Suheet/gst-compliance-faq-assistant

GST Compliance FAQ Assistant

A source-anchored RAG assistant that answers questions about Indian GST (Goods and Services Tax) compliance — returns, ITC, refunds, registration, and payments — grounded strictly in official GST FAQ documents. Every answer cites the specific FAQ section it came from, and the assistant explicitly says when it doesn't have a reliable source rather than guessing.

How it works

1. 1.Extraction: FAQ PDFs from the official GST tutorial site are parsed with pdfplumber, using font size/style heuristics to detect question boundaries, section headers, and unheaded sub-topics.
2. 2.Embedding: Rather than splitting documents into arbitrary fixed-size chunks (the common RAG default), each extracted Q&A pair is embedded as a single, semantically complete unit with BAAI/bge-m3. This avoids the classic fixed-size-chunking failure mode where a question and its answer get split across chunk boundaries, or an answer gets truncated mid-thought. Each point is stored in Qdrant with category, section, and page-range metadata for filtering.
3. 3.Retrieval: A user's question is embedded the same way and matched against the collection, with a minimum similarity threshold — weak matches are dropped before ever reaching the LLM.
4. 4.Generation: Retrieved excerpts are passed to an LLM (Groq's openai/gpt-oss-120b) with a system prompt that requires every claim to be traceable to a cited excerpt, and forbids filling gaps from general knowledge.

Stack

• —Vector DB: Qdrant Cloud (free tier)
• —Embeddings: BAAI/bge-m3 (sentence-transformers)
• —LLM: Groq (openai/gpt-oss-120b) via LangChain
• —UI: Gradio
• —Extraction: pdfplumber

Running locally

pip install -r requirements.txt

Create a .env file with:

GROQ_API_KEY=your-groq-key
QDRANT_URL=your-qdrant-cluster-url
QDRANT_API_KEY=your-qdrant-api-key

Then:

python app.py