Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.khal.ai/llms.txt

Use this file to discover all available pages before exploring further.

Brain treats your vault as a typed knowledge graph, not a bag of chunks. Documents have types (note, decision, domain, reflection, observation), links have types (semantic, tag-overlap, derived-from), and every retrieval result is uplifted by its position in that graph. This page covers the operations you use to build, maintain, and grade the graph. brain link proposes links automatically from two signals:
  1. Tag overlap — documents sharing 2+ tags are candidate links.
  2. Semantic similarity — documents with cosine similarity above threshold are candidate links.
brain link --brain my-brain
Links are stored in the brain_links table with a link type and a confidence score, and also materialized back into the markdown files as wiki-style [[document]] references (opt-in via --apply). Obsidian users get free bidirectional navigation; git users get a diff of which links were added.
Link generation is the single biggest driver of retrieval quality over time. A 500-doc vault with dense linking beats a 5,000-doc vault with no graph structure.
At query time, the retrieval pipeline runs a link-boost pass over the top-N results. For each hit, it uplifts any linked document’s score by +0.15, capped at +0.45 total per document. This means a weak match that’s linked to a strong match gets pulled into the result set — mimicking how a human researcher follows citations. See Retrieval for the scoring detail.

​
Graph analysis

brain graph reports the structural health of the graph: 
brain graph --brain my-brain
Outputs:
  • Orphans — documents with zero in-links. Candidates for deletion or re-linking.
  • Connected components — isolated subgraphs. Often indicates unrelated topics should be split into separate brains.
  • Link-degree distribution — hub-and-spoke vs. mesh structure.
  • Authority scores — PageRank-style scoring over the link graph.
A brain with more than ~15% orphans is a sign the vault has lost coherence. Either run brain link to repair, or brain forget the orphans.

​
Deduplication

brain dedup finds near-duplicate documents — the copy-paste-from-Slack pattern that silently bloats vector stores: 
brain dedup --brain my-brain --threshold 0.85
Output:
  • Groups of documents with ≥85% semantic similarity.
  • For each group: canonical pick (most-linked, highest-authority), and the others flagged for merge or deletion.
  • Audit trail written to _reflections/dedup-<date>.md so you can review before applying.
Apply a dedup decision explicitly: 
brain dedup --brain my-brain --apply --keep <doc-id> --remove <doc-id-2>,<doc-id-3>

​
Health checks

brain health lints the vault and reports structural issues: 
brain health --brain my-brain
Checks include:
  • Missing embeddings (content changed but not re-embedded).
  • Stale chunks (chunk config changed since ingest).
  • Broken [[links]] (target doc removed or renamed).
  • Frontmatter contract violations (e.g., document declares confidence: high but has no sources).
  • Empty or under-populated domains.
Run brain health --fix to apply the auto-fixable subset (re-embed, re-chunk, repair obvious link typos).

​
Quality scoring

brain score returns a 4-dimension quality grade (A–F): 
brain score --brain my-brain
The four dimensions:
DimensionWhat it measures
ContentCoverage (domain breadth), depth (avg. doc length, citations), freshness (staleness of top docs)
StructureLink density, orphan percentage, domain balance
RetrievalSweep-grade against the seeded question set — does search actually find the right docs?
HygieneDedup debt, health-check pass rate, frontmatter-contract compliance
A brain graded B- on retrieval but A on content is telling you: “your content is great, but your retrieval config doesn’t find it.” Run brain sweep next.

​
Decisions

brain decisions surfaces, edits, and applies strategy decisions recorded over time: 
# List recorded decisions
brain decisions list --brain my-brain

# Show a specific decision
brain decisions show --brain my-brain --id 2026-04-17-strategy-segment-code-queries

# Apply a decision (e.g., after rolling it back)
brain decisions apply --brain my-brain --id <id>
Decisions are markdown files under _decisions/ — auditable, diff-able, git-commitable. brain auto-kb writes to this directory automatically; you can edit the files by hand and brain decisions apply to re-materialize the config.

​
Forgetting

brain forget removes documents with a full audit trail: 
brain forget --brain my-brain --id <doc-id> --reason "outdated — superseded by <other-id>"
Forgotten docs are:
  • Removed from brain_documents, brain_chunks, and brain_links.
  • Logged to _reflections/forget-<date>.md with the reason and the content hash (so you can re-ingest if you change your mind).
  • Their outbound links re-scored against remaining docs.

​
A weekly graph-ops loop

For a brain that’s used actively but not via auto-kb:
1

Monday: brain health + fix auto-fixable

brain health --fix --brain my-brain — keeps embeddings and chunks current.
2

Wednesday: link pass

brain link --brain my-brain --apply — new content accumulated since last link run gets integrated.
3

Friday: dedup + forget review

brain dedup --brain my-brain --threshold 0.85 — read the audit, apply the merges you agree with.
4

End of week: score

brain score --brain my-brain — if anything dropped a grade, the next week starts with a sweep.
Most users skip the manual loop and run brain auto-kb --rounds 1 once a week instead.

​
What’s next

Closed-loop intelligence

The one-command autonomous version of everything on this page.

Retrieval

How the graph operations above lift retrieval quality at query time.