⚙️ System Architecture

Overview

The system integrates document ingestion, vector retrieval, and LLM reasoning to produce explainable tax recommendations.

---

Pipeline Stages

1. Data Sources:
   Finance Acts, CBDT notifications, company annual reports, competitor filings.

2. Preprocessing:

   • OCR for scanned PDFs
   • Table extraction (TableNet / PdfTable)
   • Text chunking for embeddings

3. Embeddings + Vector DB:

   • Convert text to embeddings (OpenAI/Instructor models)
   • Store in a semantic vector database (e.g., Pinecone, FAISS)

4. RAG + LLM Layer:

   • Retrieve top-k relevant legal/financial segments
   • Generate grounded responses using fine-tuned LLM

5. Output:

   • Tax regime classification
   • Line-item to tax mapping
   • Risk-of-notice score
   • Competitor benchmarking report

---

Modules

Module	Function	Tools
Data Ingestion	Extract structured data from PDFs	OCR, TableNet
Embeddings Engine	Create semantic search space	Sentence Transformers
RAG Layer	Retrieve and contextualize information	LangChain / LlamaIndex
Generation	Produce human-readable tax insights	LLM (GPT/OLLAMA)
Risk Scoring	Predict notice likelihood	ML model + heuristics
UI Dashboard	Visualize insights	Streamlit

---

Diagram

flowchart TD
A[PDFs / Financial Data] --> B[OCR + Parser]
B --> C[Vector Embeddings]
C --> D[RAG Retriever]
D --> E[LLM Generator]
E --> F[Tax Insights + Risk Score]