Kasi Search

/kasi-search

Semantic search over .kasidit/knowledge/ using local embeddings. No network calls.

Usage

python3 ~/.claude/skills/kasidit/embedding/embed_build.py             # one-time build
python3 ~/.claude/skills/kasidit/embedding/embed_search.py "<query>"  # search

What it does

• Builds a local embedding index across .kasidit/knowledge/.
• Returns top-k chunks with file path, line range, and similarity score.
• Runs entirely local — no outbound API calls.
• Falls back to grep -r if the embedding index is not built.

Flow

1. User asks a question.
2. AI runs kasi-search "<keywords>".
3. AI reads the top 3 results.
4. If relevant, answer grounds in that cached knowledge with an explicit citation.
5. If semantic score is weak (< 0.3), ignore results and proceed.

When to use

• Before answering, check whether the project has relevant cached knowledge.
• "Have we seen this before?" lookups during long sessions.
• When .kasidit/knowledge/ is large (> 20 files) and scanning by hand is slow.

When NOT to use

• Tiny knowledge base where grep is already enough.
• Top-1 match with score < 0.3 — do not force weak signal into the answer.

Anti-patterns

• Over-relying on top-1 — always consider top 3.
• Quoting a chunk without reading the full source file for context.
• Skipping citation — always cite as Based on .kasidit/knowledge/<file>.md.
• Treating low-similarity results as matches.

Fallback

grep -r "keyword" .kasidit/knowledge/

Use when embeddings have not been built or the Python environment is unavailable.

Since

Introduced in v0.8.0.

See also

• Commands (aggregate)
• Kasi-Docs
• Kasi-Scaffold