Skip to content

SKILL.md

Latest commit

 

History

History
133 lines (95 loc) · 4.97 KB

SKILL.md

File metadata and controls

133 lines (95 loc) · 4.97 KB
Raw
name knowdb-local-search
description Step-by-step guide for querying a KnowDB knowledge base using bash and grep. Use when answering questions from documents ingested into db/.

KnowDB — Local Search Guide (bash / grep)

For use by local coding agents with filesystem access. The db/ directory is the knowledge base. All commands below assume the repo root as working directory.

Step 0 — Session setup

At session start — before your first search — unconditionally generate a stable id and write it to .session_id at the repo root (overwrite; a new session gets a new id):

date -u +%Y%m%dT%H%M%SZ-$RANDOM > .session_id

Step 1 — Discover available documents

cat db/_manifest.json

Returns { "<doc_id>": { "originalFilename": "...", "title": "..." }, ... }. Each doc_id is an 8-hex string that maps to a subdirectory under db/. title is the human-readable document name — cite it (not the filename or doc_id) when referring to a document.

Step 2 — Orient within a document

cat db/<doc_id>/_index.md

Shows the full heading tree: section titles and their chunk IDs. Read this before searching to identify which chunks are relevant.

_index.md is also your breadcrumb: it maps every chunk ID to its heading title and shows the full root→chunk path. After a grep hit, read _index.md to recover where the chunk sits — its title, parent heading, and siblings — instead of cat-ing parent chunks just to orient.

If unsure which document to look in, scan all heading trees at once:

grep -ril "<keyword>" db/*/_index.md

This is fast and low-noise — headings only, no body content.

Step 3 — Targeted search

Search within one document (recommended):

grep -rinE "<keyword>" db/<doc_id>/

Search across all documents:

grep -rinE "<keyword>" db/ --include="*.md" --exclude="_index.md"

Keyword contract: the keyword is one or more literal terms joined by | (simple OR) — a|b matches a OR b. Whitespace is literal (a space matches a space, not multiple keywords). Matching is case- insensitive by default. Other regex metacharacters are not supported; behavior is undefined and gap recording falls back to a single raw- keyword topic.

Useful flags:

  • -r recursive, -i case-insensitive, -n show line numbers
  • -l list matching files only (for a quick overview)
  • -P enable Perl-compatible regex: grep -rinP "a|b" db/<doc_id>/

Each result shows the file path (<doc_id>/<chunk_id>.md) and the matching line. Read the surrounding lines before fetching the full chunk.

Step 4 — Read minimally

Need Command
Browse chunk IDs in a document ls db/<doc_id>/
Preview first line of each chunk for f in db/<doc_id>/*.md; do echo "$f:"; head -1 "$f"; done
Read one chunk in full cat db/<doc_id>/<chunk_id>.md
Read only matching lines with context grep -in -C 3 "<pattern>" db/<doc_id>/<chunk_id>.md
Find parent chunk strip the last -XX segment: 01-02-03 → parent is 01-02
Read parent chunk cat db/<doc_id>/<parent_id>.md
Read siblings ls db/<doc_id>/ | grep "^<parent_id>-"
Read the whole document (chunked nav not enough) cat $(ls db/<doc_id>/*.md | grep -v _index.md | sort) — body-only concat in chunk-ID order; no headings (unlike browser reconstruct_document) — read _index.md alongside for the heading tree

Step 5 — Lateral cross-document jump

KnowDB has no explicit [[ref]] registry — related material is found by shared terms. After reading a chunk that matters, take its salient terms and grep them across the other documents:

grep -rinlP "<term1>|<term2>" db/ --include="*.md" --exclude="_index.md" \
  | grep -v "db/<current_doc_id>/"

This surfaces connected material elsewhere in the knowledge base that a single-document (scoped) search would miss.

Core rules

  1. manifest → _index.md → scoped grep → cat — always in this order.
  2. Read grep output before fetching full chunks — the matching line is often enough.
  3. Use -C for context when a chunk is long: grep -in -C 3 "keyword" <file> returns matching lines with 3 lines of context each side.
  4. Scope every search to a doc_id subdirectory once you know the target document.
  5. Never cat a full document just to scan it — grep + _index.md first. Read the whole document only when you genuinely need it as continuous text.
  6. Chunk ID encodes position in the heading tree: 01 = first top-level section, 01-02 = its second subsection, 01-02-03 = one level deeper. Use this to navigate without reading files.
  7. _index.md is the breadcrumb — read it to learn a chunk's title, parent and siblings instead of cat-ing parent chunks to orient.
  8. Follow lateral links: after reading a strong chunk, grep its key terms across the other documents (Step 5) — it surfaces implicit cross-document connections a scoped search misses.