A knowledge and memory system for people and AI agents: a document corpus you own, and retrieval over it that searches by meaning as well as by words.

kennis is domain-agnostic. It knows nothing about any particular subject.

kennis corpus add paper.pdf --literature
kennis corpus index
kennis search "why does rank fusion beat score fusion"
kennis read yusk5agyfk --chunks 2:5

What it is

A corpus of markdown documents in three collections - literature, docs and notes - stored as ordinary files in ordinary directories, in a git repository, that you can read and edit without kennis being involved.

Over that, two retrieval legs fused into one ranking: BM25 for the words you typed, and dense embeddings for what you meant. See How retrieval works.

What makes it different

Your documents stay yours. Conversion runs locally by default and embedding runs locally by default. Sending a document to a third party requires setting a backend and supplying a key - never one without the other, and never as a default. Where your data goes lists every way anything leaves the machine.

The relevance level is measured, not invented. When kennis says a hit is high, that is a cosine similarity against cuts measured on a real corpus for that specific model. When it cannot make an absolute claim it says so in the summary line, and when it has no scale at all it shows no level rather than inventing one.

A handle never breaks. A document's identifier is surrogate and fixed at creation, so you can retitle, regroup, or move a document between collections and everything that referenced it still resolves.

The output is a payload. kennis read x > x.md gives you the document byte for byte. Colour is dropped when stdout is not a terminal, and no line is padded or re-wrapped.

Where to start

• New here? Install, then the tutorial - fifteen minutes, no PDFs, no keys.

• Want to do one thing? The how-to guides are task-shaped.

• Want to know why? How retrieval works is the one to read.

• Looking up an option? Commands, generated from the code.

• Not at a terminal? The graphical interface does the same work in a window.

• Wiring up an agent? Serve an agent - MCP over stdio, ten tools.

Status

Built and usable: the corpus, search, the command line, the per-project context bundle, packs, the MCP server and the graphical interface.

Not built: cross-machine synchronisation, and any packaged application. kennis serve is an extra rather than a base dependency, so install kennis[mcp] for it. kennis gui and PDF conversion are in the base install.

This is experimental software under active development. It keeps no backwards-compatibility shims: when something is reworked, the old path is removed rather than deprecated.