AI-Powered Workflow Assistant for Seismic Data Processing

halliburton · Energy · 2026 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Seismic workflow creation

Halliburton needed to make its Seismic Engine workflow builder accessible through natural language without losing the precision required to configure specialized geophysical processing tools.

Constraints:

manual workflow creation previously required configuring approximately 100 specialized tools
workflow generation had to preserve correct tool selection, ordering, parameters, and defaults
the assistant needed to handle both workflow requests and documentation Q&A
streaming responses were needed so users received immediate feedback
the system had to support model upgrades as Bedrock model versions changed

Architecture

Halliburton and AWS built a FastAPI service on AWS App Runner that routes each natural-language query through an Amazon Nova Lite intent router. Q&A requests go to Amazon Bedrock Knowledge Bases backed by S3 documentation and OpenSearch Serverless, while workflow-generation requests go to Claude 3.5 models on Amazon Bedrock orchestrated through LangChain and bound to 82 Seismic Engine tools. DynamoDB stores chat history and interaction logs so users can modify previous workflows over multiple turns, and streaming responses return Q&A or generated YAML workflows as they are produced.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  user-streaming-interface["user-streaming-interface"] --> fastapi-app-runner-service["fastapi-app-runner-service"]
  fastapi-app-runner-service["fastapi-app-runner-service"] --> nova-lite-intent-router["nova-lite-intent-router"]
  fastapi-app-runner-service["fastapi-app-runner-service"] --> bedrock-knowledge-base["bedrock-knowledge-base"]
  documentation-s3-corpus["documentation-s3-corpus"] --> bedrock-knowledge-base["bedrock-knowledge-base"]
  bedrock-knowledge-base["bedrock-knowledge-base"] --> claude-haiku-answer-generator["claude-haiku-answer-generator"]
  fastapi-app-runner-service["fastapi-app-runner-service"] --> workflow-generation-agent["workflow-generation-agent"]
  seismic-engine-tool-specs["seismic-engine-tool-specs"] --> workflow-generation-agent["workflow-generation-agent"]
  dynamodb-chat-history["dynamodb-chat-history"] --> workflow-generation-agent["workflow-generation-agent"]
  workflow-generation-agent["workflow-generation-agent"] --> dynamodb-chat-history["dynamodb-chat-history"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`user-streaming-interface`	Receives natural-language seismic workflow and documentation questions and streams responses back to users.	`fastapi-app-runner-service`
`fastapi-app-runner-service`	Scalable application layer that accepts queries, invokes routing, and coordinates the Q&A and workflow pipelines.	`user-streaming-interface`, `nova-lite-intent-router`, `bedrock-knowledge-base`, `workflow-generation-agent`, `dynamodb-chat-history`
`nova-lite-intent-router`	Classifies each query into workflow generation, Q&A, or general/out-of-scope intents.	`fastapi-app-runner-service`
`bedrock-knowledge-base`	Retrieves documentation and manual context for Seismic Engine Q&A using Bedrock Knowledge Bases and OpenSearch Serverless.	`fastapi-app-runner-service`, `documentation-s3-corpus`, `claude-haiku-answer-generator`
`documentation-s3-corpus`	Stores tool documentation markdown files and Seismic Engine manuals for retrieval.	`bedrock-knowledge-base`
`claude-haiku-answer-generator`	Generates cited responses for documentation and tool questions.	`bedrock-knowledge-base`
`workflow-generation-agent`	Claude-on-Bedrock agent orchestrated through LangChain that selects tools, orders operations, and writes executable YAML workflows.	`fastapi-app-runner-service`, `seismic-engine-tool-specs`, `dynamodb-chat-history`
`seismic-engine-tool-specs`	Specifications for available Seismic Engine tools, parameters, inputs, outputs, and defaults.	`workflow-generation-agent`
`dynamodb-chat-history`	Persists conversation history and interaction logs for multi-turn workflow modification.	`fastapi-app-runner-service`, `workflow-generation-agent`

Data flow

A geoscientist submits a natural-language request to the App Runner-hosted FastAPI application. Nova Lite classifies the request as workflow generation, Q&A, or general. Q&A queries retrieve tool and manual context from Bedrock Knowledge Bases and OpenSearch Serverless, then generate cited answers with Claude Haiku. Workflow queries pass tool specifications and conversation history to Claude on Bedrock via LangChain; the agent selects from 82 Seismic Engine tools, orders them, fills parameters/defaults, emits YAML, and streams the result back while DynamoDB records the conversation.

4 supporting quotes

“The backbone of our system is a FastAPI application deployed on AWS App Runner, which handles user queries through a streaming interface.”article, stated

“For Q&A requests, the system uses Amazon Bedrock Knowledge Bases with Amazon OpenSearch Serverless to provide relevant answers from indexed documentation.”article, stated

“For workflow requests, a generation agent using Anthropic’s Claude on Amazon Bedrock creates YAML workflows by selecting from 82 available Seismic Engine tools.”article, stated

“By using conversation history stored in Amazon DynamoDB, the LLM can either generate new workflows or modify existing ones according to the user’s current query.”article, stated

Why this architecture

The architecture splits routing, Q&A, and workflow generation because the tasks have different latency, accuracy, and context needs. Nova Lite is sufficient for fast intent classification, Bedrock Knowledge Bases reduces RAG infrastructure work for documentation Q&A, and Claude plus LangChain handles the higher-complexity tool-binding and YAML-generation path. DynamoDB history is required because users can revise previous workflows over multiple conversational turns, and the Bedrock abstraction lets the team upgrade models without code changes.

4 supporting quotes

“Amazon Nova Lite performed the routing task efficiently, offering a good balance between accuracy and latency.”article, stated

“We chose Bedrock Knowledge Bases because it alleviates the operational overhead of managing vector databases, chunking strategies, and embedding pipelines.”article, stated

“The models are provided with detailed tool descriptions and specifications, so that they can understand each tool’s capabilities and requirements.”article, stated

“The solution architecture supports model upgrades without code changes, so that you can use the latest model capabilities.”article, stated

Models used: Amazon Nova Lite · Claude 3.5 Sonnet V2 · Claude 3.5 Haiku · Amazon Titan Text Embeddings V2 · Amazon Bedrock · LangChain

Scale

Manual workflow creation involved approximately 100 specialized seismic tools.
“previously required manual configuration of approximately 100 specialized tools to create workflows.”article, stated
The generation agent selected from 82 available Seismic Engine tools.
“creates YAML workflows by selecting from 82 available Seismic Engine tools.”article, stated
Generated workflows completed in 5.9 to 16.6 seconds in the evaluation.
“complete workflows delivered within 5.9-16.6 seconds.”article, stated
The system reduced workflow creation time by more than 95%.
“Workflow creation time is reduced from minutes to seconds, representing over a 95% time reduction.”article, stated

Alternatives considered (roads not taken)

Keep manual configuration as the primary workflow-building path

Rejected because: Manual setup was slow, expertise-heavy, and limited accessibility for users who did not already understand the specialized seismic tools.

“This process was not only time-consuming but also required deep expertise, potentially limiting the accessibility and efficiency of the software.”article, stated

Use one model and one prompt path for all assistant requests

Rejected because: The system instead separates intent routing, Q&A, and workflow generation so simple routing and documentation questions do not need the same pipeline as tool-bound YAML generation.

“The large language model (LLM) is given a prompt to produce one of three intent labels: “Workflow_Generation”, “QnA”, and “General_Question”.The “Workflow_Generation” label is used to route queries related to workflow generation, including reading/loading datasets, data processing operations, and”article, stated

Operate a custom RAG stack for documentation retrieval

Rejected because: The team chose Bedrock Knowledge Bases to avoid owning vector database, chunking, embedding, scaling, security, and maintenance concerns.

“As a fully managed service, it handles infrastructure scaling, security, and maintenance automatically, so that our team could focus on solution development”article, stated

Patterns

`model-routing` (orchestration)

A lightweight Nova Lite router chooses between workflow generation, documentation Q&A, and out-of-scope handling before invoking heavier downstream paths.

“the Intent Router classifies the intent label of the given query by calling Amazon Nova Lite via the Amazon Bedrock API.”article, stated

`rag-retrieval` (retrieval)

Documentation questions are answered through Bedrock Knowledge Bases over tool docs and manuals rather than relying only on model memory.

“The data sources include tool documentation markdown files and Seismic Engine manuals stored in S3.”article, stated

`tool-use-orchestration` (agent-architecture)

The workflow-generation agent is bound to Seismic Engine tool specifications so it can select, order, and parameterize tool calls into executable YAML.

“the agent selects appropriate tools, determines their correct execution order, and generates a YAML workflow that addresses the user’s requirements.”article, stated

`agent-memory` (state)

DynamoDB chat history lets the assistant revise or extend previously generated workflows during multi-turn conversations.

“The workflow generation process supports multi-turn conversations, so that users can modify previously generated workflows through natural language requests.”article, stated

`eval-driven-development` (evaluation)

The team evaluated workflow generation with low- and medium-complexity query/workflow pairs derived from historical workflows and checked by experts.

“These were derived from real historical workflows and validated by subject matter experts to verify they accurately represent typical user requests.”article, stated

Tradeoffs

Route with Nova Lite before invoking specialized pipelines

Chose a lightweight intent classifier over sending every request directly to the workflow-generation or Q&A path.

Rationale: Routing gave the team a lower-latency way to select the right backend behavior while preserving more capable models for generation tasks.

“Amazon Nova Lite performed the routing task efficiently, offering a good balance between accuracy and latency.”article, stated

Use Claude Sonnet and Haiku as accuracy/speed options

Chose multiple Claude models for workflow generation and Q&A over a single fixed model for all generation.

Rationale: Sonnet produced stronger medium-complexity workflow success, while Haiku provided a faster speed/accuracy tradeoff option.

“Claude Sonnet 3.5 V2 showing superior success rates, particularly for medium complexity workflows.”article, stated

“Claude Haiku 3.5 offers faster generation times, providing a trade-off option between speed and accuracy.”article, stated

Use managed Bedrock Knowledge Bases for RAG

Chose managed retrieval infrastructure over self-managed vector database, chunking, embedding, and scaling operations.

Rationale: The team wanted to focus on the assistant and seismic workflow logic instead of operating the retrieval substrate.

“it alleviates the operational overhead of managing vector databases, chunking strategies, and embedding pipelines.”article, stated

Gotchas (what bit them)

Tool documentation retrieval needed different chunking strategies for short tool docs versus longer manuals.

Root cause: Short tool documentation benefits from complete context, while manuals need hierarchical chunks to balance granular retrieval with broader context.

Resolution: Keep tool documentation files unchunked and use hierarchical chunking with default settings for longer manuals.

“We kept tool documentation files unchunked since they’re relatively short, preserving complete context for individual tools.”article, stated

“For longer documents like Seismic Engine manuals, we used hierarchical chunking with default settings.”article, stated

Simple tool omission or wrong ordering would make generated seismic workflows unusable.

Root cause: The agent had to infer both explicit user requirements and necessary default parameters across dozens of specialized tools.

Resolution: Give the models detailed tool descriptions/specifications and evaluate against historical workflows validated by subject matter experts.

“the system considers not only the explicit requirements in the user’s query but also includes necessary default parameters when specific values aren’t provided.”article, stated

“validated by subject matter experts to verify they accurately represent typical user requests.”article, stated

Evaluation

The team tested workflow generation on a dataset of low- and medium-complexity query/workflow pairs from historical workflows validated by subject matter experts, then compared model success rates and generation times against manual new-user and experienced-user baselines.

Metrics: 84% Haiku success on simple workflows · 90% Haiku success on medium workflows · 86% Sonnet success on simple workflows · 97% Sonnet success on medium workflows · 5.9 to 16.6 second median/mean generation range reported for complete workflows · over 95% workflow creation time reduction

3 supporting quotes

“we created a comprehensive test dataset of query-workflow pairs, consisting of both low and medium complexity workflows.”article, stated

“Both models demonstrated strong performance, with Claude Sonnet 3.5 V2 showing superior success rates”article, stated

“Success rates of 84-97% surpass both new and experienced users.”article, stated

Outcomes

workflow generation success rates reached 84-97%
complete workflows were delivered within 5.9-16.6 seconds
workflow creation time dropped from minutes to seconds
users across experience levels could use natural language for Seismic Engine workflows

The proof-of-concept targeted both low- and medium-complexity Seismic Engine workflows and selected from 82 available tools, with the article reporting over 95% workflow-creation time reduction compared with manual processes.

“we achieved workflow generation success rates of 84-97% while reducing creation time by over 95% compared to manual processes.”article, stated

“The system’s ability to handle both low and medium complexity workflows”article, stated