AI is creating a seismic shift in software development. Yet, as AI revolutionizes our approach, fundamental software development principles are more important than ever. The role of the developer is transforming, driven by market demands, executive mandates, and requests from business operators to incorporate AI into software systems.
Among the many facets of AI and the software development lifecycle (SDLC) we could explore — such as security, AI-assisted programming (copilots), AIOps, and model instruction tuning — there is another critical element to factor in: pipeline orchestration. To leverage large language models (LLMs) efficiently, effectively, and practically, we must develop a comprehensive understanding of data management, processes, and prompting techniques.
While this article isn’t exhaustive, it aims to provide a solid foundation for developing practical, AI-infused applications for both front and back-office use cases.
Key Concepts
Before diving deeper, let’s define some essential terms:
- LLMs (large language models): Advanced AI models trained on vast amounts of text data, capable of understanding and generating human-like text.
- RAG (retrieval-augmented generation): A technique that combines information retrieval with text generation to produce more accurate and contextually relevant responses.
- Hallucinations: Instances where AI models generate false or nonsensical information.
- Agents: Autonomous programs that complete tasks based on goals, often utilizing LLMs for natural language understanding and generation.
- Event-driven architecture: A software design pattern where the flow of the program is determined by events such as user actions, sensor outputs, or messages from other programs.
- Governance: The framework for managing, securing, and controlling access to data and AI systems.
- Semantic retrieval: The process of finding information based on meaning and context rather than exact keyword matches.
- Syntactic retrieval: Information retrieval based on the structure and format of data rather than its meaning.
The Foundation: Data and Processes
Data is at the core of any practical AI implementation. The ability to integrate, manage, and leverage various types of data is crucial. This includes structured data (relational databases), semi-structured data (document databases, key-value stores), and unstructured data (files, images, videos, and audio).
Integrating AI with existing systems of record is critical for businesses. This involves creating, reading, and persisting data across different storage systems. It’s also essential to expose and make accessible APIs for consuming data through front-end applications or services, connecting to back-end datastores, and creating processes that transform, filter, enric,h and move data through pipelines to achieve desired outcomes.
Vector databases play a crucial role in storing and retrieving data for LLMs, particularly for semantic search and retrieval. Additionally, exposing large data endpoints from data lakes, warehouses, and marts is necessary for AI systems to access and utilize enterprise data effectively.
Event-Driven Architecture
As AI becomes more prevalent, the scale of communication between API endpoints is growing at an astronomical rate. The current API economy is just the tip of the iceberg compared to what we can expect in an AI-infused future.
Enter AI agents: autonomous, often ephemeral programs that complete tasks based on goals, without explicit instructions on how to plan, reason, or create rules. These agents, powered by language models, will be ubiquitous, getting and putting data in place, writing code, scanning, filtering, and performing various actions. As agents proliferate, they will require massively scalable communication networks.
This is where event-driven architecture (EDA) comes into play. Evolved from the publish-subscribe (pub/sub) pattern, modern EDA offers robust capabilities for handling massive scale, guaranteeing message delivery, deduplicating messages and connecting to various endpoints. Some EDA implementations can even trace data lineage, scale to many nested topics, and offer flexible data persistence options.
EDA provides the type of communication network that Internet of Things (IoT) devices use, and AI agents communicate similarly. As agents pop in and out of existence, they will require a robust data network for communication, making EDA an ideal solution.
Governance and API Management
As agents consume data from various sources, API management becomes critical. Gateways must control access to applications’ or endpoints’ data, managing not only who or what has access but also the amount and frequency of access. This is crucial for both monetization opportunities and defending against unauthorized data access.
Validating and Controlling LLM Outputs
LLMs are known to hallucinate, producing false or nonsensical information. For most organizations not developing their own foundation models, controlling the LLM’s response to prompts can be challenging. However, you can control the data and prompts fed to the LLM, as well as how the model’s output is processed and presented to the end user.
Validating data within a retrieval-augmented generation (RAG) process allows for harmonization, canonization, and vector embedding of your corporate data before presenting it to the LLM. This approach helps ensure that when combining proprietary data with the LLM’s knowledge, you maintain control over the inputs.
Evaluating data after the LLM responds is also possible — and often advisable. Ranking agents can observe the answers given to a consumer before the information is presented, adding an additional layer of quality control.
Human Interaction and Continuous Learning
The human element remains crucial in AI ecosystems. Interfaces such as chatbots or frontend applications allow users to interact with the orchestrated systems around the LLM. By creating endpoints within processes that enable humans to validate answers, you can help the orchestrated systems learn and improve from external, human interactions.
The Critical Role of Orchestration
All of these elements — data management, process control, event-driven communication, governance, validation, and human interaction — require orchestration. An LLM on its own is not grounded; it needs a carefully orchestrated ecosystem to function effectively and reliably in real-world applications.
Orchestration involves:
1. Integration: Connecting dissimilar applications, creating abstractions, and exposing APIs for data consumption.
2. Automation: Streamlining processes and reducing manual intervention in data and AI workflows.
3. Coordination: Ensuring that different components of the AI system work together seamlessly.
4. Monitoring and optimization: Continuously assessing system performance and making necessary adjustments.
By focusing on orchestration, organizations can create AI systems that are not only powerful but also practical, reliable, and aligned with business objectives.
Conclusion
As we navigate the exciting frontier of AI in software development, it’s clear that the principles of orchestration are more critical than ever. The ability to effectively integrate, automate, and coordinate various components of AI systems will distinguish successful implementations from those that fall short.
For developers, understanding and mastering orchestration is key to harnessing the full potential of AI. By grounding our approach in solid orchestration practices, we can create AI-infused applications that are not just innovative, but also practical and valuable for businesses across industries.
As you embark on your AI journey, remember that the true power of these technologies lies not in the models themselves, but in how we orchestrate them within our existing systems and processes. With careful attention to data management, communication architectures, governance and human interaction, we can build AI systems that will transform the way we work and create value.
Find out how Boomi can help you become AI-ready and implement practical AI.
This article was originally published on The New Stack, under the title “Yes, Orchestration Is for AI, Too” — you can find it here: https://thenewstack.io/yes-orchestration-is-for-ai-too/