Existing RAG systems treat closed support cases as static documents, but real troubleshooting evolves through stages. RAFT instead represents each historical case as a directed sequence of timeline entries, embeds and retrieves at the entry level, and returns the full parent-case trajectory from the matched state. An optional case-level graph links related cases through configurable similarity.
The framework was evaluated at the retrieval layer rather than as a full agent, avoiding the need for production deployments. The authors used a synthetic benchmark of 826 cases built from Microsoft Learn Windows Server documentation, plus real Apache Jira issues with human-created duplicate labels. RAFT improved Case Hit over vanilla RAG and GraphRAG at every stage, with statistically significant gains over the strongest baseline; the Jira results provided directional evidence that the advantage transfers to real case histories.
This matters because support agents need guidance from similar past cases while an investigation is still unfolding, and matching intermediate investigation states rather than whole documents is a more natural signal. The benchmark and implementation are public, which lowers the barrier for comparing retrieval approaches.