RAG Assistant for MENA Regulatory Compliance

The firm fielded the same 40–50 recurring UAE corporate-tax and VAT questions every week. Each answer required pulling two or three regulatory PDFs and reconciling cross-references between legislation — work only senior staff could do reliably.

Turnaround averaged six hours per question. Junior associates escalated constantly because retrieving the correct sub-article manually was error-prone, and the firm was capacity-bound: they could not take on new clients without hiring more senior hands.

We built a retrieval-augmented assistant against a pgvector store of cleaned regulation PDFs. Ingestion runs a 2-pass pipeline — PyMuPDF extracts raw text, then an LLM pass structures it into article-aware chunks that survive retrieval.

Composite embeddings were the unlock. Each chunk is embedded as a weighted mix of its own text (70%), RAKE-extracted keywords (10%), and LLM-generated synthetic questions (20%). Cross-article references like “see Article 23 of Cabinet Decision No. 52” stopped being invisible to retrieval.

The assistant streams answers via SSE, runs a context-condensation pass before each LLM call to keep prompt windows tight, and preserves the full conversation (no sliding window) so follow-ups stay grounded. Auth is RS256 JWT with 10-minute access tokens and 30-day refresh.