Back office ops · Production

PDI Technologies builds enterprise-grade RAG system (PDIQ) for internal knowledge access on AWS

The problem

Internal teams at PDI Technologies could not efficiently access organizational knowledge scattered across disparate systems including websites, Confluence pages, SharePoint sites, and other data sources, with no unified searchable interface for employees.

Workflow diagram · grounded in source
1
Scheduled crawler trigger
trigger
“Amazon EventBridge maintains and executes the crawler scheduler”
2
Multi-source data crawling
integration
“Crawlers are configured by Administrator to collect data from a variety of sources that PDI relies on. Crawlers hydrate the data into the knowledge base so that this information can be retrieved by end users.”
3
LLM image captioning
ai_action
“PDIQ scans and generates an image caption that explains the content of the image. This caption gets injected back into the markdown file, next to the <image> tag, thereby enriching the document content.”
4
Document chunking and summarization
ai_action
“Generate document summary – Use Amazon Nova Lite to create a summary of the entire document, constrained by the 20% token allocation. This summary is reused across all chunks to provide consistent context.”
5
Vector embedding generation and storage
integration
“Generate embeddings – Use Amazon Titan Text Embeddings V2 to generate vector embeddings for each chunk (summary plus content), which is then stored in the vector store.”
6
Similarity search retrieval
ai_action
“Using similarity search, retrieves the most relevant document chunks, which include summary, chunk data, image caption, and image link.”
7
Response generation
output
“LLM generates a response based on the data retrieved and the preconfigured system prompt.”
Reported outcome

PDIQ improved accuracy approval rate from 60% to 79%, empowered support teams to resolve customer queries significantly faster, increased customer satisfaction scores (CSAT) and net promoter scores (NPS), and reduced operational costs through serverless architecture.

Reported metrics
Accuracy approval rate60% to 79%
Customer query resolution speedresolve customer queries significantly faster
customer satisfaction (CSAT and NPS)increased customer satisfaction scores (CSAT), net promoter scores (NPS), and overall loyalty
Support staff productivityallowing limited support staff to focus on expert-level cases, which improves productivity and morale
Show all 5 reported metrics
accuracy approval rate60% to 79%
customer query resolution speedresolve customer queries significantly faster
customer satisfaction (CSAT and NPS)increased customer satisfaction scores (CSAT), net promoter scores (NPS), and overall loyalty
support staff productivityallowing limited support staff to focus on expert-level cases, which improves productivity and morale
operational overhead and costminimizing operational overhead and cost
Reported stack
Amazon S3Amazon Nova MicroAmazon Nova ProAmazon Titan Text Embeddings V2Amazon CognitoConfluenceSharePointAzure DevOps
Source
https://aws.amazon.com/blogs/machine-learning/how-pdi-built-an-enterprise-grade-rag-system-for-ai-applications-with-aws?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

PDIQ improved accuracy approval rate from 60% to 79%, empowered support teams to resolve customer queries significantly faster, increased customer satisfaction scores (CSAT) and net promoter scores (NPS), and reduced…

What tools did this team use?

Amazon S3, Amazon Nova Micro, Amazon Nova Pro, Amazon Titan Text Embeddings V2, Amazon Cognito, Confluence, SharePoint, Azure DevOps.

What results were reported?

Accuracy approval rate: 60% to 79%; Customer query resolution speed: resolve customer queries significantly faster; customer satisfaction (CSAT and NPS): increased customer satisfaction scores (CSAT), net promoter scores (NPS), and overall loyalty; Support staff productivity: allowing limited support staff to focus on expert-level cases, which improves productivity and morale (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Scheduled crawler trigger → Multi-source data crawling → LLM image captioning → Document chunking and summarization → Vector embedding generation and storage → Similarity search retrieval → Response generation.

Related back office ops cases