Knowledge bases
A knowledge base (KB) is a document store your agents can search but never write. You add files in any format; Coffer normalizes each to Markdown on disk — the source of truth — and serves them back through three retrieval modes. The Markdown files are canonical; the SQLite index is rebuildable with coffer kb reindex.
Create and fill a KB
coffer kb create handbook # default modes: keyword + grep
coffer kb ingest handbook ./onboarding.pdf # any format → Markdown
coffer kb ingest handbook ./notes.docx
coffer kb list-docs handbook # what's inside- Ingestion converts pdf, docx, pptx, xlsx, html, and more to Markdown (25 MB default cap). Re-ingesting the same source needs
--replace. - Originals are kept under the KB's
raw/folder; a document you hand-edit will not be re-converted from its raw original.
Search
coffer kb search handbook "how do I reset my password" # ranked passages (keyword)
coffer kb grep handbook "TODO" # exact / regex over the Markdown
coffer kb search handbook "vacation policy" --mode vector # semantic (needs embeddings)- grep is exact/regex with no index; keyword (the default) is FTS5 BM25 ranking; vector is semantic nearest-neighbour and is opt-in.
- Enable vector with
coffer kb set-embedding handbook --provider … --model …. A vector search with no embedding configured falls back to keyword rather than erroring.
For your agents
Every connected MCP client automatically gets read-only tools — coffer__list_knowledge_bases, coffer__search_knowledge, coffer__grep_knowledge, and coffer__read_document. There is no write tool: agents read your knowledge, they don't change it.
In the app, knowledge bases live under Resources — create a KB, drag files in, and search; the detail view shows document and chunk counts, disk usage, and which retrieval modes are indexed.