An enterprise RAG application that transforms unstructured documents into queryable, governed knowledge — enabling semantic search, conversational AI, and automated document comparison at scale.
Built as a full product (not a POC) with role-based access control, a custom web interface, and multi-step agentic workflows designed for the sensitivity requirements of enterprise finance and operations.
Enterprise AI has solved structured data well — text-to-SQL, semantic layers, dashboards. But the majority of organizational knowledge lives in unstructured documents: contracts, audit reports, policies, board materials, regulatory filings, and process documentation.
These documents are:
- Trapped in file shares and content systems — findable only if you remember the exact name or folder
- Opaque to existing AI — general-purpose chatbots can't access them, and even if they could, lack the governance controls enterprises require
- Manual to compare — identifying differences between a signed contract and a template takes hours of human review
The gap between "we have a chatbot" and "our documents are machine-readable and governed" is where this platform lives.
┌─────────────────────────────────────────────────────────┐
│ Custom Web Interface │
│ RBAC-secured · Role-aware responses │
└────────────────────────┬────────────────────────────────┘
│
┌────────────────────────▼────────────────────────────────┐
│ Orchestration Layer (LangGraph) │
│ Stateful agents · Multi-step reasoning · │
│ Workflow routing · Context management │
└──────┬───────────────────────────────┬──────────────────┘
│ │
┌──────▼──────────┐ ┌──────────▼──────────────────┐
│ RAG Engine │ │ Document Processing │
│ (LangChain) │ │ Pipeline │
│ │ │ │
│ • Retrieval │ │ • Ingestion & parsing │
│ • Generation │ │ • Structure detection │
│ • Citation │ │ • Chunking strategies │
│ • Grounding │ │ • Embedding & indexing │
└──────┬──────────┘ └──────────┬──────────────────┘
│ │
┌──────▼───────────────────────────────▼──────────────────┐
│ Vector Store + Document Store │
│ Document-level RBAC · Sensitivity classification │
│ Audit trail · Collection-based organization │
└─────────────────────────────────────────────────────────┘
| Capability | What It Does | Status |
|---|---|---|
| Document Extraction | Parses unstructured documents into structured, queryable representations | Shipped |
| Semantic Search | Retrieves relevant document sections using vector similarity rather than keyword matching | Shipped |
| Conversational Chat | Natural language Q&A over document corpora with source citation | Shipped |
| Document Generation | Creates new documents from defined templates, pre-populated with contextual data | In development |
| Document Comparison | Automated side-by-side comparison to surface discrepancies between document versions | In development |
Finance document workflows require multi-step reasoning: routing between document types, conditionally triggering different retrieval strategies, chaining extraction → comparison → summarization steps. LangGraph's stateful agent model makes these workflows explicit, debuggable, and maintainable — unlike deeply nested chain-of-calls that become opaque at scale.
Enterprise documents carry legal, financial, and regulatory sensitivity. A "bolt-on-later" approach to access control creates architectural debt that's expensive to remediate (retrofitting document-level permissions into a system designed without them). Designing RBAC into the data layer from the start ensures access control is enforced at the storage level, not just the UI level.
Embedding document intelligence in an existing tool would have constrained the UX and limited adoption. A purpose-built interface allows:
- Controlled UX design optimized for document workflows
- Direct user feedback loops for rapid iteration
- Clear product identity and discoverability
- Independence from the release cycles of host platforms
Finance teams have the highest concentration of sensitive, high-value documents combined with repetitive analytical workflows against those documents. Contract review, audit evidence gathering, policy compliance verification — these are workflows where document AI delivers measurable time savings with clear ROI justification.
| Document | What You'll Learn |
|---|---|
| System Design | Full component architecture, data flow, integration patterns |
| RAG Patterns | Chunking strategies, retrieval approaches, hallucination mitigation |
| RBAC & Governance | Enterprise access control, sensitivity classification, audit design |
| Document | What You'll Learn |
|---|---|
| Document Comparison | Multi-step comparison workflow, the hardest problem in document AI |
| Conversational Search | Citation, grounding, multi-document synthesis |
| Document | What You'll Learn |
|---|---|
| Product Decisions | Feature prioritization, UX choices, platform strategy |
| Dimension | Demo RAG | This Platform |
|---|---|---|
| Access control | None — everyone sees everything | Document-level RBAC with sensitivity classification |
| Provenance | "Here's an answer" | Cited sources with page/section references |
| Scale | Single document or small corpus | Thousands of documents across multiple collections |
| Workflows | Single Q&A turn | Multi-step agentic workflows (extract → compare → summarize) |
| Governance | None | Audit trail, data retention policies, classification enforcement |
| Reliability | Best-effort | Structured error handling, fallback strategies, confidence indicators |
| UX | Generic chat | Purpose-built interface with document-specific interactions |
| Layer | Technology |
|---|---|
| Orchestration | LangGraph (stateful multi-agent workflows) |
| RAG Framework | LangChain |
| Vector Storage | Enterprise vector database |
| Interface | Custom web application |
| Access Control | Role-based, document-level enforcement |
| Deployment | Enterprise cloud infrastructure |
| Languages | Python |
- Approaching company-wide General Availability — the broadest deployment scope in the organization's AI portfolio
- Establishes a reusable RAG foundation that other teams are building new capabilities on top of
- Reduces manual document review workflows from hours to minutes
- Extends existing enterprise GenAI capabilities (which previously covered only chat and translation) into structured document workflows
I'm Amar Chatterjee — I led this platform as Product Owner from concept through near-production delivery, driving the full product lifecycle: user research, requirements, UX/UI design, roadmap, cross-functional coordination, and pre-production validation. My background combines 14 years of financial operations expertise with data science (UC Berkeley M.S.) and hands-on AI architecture.
This repo documents the architecture and product decisions behind an enterprise Document Intelligence platform. Implementation details have been abstracted to respect organizational confidentiality.
MIT — see LICENSE for details.