How OpenMind turns conversation into a knowledge graph
Three pillars on the landing page — Talk it, Map it, Own it — say what the product does in one breath each. This page tells the longer version: what each pillar contains, how the parts hang together, and where to read more.
Talk it.
A normal chat is the input surface. You write the way you would to a teammate — "Linnea joined the apartment-search yesterday, she's leaning toward the Charlottenburg flat" — and an LLM extracts the entities, relations, decisions, and risks in the background. Drop in PDFs, Word docs, audio, or video; the same extractor walks the text or transcript. Long pastes split into segments behind the scenes so a 5,000-word handover note still flows through. The fuzzy-match layer (canonical-key + pg_trgm) means re-mentioning a person or a decision later attaches to the existing node instead of creating a duplicate.
Map it.
Every chat turn becomes one or more typed nodes (people, decisions, risks, requirements, artifacts) and edges (depends_on, part_of, conflicts_with, …). The graph lives in Postgres + pgvector, RLS-scoped to your projects. Asking it a question runs a BFS subgraph walk from semantically-relevant seeds (depth 2, ≤40 nodes) and feeds that subgraph back to the model — answers cite the exact nodes they leaned on, so nothing the model says is unanchored. Six export formats cover hand-off use cases — Markdown for LLMs, Mermaid for GitHub PRs, OPML for outliner apps, FreeMind for classic mindmap editors, PNG/SVG for slide decks.
Own it.
Apache-2.0 from day one. The same docker-compose stack that backs the hosted demo runs on a $5/month VPS or your laptop — your data, your Postgres, your keys. Bring Anthropic, OpenAI, Ollama (local), or LM Studio (local); switch per project without losing the graph. Row-level security on every table means a user's project rows are invisible to every other user — enforced by the database, not by application code. "Download my data" produces a portable ZIP of every project, conversation, message, node, and edge. Delete-account cascades through every dependent row in a single transaction.
These three loops feed each other. You talk to grow the graph, you query the graph to surface what matters, and you own the entire stack so neither the talking nor the querying lock you in. If a piece feels missing or wrong, the FAQ usually has the shortest answer; everything else lives in the per-feature help pages.