Back office ops · Production

How Infosys built a generative AI solution to process oil and gas drilling data with Amazon Bedrock

The problem

Oil and gas operations generate vast amounts of complex multimodal technical documents — well completion reports, drilling logs, and lithology diagrams — that conventional non-AI processing methods fail to handle due to specialized terminology, interconnected data relationships, and mixed text and image formats, resulting in inefficient data extraction and time-consuming manual processing.

First attempt

Three iterative RAG approaches were tried before the final design: the initial image-analysis approach worked for text but failed on image-related queries; ColBERT multi-vector embeddings proved difficult to store and manage; and fixed-size chunking improved keyword retrieval but produced fragmented long-form answers by splitting related information across chunks.

Workflow diagram · grounded in source
1
Technical document ingestion
Trigger
Well completion reports, drilling logs, and lithology diagrams are ingested into the system for processing.
▾ source quote
“well completion reports, drilling logs, and intricate lithology diagrams, contain crucial information that drives operational decisions and strategic planning”
2
Multimodal image analysis
Ai action
Amazon Nova Pro preprocesses over a thousand technical images using an iterative prompting strategy to generate comprehensive textual descriptions.
▾ source quote
“Over a thousand such technical images (including lithology diagrams, well completion charts, and drilling visualizations) were preprocessed using Amazon Nova Pro, a multimodal language model. An iterative prompting strategy was employed to generate comprehensive descriptions: Initial image analysis to extract …”
3
Hierarchical chunking and embedding
Ai action
Text content is split using a parent-child chunking hierarchy and embedded with Cohere English embeddings.
▾ source quote
“The parent-child chunking hierarchy was retained with parent chunks of 1,200 tokens and child chunks of 512 tokens. We continued using Cohere English embeddings”
4
OpenSearch vector storage
Integration
Processed text chunks and image descriptions are stored in an Amazon OpenSearch Serverless vector database.
▾ source quote
“The processed content (chunked text and complete image descriptions) was ingested into an OpenSearch Serverless vector database”
5
Hybrid search retrieval
Ai action
The system combines semantic vector search with traditional keyword-based search to retrieve the most relevant documents.
▾ source quote
“This approach combined the strengths of semantic search and traditional keyword-based search”
6
BGE result reranking
Ai action
A BGE reranker refines search results by reordering retrieved documents based on semantic similarity to the query.
▾ source quote
“implemented a BGE reranker to refine search results”
7
Domain-specific response generation
Output
Amazon Nova Pro generates domain-specific responses from the reranked retrieved content.
▾ source quote
“Amazon Nova model for domain-specific response generation”
Reported outcome

The final hybrid RAG solution achieved 92% retrieval accuracy against a human expert baseline, under 2-second average query response time, a 4.7/5 user satisfaction rating from field engineers and geologists, a 40–50% decrease in manual document processing costs, and field engineers spending 60% less time searching for technical information.

Reported metrics
Average query response timeLess than 2 seconds
Retrieval accuracy92%
User satisfaction rating4.7/5
Manual document processing cost reduction40–50%
Show all 6 reported metrics
average query response timeLess than 2 seconds
retrieval accuracy92%
user satisfaction rating4.7/5
manual document processing cost reduction40–50%
engineer time searching for technical information60%
decision-making timeSignificant reduction in decision-making time
Reported stack
Amazon BedrockAmazon Bedrock Nova ProAmazon Bedrock Knowledge BasesAmazon OpenSearch ServerlessAmazon Titan Text EmbeddingsCohere Embed English modelBGE RerankerAmazon Q DeveloperPyMuPDFOpenCVInfosys TopazColBERT
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://aws.amazon.com/blogs/machine-learning/how-infosys-built-a-generative-ai-solution-to-process-oil-and-gas-drilling-data-with-amazon-bedrock?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The final hybrid RAG solution achieved 92% retrieval accuracy against a human expert baseline, under 2-second average query response time, a 4.7/5 user satisfaction rating from field engineers and geologists, a 40–50%…

What tools did this team use?

Amazon Bedrock, Amazon Bedrock Nova Pro, Amazon Bedrock Knowledge Bases, Amazon OpenSearch Serverless, Amazon Titan Text Embeddings, Cohere Embed English model, BGE Reranker, Amazon Q Developer, PyMuPDF, OpenCV.

What results were reported?

Average query response time: Less than 2 seconds; Retrieval accuracy: 92%; User satisfaction rating: 4.7/5; Manual document processing cost reduction: 40–50% (source-reported, not independently verified).

What failed first in this deployment?

Three iterative RAG approaches were tried before the final design: the initial image-analysis approach worked for text but failed on image-related queries; ColBERT multi-vector embeddings proved difficult to store and…

How is this back office ops AI workflow structured?

Technical document ingestion → Multimodal image analysis → Hierarchical chunking and embedding → OpenSearch vector storage → Hybrid search retrieval → BGE result reranking → Domain-specific response generation.

Related back office ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Back office ops cases →