What challenges did Deltek face before AI implementation?

Answer questions accurately across multiple temporally related RFP and RFI documents Avoid stale answers when document revisions change over time Handle long PDFs and table-heavy solicitation documents in retrieval and generation

What was the impact of the AI solution?

AWS reported that Deltek’s subject matter experts evaluated the LLM responses at 96% overall accuracy The solution improved Q&A over interlinked document iterations by incorporating release-date metadata Deltek is continuing to refine the approach for broader file support and more cost-effective ingestion

Source

Evidence: Low35/100

Deltek: RAG Q&A on government solicitation documents using Amazon Bedrock + Amazon Textract + Amazon OpenSearch Service

Use case typeRAG infrastructureUpdated Jun 13, 2026

Deltek collaborated with the AWS Generative AI Innovation Center on a RAG-based solution for question answering across single and multiple government solicitation documents. The solution processes PDF documents with Amazon Textract to extract text and tables, converts tables to CSV, chunks document sections, generates embeddings with Amazon Titan Embeddings G1 – Text v1.2 on Amazon Bedrock, and indexes content plus metadata in Amazon OpenSearch Service. At query time, the system retrieves relevant chunks with semantic search, enriches prompts with metadata such as release date, and uses Anthropic Claude v2 on Amazon Bedrock to generate answers.

Organization: Deltek
Industry: Public Sector
Location: United States
Published: August 2024

Reported outcomes

96%

accuracyQuality & accuracy

Strategic outcomes

New product / capabilityEnabled RAG-based solicitation document Q&ANew product / capabilityHandled text and tables in long PDFs

Why do we believe this?Outcome claims, sources, and evidence checks

Normalized claim

Accuracy: 96%

AWS Machine Learning BlogAug 9, 2024Blog postInferred claimLow evidence strength

AWS reported that Deltek’s subject matter experts evaluated the LLM responses at 96% overall accuracy

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: Deltek
Provider: AWS
Maturity: Unknown
Linked source: AWS Machine Learning Blog

No explicit deployment-stage evidence found.

Customer identity supportedSource describes one deploymentMaturity evidence evaluated

Primary read

Use case focus

Showing 3 of 3

1Retrieval Augmented Generation (RAG)
2Document Question Answering
3Intelligent Document Processing

Answer questions accurately across multiple temporally related RFP and RFI documents
Avoid stale answers when document revisions change over time
Handle long PDFs and table-heavy solicitation documents in retrieval and generation

Amazon Textract extracts text and tables from uploaded documents
A parser converts extracted tables to CSV for better model comprehension
Document sections are chunked independently with overlap and embedded using Amazon Titan Embeddings G1 – Text v1.2
Chunks and metadata such as section name and release date are indexed in Amazon OpenSearch Service
User questions are semantically matched to chunks in OpenSearch and the retrieved context is passed to Anthropic Claude v2 on Amazon Bedrock for response generation
The prompt uses release-date ordering and chain-of-thought guidance to prioritize the latest information

Technologies

Amazon Bedrock Anthropic Claude v2 Amazon Textract Amazon OpenSearch Service Amazon Titan Embeddings G1 – Text v1.2

AWS reported that Deltek’s subject matter experts evaluated the LLM responses at 96% overall accuracy
The solution improved Q&A over interlinked document iterations by incorporating release-date metadata
Deltek is continuing to refine the approach for broader file support and more cost-effective ingestion

Architecture

Users upload solicitation documents. Amazon Textract extracts text and tables, with tables converted to CSV. The system splits documents into sections and chunks them with overlap, generates embeddings with Amazon Titan Embeddings G1 – Text v1.2, and stores embeddings plus metadata in Amazon OpenSearch Service. At query time, semantic search retrieves relevant chunks, and the prompt plus retrieved context are sent to Anthropic Claude v2 on Amazon Bedrock. Release-date metadata is used to favor the most current information across related documents.

Implementation partners1

AWS Generative AI Innovation Center

Sources & evidence1

Evidence: Low35/100Evidence strength