blink
Multiple Jev walkers navigate a repository tree one directory choice at a time.
What Jev decides
Which visible file or subdirectory should a walker inspect next.
Each walker sees only the current directory entries and uses Choice to select the most promising file or subdirectory for a natural-language query, producing traceable routes through the codebase.
- Jev role
- Route independent search walkers through the most promising repository paths.
- Input
- The natural-language query, current path, immediate child names, and walker history.
- Output
- A Choice distribution over the current directory entries.
- Decision frequency
- At every directory step for every active walker.
- Fallback
- Return ranked paths and traces for human inspection rather than asserting a final answer.
How it works
Overview
blink treats code search as navigation rather than global retrieval. Starting at the repository root, independent walkers repeatedly ask Jev which immediate child is most likely to contain the answer. The output preserves each path and probability distribution, making the search route inspectable.
The prototype does not build embeddings or a vector index. Its search quality depends on names and local directory context, and the repository does not declare a license, so reuse terms should be confirmed with the authors.
- Language
- TypeScript
- License
- Not declared
- Author
- Ellipsis
Sources
Related projects
pg-jev
Use Jev decisions directly inside PostgreSQL filters, rankings, and classifications.
What Jev decides
Whether a row matches, how strongly it matches, or which supplied label best describes it.
duckdb-jev
Ask Jev questions over DuckDB rows and keep typed SQL results.
What Jev decides
Which label fits, where a row falls on a rubric, or the probability of a yes answer.
jev-semgrep
Search files by whether each line satisfies a meaning, including across languages and negation.
What Jev decides
The probability that a line matches each meaning, including negated or composed conditions.