The popular AI orchestration frameworkLlamaIndex has introduced the Agent Document Workflow (ADW), a new architecture that the company claims surpasses the Retrieval-Augmented Generation (RAG) process and enhances the efficiency of agents.
As orchestration frameworks continue to improve, this approach provides organizations with options to enhance agent decision-making capabilities.
LlamaIndex states that ADW can assist agents in managing “complex workflows that go beyond simple extraction or matching.”
Some agent frameworks are based on RAG systems that provide agents with the information necessary to complete tasks. However, this approach does not allow agents to make decisions based on that information.
LlamaIndex provides some real-world examples to illustrate how ADW works. For instance, in contract review, human analysts must extract key information, cross-reference regulatory requirements, identify potential risks, and make recommendations. When deployed in this workflow, AI agents ideally follow the same pattern and make decisions based on the knowledge they have gained from the documents read for contract review and other documents.
LlamaIndex stated in a blog post, “ADW addresses these challenges by viewing documents as part of a broader business process.” “ADW systems can maintain state across steps, apply business rules, coordinate different components, and take action based on the content of documents—rather than just analyzing them.”
LlamaIndex previously stated that while RAG is an important technology, it is still in its infancy, particularly for enterprises seeking to leverage AI for more robust decision-making capabilities.
Understanding Decision Context
LlamaIndex has developed a reference architecture that combines its LlamaCloud parsing capabilities with agents. It “builds systems that can understand context, maintain state, and drive multi-step processes.”
To this end, each workflow has a document that acts as a coordinator. It can guide agents to use LlamaParse to extract information from data, maintain the context and state of the document process, and then retrieve references from another knowledge base. From here, agents can begin to generate suggestions for contract review use cases or other actionable decisions for different use cases.
The company states, “By maintaining state throughout the process, agents can handle complex multi-step workflows, not just simple extraction or matching. This approach enables them to build a deep context about the documents they are working with while coordinating different system components.”
Different Agent Frameworks
Agent orchestration is an emerging field, with many organizations still exploring how agents (or multiple agents) can work for them. As agents transition from single systems to multi-agent ecosystems, orchestrating AI agents and applications may become a hot topic this year.
AI agents are an extension of RAG capabilities, which are based on the ability to look up information using enterprise knowledge.
However, as more enterprises begin to deploy AI agents, they also expect them to perform many of the tasks that human employees do. Moreover, for these more complex use cases, ordinary RAG is not sufficient. One of the advanced approaches enterprises are considering isagentic RAG, which expands the knowledge base of agents. Before reaching a conclusion, the model can decide whether to look for more information, which tools to use to obtain that information, and whether the context just acquired is relevant.