their feedback shapes the next iteration — which questions still trip the retriever, which document categories need richer metadata, where the refusal logic is too cautious or not cautious enough. Working well enough for that loop to be productive. Not working well enough for us to call it done.

Beta access is open to other Indian tax practitioners who want the same kind of partnership.

How we ship a change

The studio runs an internal trust regime for CBIC RAG. Before a change ships — a new chunking rule, a retriever tweak, a model swap — it has to pass a fixed panel of evaluation gates against a curated set of gold queries and adversarial out-of-corpus queries. Did the right passage come back. Was the answer grounded in it. Did the system refuse on the questions it should have refused. Did it do all of that within an acceptable time and cost envelope.

If any gate fails, we fix the underlying spec and re-run. We don't patch and continue. That discipline is the reason the four behaviours above hold up under load, not just in a demo.

What we won't tell you

A few things we'll be transparent about not publishing.

We won't name the beta firms, even on request — until and unless they ask us to. The case-study credibility we'd gain from a logo isn't worth the relationship risk to a firm that's letting us into their workflow at this stage.

We won't publish exact recall numbers, gate thresholds, or the specific models and hardware we run. They're our work. We're glad to discuss the shape of the system in detail with serious prospects and collaborators. The internals stay internal.

And we won't promote CBIC RAG out of beta until the trust gates are clean against the full corpus. The system that ships out of beta will be the system that holds the line. Until then, this page says "in beta," and we mean it.

If you're considering an AI build for your own regulated corpus — legal, medical, financial, regulatory — the trade-offs we worked through here generalise. Book a 30-min call and we can walk through whether the same substrate fits.