Generative AI With LangChain_ Build Large Language ModelApps With Python, ChatGPT, and Other LLMs
Metadata
- Author: Ben Auffarth
Highlights
MLOps aims to increase automation, improve the quality of production models, and address business and regulatory requirements. LLMOps is a specialized sub-category of MLOps. It refers to the operational capabilities and infrastructure necessary for fine-tuning and operationalizing large language models as part of a product. While it may not be drastically different from the concept of MLOps, the distinction lies in the specific requirements connected to handling, refining, and deploying massive language models like GPT-3, which houses 175 billion parameters.The term LMOps is more inclusive than LLMOps as it encompasses various types of language models, including both large language models and smaller generative models. This term acknowledges the expanding landscape of language models and their relevance in operational contexts.FOMO (Foundational Model Orchestration) specifically addresses the challenges faced when working with foundational models. It highlights the need for managing multi-step processes, integrating with external resources, and coordinating workflows involving these models.The term ModelOps focuses on the governance and lifecycle management of AI and decision models as they are deployed. Even more broadly, AgentOps involves the operational management of LLMs and other AI agents, ensuring their appropriate behavior, managing their environment and resource access, and facilitating interactions between agents while addressing concerns related to unintended outcomes and incompatible objectives.While — location: 3800
Finally, AgentOps explicitly highlights the interactive nature of agents consisting of generative models operating with certain heuristics and includes tools. The emergence of all very specialized terms underscores the rapid evolution of the field; however, their long-term prevalence is unclear. MLOps is an established term widely used in the industry, with significant recognition and adoption. Therefore, we’ll stick to MLOps for the remainder of this chapter.Before productionizing any agent or model, we should first evaluate its output, so we should start with this. We will focus on the evaluation methods provided by LangChain. — location: 3814