Finance ops · Production

Nippon India Mutual Fund improves AI assistant accuracy using advanced RAG methods on Amazon Bedrock

The problem

Nippon India Mutual Fund needed an AI assistant that could accurately retrieve information from a large volume of enterprise documents, but naive RAG fell short: similarity-based retrieval missed relevant chunks as document volumes grew, complex structures like nested tables were not parsed correctly, compound questions degraded accuracy, and models produced hallucinated responses.

First attempt

The naive RAG approach had documented limitations at scale: retrieval accuracy degraded with document volume, complex document structures were not parsed correctly, compound questions were handled poorly, and the approach was prone to hallucinations.

Workflow diagram · grounded in source
1
Custom document parsing
ai_action
“Nippon uses custom parsing to extract the relevant details from complex elements like tables, graphs, and images. They use custom chunking to create the chunks.”
2
Embedding and vector storage
integration
“They use an embedding model from Amazon Bedrock to convert the content chunks to vector embeddings and store them in the vector database.”
3
User query embedding
ai_action
“The user's query is converted into embeddings by the Amazon Bedrock embedding model.”
4
Multi-query RAG with reranking
ai_action
“Nippon uses multi-query RAG by creating multiple variants of the query and executing them in parallel. The obtained results are reranked using a reranker model for higher accuracy.”
5
Prompt augmentation with citations
ai_action
“The prompt is augmented with the references from the source documents to create citations”
6
Final response delivery
output
“Amazon Bedrock creates and sends the final response to the user.”
Reported outcome

After implementing advanced RAG methods, Nippon saw accuracy increase by more than 95%, hallucination reduced by 90–95%, and report generation time drop from 2 days to approximately 10 minutes, with source citations added to improve user confidence.

Reported metrics
Accuracy improvementmore than 95%
Hallucination reduction90–95%
Report generation timereduced from 2 days to approximately 10 minutes
User confidence in responseimproves the user confidence in the response
Reported stack
Amazon BedrockAmazon Bedrock Knowledge BasesAmazon TextractAnthropic's Claude3 Sonnet on Amazon BedrockAmazon Bedrock reranker models
Source
https://aws.amazon.com/blogs/machine-learning/how-nippon-india-mutual-fund-improved-the-accuracy-of-ai-assistant-responses-using-advanced-rag-methods-on-amazon-bedrock?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

After implementing advanced RAG methods, Nippon saw accuracy increase by more than 95%, hallucination reduced by 90–95%, and report generation time drop from 2 days to approximately 10 minutes, with source citations a…

What tools did this team use?

Amazon Bedrock, Amazon Bedrock Knowledge Bases, Amazon Textract, Anthropic's Claude3 Sonnet on Amazon Bedrock, Amazon Bedrock reranker models.

What results were reported?

Accuracy improvement: more than 95%; Hallucination reduction: 90–95%; Report generation time: reduced from 2 days to approximately 10 minutes; User confidence in response: improves the user confidence in the response (source-reported, not independently verified).

What failed first in this deployment?

The naive RAG approach had documented limitations at scale: retrieval accuracy degraded with document volume, complex document structures were not parsed correctly, compound questions were handled poorly, and the appr…

How is this finance ops AI workflow structured?

Custom document parsing → Embedding and vector storage → User query embedding → Multi-query RAG with reranking → Prompt augmentation with citations → Final response delivery.

Related finance ops cases