Live App →

Wealth Understanding

Natural language access to portfolio context, market data, and client financial narratives powered by RAG and agentic orchestration.

Wealth Understanding

Scope — Transform static portfolio data and market information into actionable, conversational insights that wealth managers, clients, and operations teams can query in plain language.

Executive Summary

Wealth Understanding bridges the gap between raw financial data and human decision-making. By combining session-scoped RAG over ingested documents with live web search and agentic reasoning, the platform enables plain-language queries like:

“What is my exposure to mid-cap funds across all family accounts?” “Summarize the performance of my AIF allocations versus Nifty 50.” “Which clients have concentration risk above 20% in a single sector?”

The capability is delivered jointly by the Zen Chatbot Platform, Agentic Backend, and Knowledge Base (Ingestion API), with all LLM calls routed through the LLM Orchestrator.

The Problem

Wealth managers spend 30-40% of their time manually synthesizing information across:

  • Disparate portfolio statements (CAS, AIF, PMS)
  • Market news and macroeconomic indicators
  • Client meeting notes and CRM records
  • Regulatory circulars (SEBI, RBI)

Existing BI tools require predefined dashboards. Static reports cannot answer ad-hoc contextual questions. Wealth Understanding replaces this friction with conversational AI grounded in real data.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                     Wealth Understanding                         │
├─────────────────────────────────────────────────────────────────┤
│  User Query                                                      │
│       │                                                          │
│       ▼                                                          │
│  ┌─────────────┐    ┌──────────────┐    ┌──────────────────┐   │
│  │   Zen Chat  │───▶│   Agentic    │───▶│   KB / OpenSearch │   │
│  │  (Session)  │    │  Supervisor  │    │  (Hybrid Search)  │   │
│  └─────────────┘    │  (Intent +   │    └──────────────────┘   │
│                     │   Routing)   │              ▲             │
│                     └──────────────┘              │             │
│                            │                      │             │
│                            ▼                      │             │
│                     ┌──────────────┐              │             │
│                     │  Web Search  │              │             │
│                     │  (Serper/DDG)│              │             │
│                     └──────────────┘              │             │
│                                                   │             │
│                     ┌─────────────────────────────┘             │
│                     ▼                                            │
│              ┌─────────────────────┐                            │
│              │  LLM Orchestrator   │                            │
│              │  (Claude / GPT-4)   │                            │
│              └─────────────────────┘                            │
└─────────────────────────────────────────────────────────────────┘

Knowledge Sources

Source Type Refresh
Ingested Documents Vector + BM25 (OpenSearch) On upload
Session Files Scoped index per chat session Real-time
Web Search Live SERP + snippet extraction Per query
Market Data yfinance / internal feeds Daily
SEBI/RBI Circulars Ingested regulatory corpus Weekly

Intent Classification (Agentic Supervisor)

The supervisor node (GPT-4.1-nano) classifies every query into:

Intent Action
document_qa Retrieve from session-scoped RAG
general_chat Web search + general reasoning
portfolio_review Trigger portfolio analytics engine
visualization Generate charts / allocation diagrams
vision Process uploaded image with vision model

Personas & Journeys

Relationship Manager

  1. Opens Zen chat during client pre-call preparation
  2. Uploads client’s latest CAS statement
  3. Asks: “What changed in equity allocation since last quarter?”
  4. Receives a synthesized answer with percentage deltas and fund names
  5. Exports the chat thread as a PDF for client meeting notes

Client (Self-Service)

  1. Logs into white-labeled portal (B2C/B2B2C)
  2. Asks: “Explain my biggest loss this month in simple terms”
  3. Gets a guardrail-protected, jargon-free explanation
  4. Follows up: “Should I rebalance?” → Routed to portfolio review agent

Compliance Officer

  1. Queries: “Show me all clients with >₹50L exposure to unlisted securities”
  2. System ingests latest portfolio data and runs cross-client aggregation
  3. Returns a structured table with client names, amounts, and document links
  4. Exports to Excel for audit trail

Key Features

Feature Detail
Session-Scoped RAG Documents uploaded to a chat session are indexed in real-time and scoped to that session only
Hybrid Search Combines BM25 keyword search with KNN semantic search (Reciprocal Rank Fusion)
Multi-Modal Supports image upload + vision model analysis (e.g., screenshot of a portfolio table)
Web Search Augmentation Live market data and news augmentation for time-sensitive queries
Guardrails SEBI disclaimer injection, PII masking, and prohibited-advice filtering
Citations Three-tier citation model (T1: SEBI/RBI, T2: Financial Media, T3: General)
Streaming Server-Sent Events (SSE) for real-time token delivery

API Surface

Proxied through Studio Middleware at /api/v1/zen/* and /api/v1/agents/*.

Method Endpoint Purpose
POST /api/v1/chat/{id}/messages Send message with attachments
POST /api/v1/chat/invoke/stream SSE streaming response
POST /api/v1/chat/{id}/documents Session-scoped document upload
POST /api/v1/orchestrator/invoke Direct agentic routing
POST /api/v1/ingest/session-document Index document to session

Security, Compliance & Operations

  • Grounding — All answers cite source documents; hallucinations reduced via RAG constraint
  • Guardrails — Financial advice disclaimer automatically appended; no specific buy/sell recommendations generated
  • PII Masking — Indian identifiers (Aadhaar, PAN, UPI) masked pre-LLM, unmasked post-response
  • Session Isolation — Document indices are strictly scoped to session + user; no cross-tenant leakage
  • Audit — Full query + response + source chunks logged to PostgreSQL via Orchestrator