Rylox is a fast, local repository context engine for Python codebases.
It analyzes your repository, builds a structured semantic index, and retrieves relationship-aware, token-budgeted context that can be pasted into any large language model.
Unlike keyword search or sending entire files to an LLM, Rylox understands the structure of your codebase — functions, methods, classes, and how they reference each other — and selects only the code relevant to your question. Indexing and retrieval run entirely on your machine: no API keys, no cloud services, no vendor lock-in. The only network access Rylox ever makes is a one-time download of its local embedding model on first use; after that, everything is fully offline.
rylox index
rylox context "How does authentication work?" --max-tokens 20000The output is a structured Markdown document — query, entry points, execution flow, included files, and the actual relevant code — sized to fit within the token budget you set, ready to paste into ChatGPT, Claude, Gemini, Codex, or any other LLM.
Large repositories quickly exceed an LLM's context window. Keyword search often retrieves unrelated files; pasting whole files wastes tokens on irrelevant code; cloud-based embedding services require API keys, cost money per query, and send your code off your machine.
Rylox instead parses your repository's actual structure, combines dense (embedding-based) and sparse (keyword-based) retrieval, follows one hop of real call/reference relationships from the best matches, and assembles the result deterministically within a token budget — all locally.
- Tree-sitter powered Python parsing
- Semantic chunking at the function, method, and class level, including
decorators and
async def - Incremental indexing using SHA-256 content hashing — only changed files are re-parsed and re-embedded on subsequent runs
- Hybrid retrieval: dense (local sentence-transformer embeddings + FAISS) and sparse (BM25) search, combined via Reciprocal Rank Fusion
- One-hop relationship expansion (callers, callees, constructor/decorator usage) from the top-ranked matches
- Deterministic, token-budgeted context assembly — a chunk that doesn't fit the budget is skipped whole, never truncated mid-body
- Structured Markdown output with a fixed section layout
- Persistent on-disk cache for both the parsed index and computed embeddings
.gitignore-aware and configurable ignore patterns, with a configurable per-file size ceilingrylox doctorfor diagnosing your local environment and repository cache state- Offline after first run — the only network call is the initial, one-time embedding model download
- Cross-platform (Linux, macOS, Windows)
Index a repository:
rylox indexGenerate context for an LLM:
rylox context "How does authentication work?" --max-tokens 20000Rylox writes a Markdown file containing the most relevant parts of the repository for that question, within the requested token budget.
Rylox is functional end to end — index, context, clean, and doctor
all work as real commands, not stubs. That said, this is an early release
(v0.0.2, pre-1.0), and retrieval quality is actively being improved:
running Rylox against real open-source repositories has already surfaced
and fixed real bugs in how relationships between code are detected (see
KNOWN_LIMITATIONS.md for specifics and current
known gaps). If you hit a case where retrieval picks the wrong code or
misses something obviously relevant, that's useful signal — please open an
issue.
Offline first. Repository indexing and retrieval never require a network connection after the one-time initial model download.
Vendor independent. The generated context is plain Markdown — it works with any language model, not a specific provider.
Deterministic retrieval. Given the same index and query, output ordering is reproducible — driven by repository structure and fusion scoring, not opaque or randomized ranking.
Install the latest release from PyPI:
pip install ryloxUpgrade existing installation:
pip install --upgrade ryloxContributing or running the test suite:
pip install -e ".[dev]"
pytest- Python (3.9–3.12)
Additional language support is a future goal via Tree-sitter's grammar ecosystem, not yet implemented.
Rylox is licensed under the Apache License 2.0. See
LICENSE for details.