Sapien · Knowledge Brain
RAG, a knowledge graph, and a 3D vault viewer
About
Sapien's institutional memory was scattered across call transcripts, field notes, specs, and chat threads. Ali built the knowledge brain that pulled it into one system: an append-only corpus of 260+ source documents (including 120+ discovery call transcripts), a curated wiki of 130+ synthesis notes, a RAG setup where a custom agent answers questions from the actual corpus, and a knowledge graph over all of it.
The design rule that makes it compound: raw sources are append-only and never edited, synthesis lives separately, and every answer the agent produces is saved back into the wiki. Each session makes the next one smarter.
Process
Corpus architecture
Split the vault into raw-sources (append-only truth: discovery calls, field research, ICP work, market research) and wiki (synthesis: GTM docs, product specs, session briefs, team updates). A sync script pulls fresh material from the product repo so the corpus tracks reality without manual filing.
The agent
Wrote the operating instructions that make Claude Code answer from the vault instead of generic training knowledge, with a governance spec for what gets written back and where. Planning, strategy, and team updates now draw on 1 accurate source of truth.
The knowledge graph
Ran the full corpus through a graph-extraction pipeline: 617 nodes, 587 edges, and 79 detected communities, with 88% of edges extracted directly from the text. The graph report surfaces the most-connected concepts, so the team can see which ideas actually anchor the company's thinking.
The 3D viewer
Built a Three.js viewer that renders the whole vault as an interactive 3D force-directed graph, with layout modes (brain, orbit, spiral, organic) and a time-lapse that replays notes in creation order. 1 command launches it locally, no API keys, so any teammate can explore the vault in 3D.
Outcome
Adopted as the team's shared knowledge system, cloned under the company GitHub org with its own onboarding guide. New teammates go from zero to asking the corpus real questions in minutes, and strategy discussions cite the vault instead of memory.