epistract

Epistract

A two-layer knowledge graph framework for the analyst working inside a question.

📄 Paper: Epistract: A Two-Layer Knowledge Graph Framework with an Epistemic Super-Domain Layer — paper/v2/main.pdf (April 2026, framework v3.2.x).

🎥 Demo (5 min): Epistract at Knowledge Graph Conference 2026 (May 2026, framework v3.2.2).

I have spent the last decade building enterprise knowledge graphs — Anzo from the Cambridge Semantics era, AWS Neptune / RDF stacks, Neo4j deployments at multiple companies. They worked. They answer the questions executives ask: what do we know across the organization, where do teams' work intersect, what coverage do we have on a topic. They are the right tool for breadth.

The audience this framework is built for is a different one. The biomedical researcher screening a 30-paper landscape for a target-validation decision. The regulatory specialist comparing seven FDA product labels for a class. The CI analyst working through a competitor's patent stack. The contracts reviewer triaging a vendor portfolio. Their questions are bounded by a corpus they have curated for a specific decision, and the graph they need exists for that question — built fast, queried hard, archived when the decision is made. Maintaining a permanent enterprise graph is the wrong tool for that workflow: too slow to build, too broad to be precise, too costly to retire.

Epistract is built for that workflow. Point it at a corpus, get a structured graph that answers depth questions about that corpus with citations, plus an epistemic layer that distinguishes peer-reviewed claims from patent forward-looking language from hedged hypotheses. The whole loop — corpus → extraction → graph → query → archive — runs end to end in a single specialist's working session, inside Claude Code. No platform team. No ontology committee. No maintained mega-graph.

What makes this practical now is the multi-agent AI harness. Claude Code runs the extraction agents in parallel; the workbench chat reads the same graph data the narrator wrote; one persona drives both reactive Q&A and a proactive briefing. Epistract is one tool inside that harness — it does graph construction and epistemic analysis. Claude Code does corpus acquisition, deep research, and archival.

---

What it looks like

Epistract running on a 34-document GLP-1 corpus (10 patents + 24 PubMed papers): 278 entities, 855 typed relations, 10 communities, built in ~22 minutes. Each node colored by entity type, each edge a typed relation with a verbatim source quote and confidence score. Pan, zoom, filter, click any node for neighborhood context. The interactive workbench is /epistract:dashboard.

The chat panel is grounded in the same graph data the narrator wrote — ask about prophetic claims, contested indications, or coverage gaps and it answers with citations to the originating documents.

More screenshots: docs/WORKBENCH.md.

---

What Epistract is — and isn't

Is

• A consolidation layer for a small-to-medium corpus (7–34 documents in the eight validated scenarios)
• A two-layer graph: brute facts from the corpus + epistemic classification on every relation
• A multi-agent harness that runs end to end inside Claude Code
• Pluggable — four pre-built domains; new domains via a 15-minute wizard from sample documents

Isn't

• A persistent enterprise KG store — use /epistract:export → Neo4j / Neptune / SQLite for that
• A document acquisition tool — Claude Code's web tools, MCP connectors do that; Epistract starts after acquisition
• A general-purpose answer engine — it answers depth questions about the corpus you provided, not the open web
• Self-learning across runs — refinements are human-mediated; Issue #15 tracks the aspirational compounding mechanism

---

The two layers

Layer 1 — Brute facts. Entities and typed relations extracted from the corpus, each with a confidence score and a verbatim source quote. Pydantic-validated at write time — no silent drops.

Layer 2 — Epistemic super-domain. Every relation carries a categorical status: asserted (stated with quantitative evidence), prophetic (patent forward-looking language), hypothesized (hedged wording), contested (multiple sources with conflicting confidence), contradiction (opposing evidence), negative (explicit absence), speculative. Domains can extend this — the FDA domain ships a four-level evidence-tier classifier (established / observed / reported / theoretical) that runs alongside the v3 vocabulary.