Join our daily and weekly newsletters to receive the latest updates and exclusive content on our industry-leading AI coverage. Find out more
A popular AI orchestration framework Llama Index introduced a modern Agent Document Workflow (ADW) architecture that the company says goes beyond search-assisted generation (RAG) processes and increases agent productivity.
As orchestration structures continue to improve, this method may offer organizations the opportunity to escalate agent decision-making capabilities.
LlamaIndex says ADW can facilitate agents manage “complex workflows beyond simple extraction or matching.”
Some agent structures rely on RAG systems that provide agents with the information they need to perform tasks. However, this method does not allow agents to make decisions based on this information.
LlamaIndex provided some real-world examples of ADW working well. For example, when reviewing contracts, analysts must extract key information, relate regulatory requirements, identify potential risks, and generate recommendations. AI agents deployed in this workflow would ideally follow the same pattern and make decisions based on the documents they read to review the contract and knowledge from other documents.
“ADW addresses these challenges by treating documents as part of broader business processes,” LlamaIndex said in: blog entry. “An ADW system can maintain state across stages, apply business rules, coordinate various components, and take action based on document content – not just analyzing it.”
LlamaIndex has previously stated that RAG, while an crucial technique, remains primitive, especially for enterprises seeking more strong decision-making capabilities using AI.
Understanding the decision-making context
LlamaIndex has developed reference architectures that combine LlamaCloud’s analytics capabilities with agents. “It builds systems that can understand context, maintain state, and control multi-step processes.”
To do this, each workflow has a document that acts as a coordinator. It can direct agents to apply LlamaParse to extract information from data, maintain the document’s context and process state, and then retrieve reference material from another knowledge base. From this point, agents can start generating recommendations for the contract review apply case or other actionable decisions for various apply cases.
“By maintaining state throughout the process, agents can support complex, multi-step workflows that go beyond simple extraction or matching,” the company said. “This approach allows them to build deep context around the documents being processed while coordinating the operation of various system components.”
Different agent structures
Agent orchestration is an emerging field, and many organizations are still exploring how agents – or multiple agents – work for them. Coordinating agents and AI applications may become a broader topic this year as agents move from single systems to multi-agent ecosystems.
AI agents are an extension of the RAG offer, i.e. the ability to search for information based on the company’s knowledge.
But as more enterprises begin to deploy AI agents, they want them to also perform many of the tasks that employees perform. In more complicated cases, “vanilla” RAG will not be enough. One advanced approach that enterprises are considering is agent-based RAG, which expands agents’ knowledge base. Before they get the result, models can decide whether they need to find more information, what tool to apply to get it, and whether the context they just retrieved is relevant.