Dev builds source-cited RAG for Indian GST FAQs with per-Q&A chunking
Suheet · reddit · 2026-09-07
A developer open-sourced a RAG assistant grounded in official Indian GST compliance FAQ documents, with a live demo.
- Semantic chunking: instead of fixed-size splits, each FAQ is extracted as one Q&A pair embedded as a unit (bge-m3 → Qdrant), avoiding answers cut across chunk boundaries
- Similarity threshold before anything reaches the LLM, shrinking the hallucination surface
- Mandatory citations plus "I don't know" fallback via system prompt; held up well against adversarial/off-topic tests
The author is soliciting feedback on whether per-Q&A chunking has a better-established pattern.
More from coding & agent
- Magento founder: Instinct agent books hotels and dinners via WhatsApp voice notes — vaibhavbetter · 2026-09-07
- Developer uses Claude Code to build a tiny 10k tok/s CPU model, shares 3 findings — GregoryDiamos · 2026-09-07
- GPT-6 Astra reverse-engineers Simpsons Hit & Run into a playable open-source three.js browser port — CtrlAltDwayne · 2026-09-07
- Creator tries Codex x PixVerse agent workflow to generate Four Gods and Phoenix video — Hailuo_AI · 2026-09-07
- LoopArena: Open Benchmark Tests Which LLMs Make the Best Runtime Controllers for Coding Agents — PepsiBetter · 2026-09-07
- AICanvas MCP 0.2.5 released, author laments slow enterprise adoption — ui_nerd · 2026-09-07