Seismic information evaluation is a vital part of vitality exploration, however configuring complicated processing workflows has historically been a time-consuming and error-prone problem. Halliburton’s Seismic Engine, a cloud-native software for seismic information processing, is a strong device that beforehand required handbook configuration of roughly 100 specialised instruments to create workflows. This course of was not solely time-consuming but additionally required deep experience, probably limiting the accessibility and effectivity of the software program.
To handle this problem, Halliburton partnered with the AWS Generative AI Innovation Middle to develop an AI-powered assistant for Seismic Engine. The answer makes use of Amazon Bedrock, Amazon Bedrock Information Bases, Amazon Nova, and Amazon DynamoDB to remodel complicated workflow creation into conversations. Geoscientists and information scientists can configure processing instruments by pure language interplay as an alternative of handbook configuration.
On this submit, we’ll discover how we constructed a proof-of-concept that converts pure language queries into executable seismic workflows whereas offering a question-answering functionality for Seismic Engine instruments and documentation. We’ll cowl the technical particulars of the answer, share analysis outcomes exhibiting workflow acceleration of as much as 95%, and focus on key learnings that may assist different organizations improve their complicated technical workflows with generative AI.
Our collaboration with AWS has been instrumental in accelerating subsurface interpretation workflows. By integrating Amazon Bedrock providers with Halliburton Landmark’s DS365 Seismic Engine, we had been in a position to cut back historically time‑consuming workflow‑constructing duties by an order of magnitude. This generative AI–powered workflow assistant not solely improves effectivity and accuracy but additionally makes our superior geophysical instruments extra accessible to a broader vary of customers. The scalable cloud‑native structure on AWS has enabled us to ship a seamless, conversational expertise that basically improves productiveness throughout subsurface workflows.
— Phillip Norlund, Supervisor of Subsurface Applied sciences, Halliburton Landmark
— Slim Bouchrara, Senior Product Proprietor of Subsurface R&D, Halliburton Landmark
Answer overview
Our undertaking aimed to deal with two key targets: reworking pure language queries into executable seismic workflows, and offering an clever query and reply (Q&A) system for Seismic Engine documentation. To realize this, we developed an answer utilizing Amazon Bedrock that allows geoscientists to work together with complicated seismic instruments by pure dialog.The spine of our system is a FastAPI software deployed on AWS App Runner, which handles consumer queries by a streaming interface. When a consumer submits a question, an intent router powered by Amazon Nova Lite analyzes the request to find out whether or not it’s searching for workflow technology or technical info. For Q&A requests, the system makes use of Amazon Bedrock Information Bases with Amazon OpenSearch Serverless to offer related solutions from listed documentation. For workflow requests, a technology agent utilizing Anthropic’s Claude on Amazon Bedrock creates YAML workflows by deciding on from 82 accessible Seismic Engine instruments.
To take care of context and allow multi-turn conversations, we built-in Amazon DynamoDB for chat historical past and interplay logging. The system helps streaming responses for each Q&A and workflow technology, offering speedy suggestions to customers because the system processes their requests. This structure permits complicated technical workflows to be created and modified by pure dialog, whereas sustaining the exact management required for seismic information processing. The next diagram illustrates the answer structure.
Question routing and intent classification
After the consumer’s question is supplied to the system, the Intent Router classifies the intent label of the given question by calling Amazon Nova Lite by way of the Amazon Bedrock API. The big language mannequin (LLM) is given a immediate to provide certainly one of three intent labels: “Workflow_Generation”, “QnA”, and “General_Question”.The “Workflow_Generation” label is used to route queries associated to workflow technology, together with studying/loading datasets, information processing operations, and varied requests involving manipulating particular datasets. The “QnA” intent label is used for questions associated to particular instruments, requests for pattern workflows, or questions on Seismic Engine documentation. The “General_Question” label is reserved for queries unrelated to Seismic Engine operations or workflows.In our implementation, Amazon Nova Lite carried out the routing job effectively, providing a very good steadiness between accuracy and latency.
Query answering implementation
The Q&A element handles Seismic Engine-related queries by utilizing Amazon Bedrock Information Bases, a totally managed service for end-to-end Retrieval Augmented Technology (RAG) workflow. We selected Bedrock Information Bases as a result of it alleviates the operational overhead of managing vector databases, chunking methods, and embedding pipelines. As a totally managed service, it handles infrastructure scaling, safety, and upkeep robotically, in order that our group may give attention to answer growth relatively than RAG infrastructure operations. The service gives native assist for a number of chunking methods together with hierarchical chunking, which maintains parent-child relationships to steadiness granular retrieval with broader doc context.The information sources embrace device documentation markdown information and Seismic Engine manuals saved in S3. We stored device documentation information unchunked since they’re comparatively brief, preserving full context for particular person instruments. For longer paperwork like Seismic Engine manuals, we used hierarchical chunking with default settings. We use Amazon Titan Textual content Embeddings V2 for embedding technology and OpenSearch Serverless because the vector database. The system additionally shops metadata similar to file names, URLs, and doc varieties for every listed merchandise for downstream use.For each retrieval and response technology, we use Amazon Bedrock Information Bases’ retrieve_and_generate API with Claude 3.5 Haiku because the mannequin. The system helps multi-turn conversations by sustaining session context, and responses are formatted with inline citations for enhanced traceability.
Observe: This answer was developed and evaluated utilizing Claude 3.5 Sonnet V2 and Claude 3.5 Haiku. Since then, these fashions have been succeeded by Claude Sonnet 4.5 and most just lately Claude Sonnet 4.6, in addition to Claude Haiku 4.5, all accessible by Amazon Bedrock. The answer structure helps mannequin upgrades with out code adjustments, so as to use the newest mannequin capabilities.
This strategy permits our system to offer context-aware, related solutions to consumer queries about Seismic Engine instruments and workflows.
Workflow technology
For queries labeled as “Workflow_Generation”, our answer makes use of LLM brokers to transform pure language into executable YAML workflows. The agent is sure with 82 instruments accessible on Seismic Engine. Based mostly on the consumer’s question and power specs that outline inputs, parameters, and outputs, the agent selects acceptable instruments, determines their right execution order, and generates a YAML workflow that addresses the consumer’s necessities. The next determine illustrates the workflow technology course of.
We used each Claude 3.5 Sonnet V2 and Claude 3.5 Haiku in our implementation, orchestrated by the LangChain framework for agent administration and power binding. The fashions are supplied with detailed device descriptions and specs, in order that they’ll perceive every device’s capabilities and necessities. When producing workflows, the system considers not solely the specific necessities within the consumer’s question but additionally contains mandatory default parameters when particular values aren’t supplied.The workflow technology course of helps multi-turn conversations, in order that customers can modify beforehand generated workflows by pure language requests. By utilizing dialog historical past saved in Amazon DynamoDB, the LLM can both generate new workflows or modify current ones in response to the consumer’s present question.
Analysis
To guage our answer’s effectiveness, we created a complete check dataset of query-workflow pairs, consisting of each low and medium complexity workflows. These had been derived from actual historic workflows and validated by subject material consultants to confirm they precisely characterize typical consumer requests.
Workflow technology outcomes
| Mannequin | Complexity | Success Price | Imply Technology Time (s) | Median Technology Time (s) |
| Claude Haiku 3.5 | easy | 84% | 8.3 | 5.9 |
| medium | 90% | 12.4 | 9.1 | |
| Claude Sonnet 3.5 V2 | easy | 86% | 11.2 | 11.5 |
| medium | 97% | 15.8 | 16.6 |
Each fashions demonstrated sturdy efficiency, with Claude Sonnet 3.5 V2 exhibiting superior success charges, significantly for medium complexity workflows. The system delivers responses by streaming, offering customers with speedy suggestions because the workflow is generated, with full workflows delivered inside 5.9-16.6 seconds. Claude Haiku 3.5 affords sooner technology instances, offering a trade-off choice between pace and accuracy.
Comparability to baseline efficiency
| Consumer Kind | % Success | % Failure | Time to Construct Easy Circulate (min) | Time to Construct Advanced Circulate (min) |
| New Consumer | 70% | 20% | 4 | 20 |
| Skilled Consumer | 85% | 10% | 2 | 5 |
| Our Answer | 84-97% | 3-16% | 0.13-0.26 | 0.21-0.28 |
Our generative AI answer demonstrates the next enhancements:
- Success charges of 84-97% surpass each new and skilled customers.
- Workflow creation time is decreased from minutes to seconds, representing over a 95% time discount.
These outcomes display that customers throughout expertise ranges can improve productiveness by over 95%, whereas sustaining or exceeding the accuracy of handbook workflow creation.
Conclusion
On this submit, we confirmed how we used Amazon Bedrock to remodel complicated technical processes into pure conversations. By implementing an AI-powered assistant with built-in Q&A capabilities, we achieved workflow technology success charges of 84-97% whereas lowering creation time by over 95% in comparison with handbook processes. The system’s capacity to deal with each low and medium complexity workflows, mixed with its contextual understanding of Seismic Engine instruments, demonstrates how generative AI can enhance industrial software program usability with out compromising accuracy.
This strategy additionally generalizes properly to different domains with complicated, multi-step agentic workflows requiring specialised device information and configuration. As subsequent steps, think about exploring multi-agent architectures utilizing frameworks like Strands Brokers SDK with Amazon Bedrock AgentCore for improved accuracy by specialised sub-agents.
Concerning the authors
