Context
Tax questions are a bad fit for a plain chatbot. The answer has to be right, it has to cite the actual section, and the arithmetic has to be exact. A model that confidently invents a clause is worse than no tool at all. VittBot is an attempt to get the helpfulness of a chat interface without letting the model freelance on the law or the numbers.
Problem
Two failure modes had to be designed out. First, hallucinated law: the assistant should never state a rule it cannot point to in the Income Tax Act 2025 or a CBDT circular. Second, fuzzy math: tax calculations like regime comparison and HRA exemption are deterministic, so a language model should not be the thing computing them. The scope is India only, every rule and disclaimer assumes Indian tax law.
Approach
Retrieval is hybrid. A BM25 lexical index catches exact section numbers and legal phrasing, and a BGE-M3 dense index catches the cases where a user asks in plain language. Together they pull the right passage so the model answers from cited text instead of memory. The calculations live in a separate JSON-configured rule engine: the old versus new regime comparison for FY 2025-26 under Section 115BAC, and the HRA exemption under Section 10(13A) with its metro and non-metro split. The model explains, the rule engine decides the figures.
Outcome
The MVP runs as a Streamlit app a CA can actually sit in front of, with section preset buttons, a citation expander so the source is one click away, and a calculator tab for the deterministic work. Keeping retrieval and calculation out of the model is what makes the answers trustworthy enough to hand to a professional.
Next Iteration
- Widen the corpus beyond the selected circulars as more are published.
- Add a saved-client context so repeated questions for the same filer carry forward.