Skip to content

Metis#

Metis turns the documents on your drives into knowledge bases you can ask questions of. A knowledge base reads Markdown, text, HTML, PDF and Word files, indexes them on one of your machines with an embedding model you serve with Eos, and answers each question with the passages that answer it best — every one with a citation: the drive, the document, the page or headings, the line and a snippet. The documents and the questions never leave your machines.

flowchart LR
  D[Drives: Markdown, text,<br/>HTML, PDF, DOCX] --> I[Knowledge base<br/>on your machine]
  E[Eos embedding model] --- I
  R[Eos reranker] --- I
  Q[Question:<br/>console, astra, API, MCP] --> I
  I --> A[Passages with citations]

What it does#

  • Hybrid search by default. Each question is matched by meaning (vectors from your embedding model) and by its words (BM25), the two lists are fused, and a reranker orders the best candidates. A search takes tens of milliseconds on a CPU.
  • Citations always. Every result says where it comes from, so a person or a model can check it.
  • Incremental. When files change, only the changed files are read again; files that are gone are dropped. A drive backed by a data source is indexed again when it has a new revision.
  • Measured. An evaluation set — questions and the documents that answer them — gives recall@k and MRR, so "it finds it" is a number.
  • Every door. The console (Metis), astra metis, the API and your AI assistant over MCP.

Start here#