Metis

A knowledge learning and retrieval engine.

Metis transforms books, articles, and videos into structured, retrievable expert knowledge that any language model can use to reason like a domain expert.

It is not an agent, a chatbot, or a domain-specific expert system. It is the tool that makes building such systems possible.

The problem

Large language models know a lot, but their knowledge is broad and shallow. Ask one about a niche topic — childhood bilingual language acquisition, specialized regulatory compliance, obscure engineering disciplines — and you get generic, wishy-washy advice.

Meanwhile, deep domain expertise exists in books, papers, and the minds of practitioners. But it's locked in formats that machines can't reason with.

The idea

Feed Metis a book. It reads, comprehends, and extracts structured knowledge — not summaries, not chunks of text, but what it learned: claims, frameworks, procedures, heuristics, mental models, and the relationships between them.

The output is a knowledge graph of atoms — small, precise, typed units of knowledge inspired by frame semantics in linguistics. Each atom has named roles, conditions, confidence scores, and source provenance.

When a question comes in, Metis retrieves the relevant atoms, traverses the graph to pull connected knowledge, detects gaps, and composes a structured context package. Any LLM can then use this package to reason at an expert level.

Two sides

Learn. Content in, structured knowledge out. The learning pipeline ingests raw material, extracts atoms, and integrates them into a growing knowledge graph.

Apply. Question in, expert context out. The retrieval engine identifies relevant atoms, follows connections, flags missing knowledge, and delivers a structured brief for the downstream model.

Status

Metis is in early design. We are working with the garage door open — follow the log for design decisions, experiments, and milestones.