How Narwal Helped a Leading Credit Reporting Company Reduce Knowledge Search by 99% While Building an Enterprise AI Foundation

Summary

A leading credit reporting and financial services company needed a fast, governed way to search across a massive, fragmented body of internal knowledge, and later, a reusable retrieval foundation that could support AI Advisor agents across the enterprise. Narwal delivered both. The first phase was a governed enterprise AI platform that turned unstructured internal knowledge into a fast, auditable, customer-facing FAQ assistant, cutting search time from hours to under 30 seconds while improving retrieval precision from 60% to 75%. The second phase, now underway, is extending that same retrieval philosophy into a centralized ingestion and retrieval platform on Google Cloud, giving the company a shared foundation that future AI Advisor agents can build on instead of starting from scratch

Customer Challenge

The company’s internal knowledge was extensive but scattered, spread across more than 50GB of unstructured content with no fast or governed way for teams to search it. As AI initiatives expanded, that gap widened.

• Fragmented Knowledge Access: Product and methodology information lived across disconnected sources, making it difficult for teams to find accurate answers quickly.
• Governance Requirements: Any customer-facing or internal AI assistant needed auditable, company-approved knowledge access, not open-ended search.
• Retrieval Quality Concerns: Early search approaches risked surfacing incomplete or inconsistent answers, which is a significant liability in a regulated, credit-driven business.
• Scaling Beyond a Single Use Case: As the company moved toward multiple AI Advisor agents, it needed a reusable retrieval foundation rather than a one-off build for each new use case.

Narwal’s Solution

Narwal approached this as a two-phase retrieval journey: prove the model on internal knowledge search, then scale it into a platform other AI Advisor agents could draw on.

Phase One: Enterprise Knowledge Discovery RAG Platform

Narwal built a large-scale retrieval-augmented generation system on Vertex AI to make internal knowledge searchable and governed.

• Hybrid Retrieval Architecture: Combined BM25 and FAISS-based vector search with reranking to surface relevant results consistently across a large, varied knowledge base.
• Governed FAQ Assistant: A Python and FastAPI backend, with Redis handling fast-access caching, powered the customer-facing assistant.
• Built-In Quality Control: An LLM-as-Judge evaluation framework continuously checked answer quality and flagged weak responses before they reached users.

Phase Two: AI Advisor Ingestion & Retrieval Platform (Active)

Building on what phase one proved out, Narwal is designing a centralized data-ingestion and retrieval platform on Google Cloud to give every AI Advisor agent a shared, governed source of truth.

• Centralized Data Ingestion: Aggregates the company’s product and methodology content into one enriched knowledge store, built on BigQuery, GCS, Java 17, and Spring Boot.
• Reusable API Layer: Exposes ingestion and CRUD APIs, a semantic layer, and hybrid search with reranking and MMR, so new AI Advisor agents can plug in instead of rebuilding retrieval from scratch.
• Milestone-Based Delivery: Structured across solution design, GCP landing zone and raw ETL, preprocessing and vectorization, API development, semantic-layer integration, hybrid search testing, and validation through Dev, QA, UAT, and Production.
• Scoped MVP Rollout: Targets at least 25 key product sheets alongside broader product information, with rollout targeted for late 2026.

Business Outcomes

• Faster, More Accurate Knowledge Search: Retrieval precision improved from 60% to 75%, and hallucinations dropped by approximately 40%, giving the company a safer, more auditable support experience.
• Dramatic Reduction in Search Time: Enterprise knowledge search time fell from hours to under 30 seconds.
• Governed, Customer-Facing Assistant Delivered: The company gained a company-approved knowledge assistant with built-in evaluation and quality controls.
• Reusable Retrieval Foundation Underway: Rather than building retrieval logic separately for each agent, the company is establishing a shared ingestion and retrieval platform that new AI Advisor agents can adopt directly as they launch.
• Faster Path to Future AI Initiatives: With centralized ingestion, semantic search, and citation-based observability in place, the company will be positioned to extend AI Advisor capabilities to new domains without re-solving the same retrieval problems.

Why Narwal

Building a single RAG assistant is one challenge. Designing a retrieval foundation that other AI agents can reuse at enterprise scale is another. Narwal brought both together for this engagement, first proving the retrieval approach on a real, governed use case with measurable gains in precision and speed, then generalizing it into a platform built for reuse, with the semantic modeling, hybrid search architecture, and evaluation discipline needed to keep answers accurate as the platform scales.