How to Build a customer support Agent Using LlamaIndex in TypeScript for fintech

A customer support agent for fintech answers account, payments, card, and compliance questions using your internal knowledge base instead of guessing. That matters because in fintech, a bad answer is not just a bad UX issue; it can become a compliance incident, a chargeback dispute, or a trust problem.

Architecture

• •

  Ingestion layer

  • •Pulls policy docs, FAQ pages, dispute workflows, and product manuals from approved sources.
  • •Normalizes content into chunks with metadata like source, region, product, and lastReviewed.
• •

  Vector index

  • •Stores embeddings for semantic retrieval over support content.
  • •Use this for “how do I…?” questions where exact keywords are unreliable.
• •

  Retriever + reranker

  • •Retrieves the top candidate chunks and optionally reranks them by relevance.
  • •Critical for fintech because you want the answer grounded in the right policy version.
• •

  Response synthesizer

  • •Turns retrieved context into a concise support answer.
  • •Should be constrained to cite sources and avoid inventing policy details.
• •

  Escalation layer

  • •Detects high-risk intents like disputes, fraud claims, chargebacks, KYC/AML issues, and account closure.
  • •Routes to human agents or specialized workflows instead of answering directly.
• •

  Audit logging

  • •Records question, retrieved sources, model output, confidence signals, and escalation decisions.
  • •Needed for compliance review and incident reconstruction.

Implementation

1) Install the core packages

Use LlamaIndex TS with an OpenAI embedding/model setup. In production you can swap the model provider later; the retrieval pattern stays the same.

npm install llamaindex dotenv

Set your environment variables:

export OPENAI_API_KEY="your-key"

2) Load support documents and build an index

This example uses Document, VectorStoreIndex, and Settings. The key pattern is to attach metadata that helps with filtering later: region, product line, and document type.

import "dotenv/config";
import { Document, Settings, VectorStoreIndex } from "llamaindex";

async function main() {
  Settings.chunkSize = 800;
  Settings.chunkOverlap = 120;

  const docs = [
    new Document({
      text: `
        Card replacement policy:
        - Lost cards are blocked immediately.
        - Replacement cards ship in 3-5 business days.
        - Expedited shipping is available only for premium accounts.
      `,
      metadata: {
        source: "support-policy",
        product: "cards",
        region: "us",
        docType: "policy",
        lastReviewed: "2025-01-10",
      },
    }),
    new Document({
      text: `
        Chargeback timeline:
        Customers must report unauthorized card transactions within 60 days.
        We may request supporting evidence before opening a case.
      `,
      metadata: {
        source: "support-policy",
        product: "payments",
        region: "us",
        docType: "policy",
        lastReviewed: "2025-01-12",
      },
    }),
    new Document({
      text: `
        KYC review:
        If identity verification fails twice, escalate to manual review.
        Do not advise customers to submit altered documents.
      `,
      metadata: {
        source: "compliance-handbook",
        product: "onboarding",
        region: "eu",
        docType: "compliance",
        lastReviewed: "2025-01-08",
      },
    }),
  ];

  const index = await VectorStoreIndex.fromDocuments(docs);
}
main();

3) Build a retriever-driven support query flow

For fintech support you should not just call a chat model directly. Retrieve first, then answer from evidence. The code below uses asQueryEngine() so you get retrieval plus synthesis in one path.

import "dotenv/config";
import {
  Document,
  MetadataFilters,
  MetadataFilter,
  Settings,
  VectorStoreIndex,
} from "llamaindex";

async function buildAgent() {
  Settings.chunkSize = 800;
  Settings.chunkOverlap = 120;

  const docs = [
    new Document({
      text: "Customers can reset their PIN in the mobile app under Card Settings.",
      metadata: { product: "cards", region: "us", docType: "faq" },
    }),
    new Document({
      text:
        "For disputed card charges older than 60 days, route the case to manual review.",
      metadata: { product: "payments", region: "us", docType: "policy" },
    }),
    new Document({
      text:
        "If a user asks about closing an account while a balance remains open, escalate to operations.",
      metadata: { product: "accounts", region: "eu", docType: "policy" },
    }),
  ];

  const index = await VectorStoreIndex.fromDocuments(docs);

  const filters = new MetadataFilters({
    filters: [
      new MetadataFilter({ key: "region", value: "us" }),
      new MetadataFilter({ key: "product", value: "cards" }),
    ],
  });

  const queryEngine = index.asQueryEngine({
    similarityTopK: 3,
    filters,
  });

  const response = await queryEngine.query({
    query:
      'How does a customer reset their PIN? Only answer using the provided support content.',
  });

  console.log(response.toString());
}

buildAgent();

4) Add escalation rules before generating an answer

This is where most fintech teams get it wrong. You need deterministic checks for regulated topics before letting retrieval synthesize a response. Use simple intent classification rules first; keep them explicit and auditable.

import { QueryEngineTool } from "llamaindex";

const riskyPatterns = [
  /fraud/i,
  /unauthorized/i,
  /chargeback/i,
  /aml/i,
  
/// truncated?

---

## Keep learning

- [The complete AI Agents Roadmap](/blog/ai-agents-roadmap-2026) — my full 8-step breakdown
- [Free: The AI Agent Starter Kit](/starter-kit) — PDF checklist + starter code
- [Work with me](/contact) — I build AI for banks and insurance companies

*By Cyprian Aarons, AI Consultant at [Topiax](https://topiax.xyz).*