NAME

KB — knowledge base management

SYNOPSIS

DESCRIPTION

Harvey keeps a SQLite knowledge base at /agents/knowledge.db. It stores structured notes about experiments and concepts so you can search and inject that context into conversations without relying on the model’s general knowledge.

The knowledge base is independent of the RAG store (/help rag). KB holds hand-authored structured records; RAG holds embedded chunks from ingested documents. Use both: /kb inject to bring structured records into context, and RAG to retrieve relevant document passages automatically.

CONCEPTS

Project — a named container for a body of work. One project can be “active” at a time; /kb observe attaches to the active project.

Observation — a timestamped note attached to a project. Each observation has a kind:

              note        — general remark
              finding     — empirical result
              decision    — a choice made and its rationale
              question    — open question to return to
              hypothesis  — testable prediction

Concept — a named idea or term that can be referenced across multiple projects and observations.

SUBCOMMANDS

/kb status Show the database path, project count, and observation count.

/kb search TERM [TERM…] Full-text search (FTS5) across all observations and concepts. Supports quoted phrases and prefix wildcards:

  /kb search RAG embedding
  /kb search "context window"
  /kb search grpc*

/kb inject [PROJECT] Format the knowledge base as Markdown and add it to the conversation as a user message. With no argument, injects the active project (or all projects if none is active). With a project name, injects only that project.

  /kb inject
  /kb inject harvey

/kb project list List all projects with ID, name, and status. The active project is marked with *.

/kb project add NAME DESCRIPTION Create a project and set it as the active project.

  /kb project add harvey "terminal coding agent for Ollama"

/kb project use ID Set an existing project as the active project by numeric ID.

/kb observe [KIND] TEXT Record an observation against the active project. KIND defaults to “note” if omitted. Valid kinds: note, finding, decision, question, hypothesis.

  /kb observe finding RAG threshold of 0.3 eliminates noise on granite3-moe
  /kb observe decision switched embedding model to nomic-embed-text
  /kb observe question does bge-m3 outperform nomic on code retrieval?

/kb concept list List all concepts with ID and description.

/kb concept add NAME DESCRIPTION Add a named concept to the knowledge base.

  /kb concept add RAG "retrieval-augmented generation"
  /kb concept add "context window" "token budget for a single LLM call"

LEARNING MODE

/kb learn turns Harvey’s own recorded sessions and hand-off notes into searchable knowledge. Every step ends with a human decision: nothing a model writes is trusted, or found by /kb search, until you accept it. All four steps work on the active project (/kb project use ID).

/kb learn ingest [–min-words N] [–all] [–dry-run] List the hand-off notes and recorded sessions that are not yet in the knowledge base, newest first, and ingest the ones you choose. Hand-offs are always offered; a session must have at least 200 words (–min-words changes that, –all offers every session). New sections start unsummarized.

/kb learn draft [–limit N] [–dry-run] [@model] Ask a model to draft a summary of each unsummarized section, then of each whole document from its section summaries. The model is @model if given, else learn_model in agents/harvey.yaml, else the active model. Default limit is 25 items (–limit 0 drafts everything queued); a small local model takes about a minute per item, so a full run over many documents is long. Each draft gets a confidence score, which is only how many of the known concepts in the source the draft also names. It is not a measure of whether the draft is faithful.

/kb learn review [@model] Walk the drafts one at a time, showing the source excerpt, size, known concepts mentioned, confidence and which model wrote it:

  [a]ccept   trust the summary; it becomes searchable
  [e]dit     open $EDITOR on the draft (the edit is recorded as "human")
  [r]edraft  ask the model again
  [s]kip     leave it drafted (Enter also skips)
  [q]uit

Bare /kb learn drafts, then reviews. A summary whose source changed after it was written is marked STALE.

/kb learn concepts [–limit N] Suggest new concepts from the ingested documents (most distinctive terms that are not yet concepts; default 20). You pick which become concepts, by number, range (1-3), all or none. Each pick is added to the knowledge base at once. Then, for each project document that mentions a picked concept more than once, Harvey shows the changed lines and asks before writing: [[Concept]] links, and footnotes for near-miss spellings.

A write goes through the permissions table (/permissions) like any other file write, and a path Harvey may not write is refused before you are asked. The document is re-ingested straight afterwards; if that fails, the file is put back. Fountain documents (sessions and hand-offs) are never rewritten: they are records, and [[…]] in them reads as a note. Tagging changes source text, so summaries of changed sections are marked stale; the closing line says how many.

WORKFLOW EXAMPLE

  /kb project add myapp "Go CLI for processing audio files"
  /kb observe decision using ffmpeg via exec.Command, not a Go binding
  /kb observe finding ffmpeg probe takes ~80 ms per file on Pi 4
  /kb observe question can we batch probe calls to reduce overhead?
  /kb concept add ffmpeg "audio/video processing CLI"
  /kb inject

After /kb inject the model sees the full project record as context and can answer questions about it, suggest next steps, or help resolve open questions.