Ask about Indian income tax and LexRAG answers from the current Income-tax Act, 2025 — instead of confidently quoting a section that was repealed years ago.
India replaced the Income-tax Act, 1961 with the Income-tax Act, 2025 (in force April 2026). Section numbers everyone memorised changed — but most documents an assistant ingests still describe the old law.
“Claim your life-insurance & PPF deduction under Section 80C of the Income Tax Act, 1961.”
Confident. Well-written. Wrong as current law.
The deduction is under Section 123, read with Schedule XV, capped at ₹1,50,000.
“80C” has zero operative hits in the 2025 Act.
A plain-English query matches the friendly old booklet better than dense legalese — so it retrieves “80C” and answers 80C.
Booklet vs statute look identical to an embedding. Nothing flags one as repealed education and the other as governing law.
Vector search smears rare anchors like “Schedule XV” or “Section 123” — it misses the very thing that makes the answer correct.
Most RAG scores (like RAGAS faithfulness) only check that the answer is grounded in the retrieved text. A system that retrieves the stale booklet and answers “80C” scores 0.9+ on faithfulness — perfectly grounded, in the wrong source.
Prefer the authoritative & current statute on points of law — without throwing away the plain-language source for everything else.
Each chunk is tagged — statute: authoritative · 2026; booklet: explanatory · 2020.
When a stale explainer and the statute both match, the statute leads — on points of law. Booklet-only questions still answer from the booklet.
BM25 nails exact anchors that embeddings blur (Schedule XV, section numbers), fused with dense recall via RRF.
The model is told the statute governs and must never present a repealed section as current — and to decline cleanly when the answer isn’t in the docs.
Query-expansion (old→new terms) → HyDE → Weaviate hybrid (BM25 + Jina vectors + RRF) → currency-aware selection → grounded generation with refusal.
STALE-LAW CORRECTNESS
held on a fresh held-out set it was never tuned on
RETRIEVAL PRECISION
+49% — less irrelevant context in the prompt
32 cases (19 tuning + 13 held-out). Correctness gate judged by GPT-4o (cross-validated against Gemini); grounding metrics by RAGAS. Ties on the easy categories (simple lookup, refusal) — as expected.
It answers only from the Income-tax Act, 2025 and the SEBI investor-education booklet. Off-topic questions are declined.