SalesTaxGST

Problem#

A Sales Tax/GST consultant needed an AI platform that could work directly over their real document corpus — case files, filings, and reference material — to help answer questions, maintain context across ongoing conversations, and assist in drafting appeals and similar documents. Unlike a general-purpose assistant, this had to be grounded in the consultant's own, large, real-world set of documents, not a small curated demo corpus.

My Contributions#

I built the Docling-based ingestion pipeline, the hierarchical RAG system, cross-chat conversational context, the document-drafting assistance features, and a two-way file sync integration that also served as a backup layer for the client's data.

Architecture#

At a high level: documents are ingested via Docling, indexed through a hierarchical RAG system designed to handle a large, real-world corpus (rather than a small demo set), and made available to a conversational interface that persists context across chats and can assist in drafting appeals and similar tax documents. A two-way file sync sits alongside the core platform, keeping the client's documents backed up and in sync with the system.

Technical Decisions#

The corpus scaled to roughly 90GB of files, which put real strain on the original infrastructure. Handling that required migrating to more scalable cloud infrastructure, alongside adopting a hierarchical RAG architecture to improve retrieval quality once the corpus was large enough that flat retrieval stopped being sufficient.

Engineering Challenges#

The hardest problem wasn't RAG quality itself — it was file management at real scale, before RAG questions even came into play. Handling roughly 90GB of client files put real strain on the original infrastructure, which required migrating to more scalable cloud infrastructure. Improving retrieval with a hierarchical RAG architecture came after that infrastructure problem was addressed, as the next layer of the same underlying challenge: making a genuinely large, real-world document corpus both storable and retrievable.

Lessons Learned#

The clearest lesson was underestimating real-world data scale early on. The corpus size, and the strain it put on the original infrastructure, wasn't something the initial setup was sized for — and that gap only became visible once real client data started accumulating, not during initial design. Sizing infrastructure for the data volume a real client will actually generate, not a demo-scale assumption, is the takeaway that carries forward to future projects.

Tech Stack#

Docling for document ingestion, a hierarchical RAG architecture for retrieval over the corpus, and cloud infrastructure sized to the client's real data volume, alongside a two-way file sync for backup. Further implementation detail is not shared here, consistent with this project's proprietary status.

Related Content#

AIPathshala — another AILACDS professional project, also built on Docling-based ingestion and retrieval-augmented generation, for a different domain (education rather than tax consulting).