completed production-ai

FraudShield RAG

Retrieval-augmented fraud investigation system with live API and UI deployment on Cloud Run, grounded responses, and source citations.

Started: March 28, 2026

Completed: April 5, 2026

GitHub Repository Live Demo

Evidence & Verification

Metrics and claims on this page are tied to the linked artifacts below (repository docs, experiment outputs, and deployment pages when available).

2

Cloud Run Services

3

Api Endpoints

2

Retrieval Stages

8/8

P0 Checks

Technologies

Python FastAPI LangChain Qdrant Streamlit Docker

Overview

FraudShield RAG is a document-grounded assistant for fraud investigations. It ingests fraud documents (PDF/CSV/text), builds vector embeddings in Qdrant, and answers analyst questions with cited evidence.

Live Deployment

UI: https://fraudshield-demo-5tphgb6fsa-as.a.run.app
API docs: https://fraudshield-api-5tphgb6fsa-as.a.run.app/docs
Latest ready revisions:
- fraudshield-demo-00004-z9d
- fraudshield-api-00013-fz7
Traffic split: 100% on latest revisions

API Surface

POST /ingest for file ingestion and chunking
POST /query for retrieval + generation
GET /health for service readiness and vector backend checks

Retrieval Pipeline

The service uses a two-stage retrieval strategy:

Dense vector recall (top_k) from Qdrant.
Cross-encoder rerank for higher precision before final LLM response generation.

Default chunking is configured at 512 with 50 overlap to balance context quality and retrieval speed.

GitHub Status

Repository: fraudshield-rag
Latest main commit verified on April 5, 2026: 68a04cd