Terraphim AI builds knowledge graphs from your documents and searches them with sub-millisecond automata. Everything runs on your machine. Nothing leaves your control.
Terraphim AI indexes your personal knowledge, team wikis, and public repositories through pluggable data sources called haystacks. Each haystack -- whether a local folder, a Confluence space, a Discourse forum, or an email inbox -- feeds into a role-specific knowledge graph.
The system uses Aho-Corasick automata for pattern matching, BM25 and graph-based relevance functions for ranking, and thesaurus-driven concept expansion for semantic recall. All computation happens locally. Built by Applied Knowledge Systems Ltd.
Rolegraph builds per-role knowledge graphs with concept nodes and semantic edges. Each role maintains its own graph, thesaurus, and relevance configuration. Personalised search without centralised profiling.
Pluggable indexers for local files (ripgrep), Atlassian Confluence, Jira, Discourse, JMAP email, Quickwit, and MCP protocol. Each source implements a common interface. Add your own in a single Rust file.
Aho-Corasick automata compiled from thesaurus entries. LeftmostLongest matching. Autocomplete in under 5ms. Compiles to WebAssembly for browser deployment at roughly 200KB compressed.
Three scoring strategies: TitleScorer for fast text matching, BM25Plus for statistical relevance, and TerraphimGraph for full semantic graph traversal with thesaurus expansion.
Firecracker microVMs provide sub-2-second sandboxed execution. Untrusted code, web requests, and file operations isolated in ephemeral virtual machines. VM pooling minimises startup overhead.
Model Context Protocol server exposes autocomplete, text processing, thesaurus management, and graph connectivity as tools. stdio for local development, SSE for production. OAuth optional.