Code semantic search
Skill yigityildiz0/universal-ai-skill-library/skills/common/code-semantic-search
531 searchable AI Agent Skills for Claude Code, OpenAI Codex, and OpenCode — EN/TR catalog, platform and risk notes, direct ZIPs, and curated bundles.
npx -y skills add yigityildiz0/universal-ai-skill-library --skill code-semantic-searchAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
3 things to look at
- 18 days oldThe repository was created 18 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Design, evaluate, or operate privacy-aware semantic and hybrid search over a large codebase using structural chunks, lexical retrieval, optional embeddings.
The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.
SKILL.md
2.5 KB, as published. Nobody here has run it
Semantic Code Search
Start with repository search, symbol indexes, and targeted reads. An index is justified when the corpus is too large, natural-language concepts do not map to identifiers, or repeated discovery cost exceeds indexing cost.
Requirements
- Define queries and relevance judgments before choosing a backend.
- Inventory languages, generated/vendor/secrets paths, repository size, update rate, data boundary, hardware, and latency target.
- Chunk on symbols and syntax boundaries when a verified parser exists; preserve file path, symbol, line range, language, commit/hash, and parent context.
- Use lexical/BM25 and exact identifier matching as a strong baseline. Add embeddings only when evaluation shows a measurable recall gain.
- Prefer local processing for private code. Any external embedding/vector/reranking service requires explicit authorization, retention/privacy review, and secret filtering.
- Keep embedding model/backend configurable and record its exact version, dimensions, tokenizer, normalization, and index schema. Do not choose a provider/model automatically.
- Deduplicate chunks, cap generated/minified content, and exclude secrets before indexing.
- Incrementally update by content hash; delete stale chunks when files disappear or symbols move.
Retrieval pipeline
Normalize the query without destroying identifiers, retrieve lexical and semantic candidates, fuse ranks, rerank only a small bounded set, diversify near-duplicates, and return source-linked snippets. Never answer from vector text without reopening the current source file; the index may be stale.
Evaluation
Create a labeled set of realistic "where/how/what depends on" questions. Measure recall@k, MRR/nDCG, exact-symbol recall, stale-result rate, latency, index size, build time, and privacy failures. Compare lexical-only, semantic-only, and hybrid. Test renamed symbols, cross-language calls, common words, generated code, and deleted files.
Completion gate
- every result links to a current path/symbol/range and commit/hash;
- direct search fallback remains available;
- private code stayed inside the approved boundary;
- index/update/delete behavior was tested;
- chosen complexity beats the baseline on the evaluation set;
- setup, dependencies, rebuild, and removal steps are documented.