ck

name	ck
description	Use when searching code with ck/seek for exact grep-style matches, semantic search, lexical search, hybrid search, index management, JSONL output for agents, or MCP server mode.
metadata	{"short-description":"Semantic, lexical, and hybrid code search"}

ck

Use ck (seek) when the task is about finding code by meaning or by exact text, especially in large repos where index-backed search helps.

Search modes

Mode	Flag	Index needed	Best for
Regex (default)	--regex or none	No	Exact identifiers, literals, paths
Semantic	--sem	Auto-built	Conceptual queries ("error handling", "retry logic")
Lexical	--lex	Auto-built	Ranked full-text phrases
Hybrid	--hybrid	Auto-built	Keyword precision + semantic recall

When to use

• Use ck --sem for conceptual searches like "error handling", "authentication logic", or "retry logic".
• Use ck --lex when you want BM25-ranked full-text search.
• Use ck --hybrid when you want keyword precision plus semantic recall.
• Use plain ck with -n, -r, -i, -w, -F, -C, -A, and -B for grep-compatible exact search.
• Use --jsonl for agent pipelines and --serve when an MCP client should query the repo directly.

When not to use

• Do not use semantic search for exact identifiers, literal strings, paths, or config keys. Use grep-style search instead.
• Do not guess at stale index state; check --status or --status-verbose, and rebuild with --index when results look off.
• Do not use --serve unless the workflow really needs MCP integration.

Workflow

1. Start with the narrowest useful search.
   • ck "TODO" src/
   • ck -n -r "fn main" .
2. Switch to meaning-based search when exact text is missing.
   • ck --sem "error handling" src/
3. Tighten or broaden results with scores, thresholds, and limits.
   • ck --sem --scores --threshold 0.7 "validation" src/
   • ck --hybrid --limit 5 "timeout" src/
   • ck --sem --limit 5 --threshold 0.8 "auth" src/
4. Improve relevance with reranking.
   • ck --sem "retry logic" --rerank src/
   • ck --sem "login" --rerank-model bge src/
5. Use JSONL or JSON for tools and agents.
   • ck --jsonl --sem --no-snippet "authentication" src/
   • ck --json --sem --limit 5 "bug fix" src/
6. Check or manage the index when needed.
   • ck --status .
   • ck --status-verbose .
   • ck --index --model nomic-v1.5 .
   • ck --index --model jina-code .
   • ck --switch-model nomic-v1.5 .
   • ck --clean-orphans .
   • ck --clean .
   • ck --add path/to/file.rs
7. Expose ck to an MCP client when asked for agent integration.
   • ck --serve

Defaults

• Semantic search defaults: top 10 results, threshold >= 0.6.
• --topk and --limit are aliases for limiting result count.
• Hybrid RRF scores are in 0.01–0.05 range (lower than semantic 0.0–1.0).

Quick reference

• Exact text: ck "pattern" path/
• Semantic: ck --sem "what the code does" path/
• Lexical: ck --lex "multi word phrase" path/
• Hybrid: ck --hybrid "keyword + concept" path/
• Agent JSONL: ck --jsonl --sem "concept" path/ --no-snippet
• Agent JSON: ck --json --limit 5 "query" path/

Practical notes

• ck is grep-compatible, but semantic and lexical modes auto-index the target path.
• .ck/ is a cache directory. It can be rebuilt safely with --index or removed with --clean.
• Use --exclude, --no-ignore, or --no-ckignore only when you need to override default file filtering.
• Available embedding models: nomic-v1.5 (8k context), jina-code (code-specialized).