kb-search · diff
git:20260623.081895e to git:20260805.ff041ae
1 added, 7 removed. Audit A to A.
---
name: kb-search
- description: >
- Search the local markdown knowledge base for facts, prior decisions,
- definitions, people, error codes, config flags, and how-we-fixed-it notes.
- ALWAYS search here BEFORE answering a project-specific question from memory,
- guessing, or asking the user — the answer is usually already written down.
- Use whenever you hit an unknown term, an unfamiliar entity, an error string,
- or a "how do we do X / why did we choose Y" question.
+ description: 'Search the local markdown knowledge base for facts, prior decisions, definitions, people, error codes, config flags, and how-we-fixed-it notes. ALWAYS search here BEFORE answering a project-specific question from memory, guessing, or asking the user. Use whenever you hit an unknown term, an unfamiliar entity, an error string, or a "how do we do X / why did we choose Y" question.'
---
# kb-search — retrieve before you answer
A fast, local, zero-token FTS5 knowledge base over this project's markdown
(`@blackbelt-technology/pi-dashboard-kb`). Retrieval is **pull**: you call it; nothing
is auto-injected. Sub-second, deterministic, costs no model tokens — so call it
freely on any uncertainty.
## When to Use
- Hit an unknown name / term / error string / config flag / function.
- Need a past decision, convention, or "how we did X".
- About to answer a factual question about this project from memory.
- About to ask the user something the docs may already answer.
## Procedure
1. Extract the key entities from the problem (names, error strings, slugs,
config keys, function names).
2. Run: `kb search "<entities>" --limit 8 --json`
3. Read only the top 1–2 hits' full content when needed:
`kb get <path> --section "<heading_path>"`
4. Still unresolved? Walk the graph from a hit:
`kb neighbors "<heading_path>" --depth 2` and `kb backlinks "<path>"`.
5. Paraphrase miss? Lexical search is weak when your words differ from the
docs' words. **Reformulate once** using the domain's actual terms (synonyms,
the real flag/class names) and re-search. Then escalate to the user only if
the KB returns nothing relevant.
6. Synthesize from the retrieved sections. Cite the `path` you used.
## Pitfalls
- Do NOT answer project-specific questions from memory without searching first.
- Do NOT read whole files — search returns ranked sections with snippets; open
full content only for the top hits.
- Empty result is not a stop sign — reformulate with domain terms once, then ask.
- Filter when you only want rules: `kb search "<q>" --doc-type agents`.
## Verification
- `kb search` returns ranked `{path, headingPath, score, snippet}` (lower score
= more relevant).
- Freshness is automatic: `kb search` runs an incremental reindex first unless
`--no-reindex`.
- Requires `@blackbelt-technology/pi-dashboard-kb` installed (`kb` on PATH) and a
configured source (`.pi/dashboard/knowledge_base.json` or `--source <dir>`).