LlamaIndex Launches Legal Knowledge Base App with Agentic Retrieval

LlamaIndex has launched legal-kb, a public reference application available on GitHub, designed as a knowledge base for legal documents. This project utilizes LlamaIndex Index v2, also known as the LlamaParse Platform, to showcase a pattern the team refers to as a Retrieval Harness for agentic retrieval. Unlike traditional single-shot retrieval methods that perform one embedding search per query, this approach equips an agent with filesystem-style tools. These tools enable the agent to navigate and query a large, dynamic knowledge base to accomplish specific tasks. The tools provided mirror common engineering operations such as semantic and keyword search, regular expression (regex) grep, file searching, and file reading.

The legal-kb application functions as a working TanStack Start web app. Users can sign in, establish projects, upload files, and interact with an agent through chat. Each project is mirrored as a managed LlamaCloud Index v2, with uploaded files automatically parsed and indexed in the background. The chat agent then queries this live index during each interaction turn. The Retrieval Harness establishes a persistent data pipeline over the documents, connecting to a data source, indexing it, and ensuring it remains updated. This pipeline exposes a set of tools to the agent, deliberately designed to resemble filesystem operations. These generic tools allow users to integrate the harness with their own custom agents.

The agent within legal-kb is equipped with four specific tools, each mapped to a corresponding Index v2 retrieval API. The 'retrieve' tool performs a hybrid semantic search, allowing for optional reranking and returning document chunks along with citations. The 'findFiles' tool facilitates searching for files by exact name or substring, with automatic pagination. The 'readFile' tool enables the retrieval of raw file content based on a file ID, with options for specifying offset and maximum length. The fourth tool, 'grep', allows for regex-based searching within files, further enhancing the agent's ability to extract specific information from the legal documents.