CocoIndex: fresh agent context and code search demo

CocoIndex co-founder Linghua Jin explains how fresh data views can support agents that work across hours or days. Her approach is to declare the desired output from source data and let the framework manage updates. She compares this abstraction to React: developers describe what an agent should see, while the engine handles the work needed to keep that view current.

The discussion covers lineage tracking, change propagation and stale data cleanup. Jin says these mechanisms let CocoIndex recompute the affected parts of a pipeline rather than repeatedly process the whole corpus. She describes a Rust core with a Python SDK, discusses a Rust SDK and names TypeScript support as planned.

For a coding assistant, the proposed views include hierarchical summaries, semantic search and symbol graphs that describe dependencies and the impact of a change. Jin argues that grep can be sufficient for small projects, while larger codebases and team knowledge need more selective retrieval. Progressive disclosure gives an AI agent a small initial view with ways to request further detail.

The separate CocoIndex Code tool, built on the framework, provides the demo. It searches the OpenClaw codebase, adds a function and shows it appearing in search results. A comparison shows Claude completing a query sooner with CCC search while the other session continues searching. This is one demonstration, not a general benchmark. Jin describes the tool as open source, with a CLI and skill integration for Claude or Codex.