LlamaIndex and LlamaParse: document agent workflows

Tuana Çelik explains how LlamaIndex developed from an open source Python framework for retrieval-augmented generation into tooling for document-focused agents. The talk contrasts LLM-selected tools with structured workflows that restrict which steps an AI agent can take.

The workflow explanation covers events, start and stop events, and the step decorator. Steps can branch or loop. Çelik then introduces LlamaParse for processing complex documents, where tables, charts and page layout carry context that plain text extraction can lose. A chart-parsing example returns chart data as a table item for a downstream model. Some slide examples are explicitly pseudocode.

The Agent Builder demo splits a resume book into individual resumes and extracts structured fields. Viewers see extraction schemas, split categories and generated Python workflow code. Çelik describes deployment through an editable GitHub repository, with merges to main triggering redeployment. The builder was in beta at the time of the talk, and she acknowledges occasional instability.

The Q&A clarifies that the resume demo follows a flow Çelik defined, delegating only extraction and splitting decisions. She recommends choosing the degree of model autonomy according to the stakes. The deployment examples use the Llama Agents platform; the talk does not establish an offline or self-hosted setup.