xoai/sage-wiki

What it solves

sage-wiki is a graph-based memory and knowledge base designed for collaborative use between AI agents and humans. It solves the problem of fragmented information by transforming raw documents (PDFs, emails, code, etc.) into an interlinked wiki with a knowledge graph, allowing for multi-hop relational queries that standard vector search cannot handle.

How it works

The system uses an LLM-powered "compiler" pipeline to process source documents into Obsidian-compatible markdown articles. It extracts entities and typed relations to build a knowledge graph.

Key mechanisms include:

• Hybrid Search: Combines lexical (BM25), vector, and graph proximity to find information.
• Evidenced Graph: An optional pass that extracts triples (subject-relation-object) with confidence scores and direct citations to the source text.
• Entity Resolution: Merges surface-form variants (e.g., "K8s" and "Kubernetes") into a single canonical node.
• MCP Integration: Exposes 19 tools via the Model Context Protocol (MCP), allowing AI agents to search, capture, and query the wiki directly.

Who it’s for

• Individuals: Users with existing Obsidian vaults who want an AI-enhanced personal knowledge base.
• Teams: Groups sharing a wiki via Git or a self-hosted server to collaborate on knowledge.
• Companies: Organizations requiring scalable storage (PostgreSQL/pgvector) and authenticated access for large-scale company knowledge graphs.

Highlights

• Multi-format Support: Ingests everything from PDFs and Word docs to emails, transcripts, and images (via vision LLMs).
• Agent-Human Synergy: Provides a TUI, Web UI, and MCP tools so both humans and agents can interact with the same data.
• Provenance and Trust: Every evidenced relation records its source document, and query outputs can be quarantined for verification.
• Scalable Architecture: Supports tiered compilation to handle over 100,000 documents efficiently.
• Local-First: Can run on local models (e.g., via Ollama) for zero-cost, private operation.