skylakegrep 0.2.0

Free local semantic code search using Ollama

Quickstart  ·  30s demo  ·  Performance  ·  How it works  ·  Releases  ·  Docs

---

skygrep is a fully-offline semantic code-search CLI for natural-language questions about your codebase. Ask in plain English, get the right file and line range. Indexing, retrieval, and optional answer synthesis all run locally against your own Ollama server. No remote service, no subscription, no data leaves your machine.

In 30 seconds

$ pip install skylakegrep
$ ollama pull nomic-embed-text qwen2.5:1.5b qwen2.5:3b   # one-time
$ cd ~/your-project

$ skygrep "where is the cascade tau threshold defined?"
=== skylakegrep/src/storage.py:578-602 (score: 0.781) ===
CASCADE_DEFAULT_TAU = 0.015

def cascade_search(...
[0.51s · cascade=cheap (gap=0.020 τ=0.015) · index 20s ago · 36 files · L2 symbols on · graph prior on]

That is the entire happy path. First query in a fresh project completes in under 1 s via a ripgrep fallback while a background process builds the semantic index. Every query after that uses the full cascade with a local LLM kept warm in memory.

Quickstart

pip install skylakegrep
ollama pull nomic-embed-text qwen2.5:1.5b qwen2.5:3b   # ~3 GB total

# One-time: register skylakegrep with detected LLM CLIs
# (Claude Code / Codex / OpenCode / Gemini CLI / Cursor)
skygrep setup

cd /your/project
skygrep "<your question>"
skygrep doctor                # verify runtime + models + index + integrations
skygrep stats                 # show current project's index info

skygrep derives the project root from git rev-parse --show-toplevel (falling back to the working directory) and keeps a per-project index under ~/.skylakegrep/repos/. Subcommand names (index, doctor, stats, watch, serve, setup, enrich) take precedence — anything else is treated as a query, so skygrep "stats and metrics" (quoted) is unambiguous.

skygrep setup writes a small markdown snippet to each detected LLM CLI's user-level instructions file (e.g. ~/.claude/CLAUDE.md, ~/.codex/AGENTS.md, ~/.gemini/GEMINI.md) telling the agent to prefer skygrep for natural-language code search and fall back to rg otherwise. Snippets are delimited by markers; skygrep setup --uninstall removes them cleanly without touching your other instructions.

Performance

Public, reproducible benchmark on three popular open-source codebases. 30 hand-labelled questions total (10 per repo). Anyone can clone the repos and re-run with one command — see benchmarks/public_oss_bench.py and docs/parity-benchmarks.md.

Repo	Language	LOC ≈	Tasks	skygrep recall	rg recall	Token reduction
Django	Python	524 K	10	10 / 10	10 / 10	703 ×
Tokio	Rust	80 K	10	10 / 10	10 / 10	61 ×
React	JS+TS	270 K	10	10 / 10	10 / 10	773 ×
Aggregate			30	30 / 30 (100 %)	30 / 30	60×–770×

Honest reading:

• Hit-rate parity across all three (10/10 each on Django, Tokio, React). The two React misses that previously surfaced (test-fixture path bias on react-007, devtools-vs-reconciler on react-010) were resolved by upgrading the embedding substrate to bge-m3 and extending the non-canonical-path filter — see docs/parity-benchmarks.md for the failure analysis and the resolution.
• rg "100 %" is a recall-ceiling baseline. It returns 20 M+ tokens per task (term-OR scan with 2-line context windows). Yes the answer is in the dump; no, the agent has to read the 20 M tokens to find it.
• skygrep delivers the answer ranked top-10 in 30 / 30 cases while emitting 60×–770× less context for the agent's LLM round-trip downstream. That is the user-facing claim.

Recall counts a query as a hit when at least one returned chunk matches the canonical expected path or any of the question's expected_alternatives.

Cascade-only ablation (Django, in-bench numbers)

Tier	What it does	Cold first query	Warm avg s/q
cascade ⭐ default	rg prefilter → file-mean cosine → escalate to HyDE only when uncertain	~10 s (Ollama loads)	0.5–2 s (warm)
cascade-cheap	early-exit only, no LLM call	<1 s	<0.2 s
cascade + small HyDE (OLLAMA_HYDE_MODEL=qwen2.5:1.5b)	uses 1.5 B for HyDE — faster, slightly lower recall	~5 s	2.0 s
chunk + rerank	classic chunk cosine + cross-encoder rerank	~10 s + 30 s reranker load	~10 s
ripgrep raw	rg -il -F token-OR (file membership only)	<1 s	<1 s

The cascade is bimodal by design: ~80 % of queries take the cheap path (file-mean cosine, no LLM call) and complete in under 200 ms warm; the remaining ~20 % escalate to a HyDE-augmented retrieval and complete in the 1–2 s band. With Ollama models kept resident in memory (OLLAMA_KEEP_ALIVE=-1, the 0.6.0 default) the second query in a shell session no longer pays the 5–10 s Ollama cold-load.

A second self-test benchmark compares skygrep against a simulated grep agent over 30 navigation tasks against this repo: 30 / 30 recall at top-k 10 with 2× total-token reduction and 2.9× context-token reduction vs the agent baseline.

How it works

your query
    │
    ▼
┌─────────────────────────────────────────────────────────────────┐
│ 1.  ripgrep prefilter      Fast surface-token narrowing          │
│ 2.  file-mean cosine       Rank files by mean of chunk vectors   │
│ 3.  cascade decision       Confident? return cheap. Else escalate│
│ 4.  HyDE escalation        LLM rewrites query → cosine union     │
│ 5.  symbol + graph         Tree-sitter symbol boost + PageRank   │
│     (L2 + L4)              tiebreaker on near-tied candidates    │
└─────────────────────────────────────────────────────────────────┘
    │
    ▼
top-K chunks (path · line range · score · snippet)

Each layer is offline-paid and query-time-free where possible: embeddings are precomputed at index time, symbol extraction runs once per project, the file-export PageRank is one regex pass over the corpus. Only the cascade's HyDE-escalation path makes a query-time LLM call, and it only runs on the ~20 % of queries the cheap path is uncertain about.

The full architecture diagram and module-by-module walk-through is at docs/skylakegrep-0.1.0.md and docs/roadmap.md.

When to use what

You want	Use
Find code by concept ("how does X work?")	skygrep "<query>"
Find code with a known token	rg <token> (it's faster, no setup)
Synthesize an answer with citations	skygrep "<query>" --answer
Decompose a broad question	skygrep "<query>" --agentic --max-subqueries 3 --answer
Machine-readable output for an agent	skygrep "<query>" --json
Re-rank candidates with a cross-encoder	skygrep "<query>" --no-cascade --rerank
Continuously index a watched dir	skygrep watch /path
Keep the cross-encoder warm across queries	skygrep serve & ; skygrep "<q>" --daemon-url http://127.0.0.1:7878

Configuration

Variable	Default	Effect
OLLAMA_URL	http://localhost:11434	Ollama server URL.
OLLAMA_EMBED_MODEL	nomic-embed-text	Embedding model. Switching requires skygrep index --reset.
OLLAMA_LLM_MODEL	qwen2.5:3b	Used for --answer and --agentic.
OLLAMA_HYDE_MODEL	qwen2.5:3b	Used for cascade-escalation HyDE. Falls back to OLLAMA_LLM_MODEL if not installed. Set to qwen2.5:1.5b for ~30 % speedup at the cost of 1 task on 16-task Rust.
OLLAMA_KEEP_ALIVE	-1	Passed to every Ollama call. -1 keeps models resident indefinitely (recommended).
SKYGREP_DB_PATH	per-project	When set, skygrep treats the index as curated and disables auto-mutation.
SKYGREP_AUTO_PULL	unset	Set yes to auto-ollama pull missing models without prompting.
SKYGREP_AUTO_REFRESH_THROTTLE_SECONDS	30	Skip the mtime scan if the previous refresh ran more recently.
SKYGREP_RERANK_MODEL	mixedbread-ai/mxbai-rerank-large-v2	Cross-encoder for --rerank.
SKYGREP_RERANK_POOL	50	Candidate pool before reranking.

Releases

This is the first public release of skylakegrep. See docs/skylakegrep-0.1.0.md for the full description of capabilities, architecture, CLI flags, and environment variables.

• 0.1.0 — first public release. LLM-driven query routing (filename / lexical / semantic cascade tiers, intent-aware merge), confidence-gated semantic cascade with HyDE escalation + cross-encoder rerank + PageRank tiebreaker, lazy PDF / docx content extraction (pdftotext + pypdf fallback, optional --ocr), framed Pygments-highlighted card rendering, four --detail levels, skygrep setup auto-registration with major LLM CLIs (Claude Code, Codex, OpenCode, Gemini CLI, Cursor). PolyForm Noncommercial 1.0.0 license.

CLI reference

skygrep setup    [--list|--uninstall|--yes] # register with Claude Code / Codex / OpenCode / Gemini / Cursor
skygrep "<query>" [OPTIONS]                 # bare-form search
skygrep search   "<query>" [OPTIONS]        # explicit search
skygrep doctor                              # health check
skygrep stats                               # project index info
skygrep index    [PATH] [--reset]           # explicit reindex
skygrep watch    [PATH] --interval N        # poll for changes
skygrep serve    [--host H] [--port P]      # warm-reranker daemon
skygrep enrich   [--max N] [--batch B]      # opt-in doc2query enrichment

skygrep search options

Option	Default	Effect
-m, -n, --top	5	Number of final results.
--json	off	Emit a JSON array; suppresses human formatting.
--answer	off	Synthesize an answer from retrieved snippets via Ollama.
--content / --no-content	on	Show or hide snippet bodies in human output.
--language	—	Restrict to one or more language keys; repeatable.
--include / --exclude	—	Glob filter (repeatable).
--cascade / --no-cascade	on	Confidence-gated retrieval. Off = chunk-only legacy path.
--cascade-tau	0.015	Top-1 / top-2 file-mean cosine gap above which to early-exit.
--rerank / --no-rerank	on	Cross-encoder rerank on the non-cascade path.
--rerank-pool	50	Candidate pool before reranking.
--rerank-model	env or default	HuggingFace cross-encoder id.
--hyde / --no-hyde	off	Force HyDE outside the cascade (rare; cascade decides per query).
--multi-resolution / --no-multi-resolution	on	File-level cosine top-N → chunk-level inside those files.
--file-top	30	Files surfaced by file-level retrieval.
--lexical-prefilter / --no-lexical-prefilter	on	Use ripgrep to narrow the candidate file set.
--lexical-root	cwd / git toplevel	Root directory ripgrep scans.
--lexical-min-candidates	2	Fall back to corpus-wide cosine when ripgrep returns fewer files.
--rank-by	chunk	chunk (per-file diversity cap) or file (one chunk per file).
--auto-index / --no-auto-index	on	Auto-build the index on first query and refresh on mtime change.
--daemon-url	—	Send the search to a running skygrep serve daemon.
--agentic	off	Decompose into subqueries via Ollama before search.
--max-subqueries	3	Upper bound on agentic subqueries.
--semantic-only	off	Skip lexical reranking; rank by cosine alone.

Capability matrix (every feature, when introduced)

Capability	Since
Semantic code search via local Ollama	0.1.0
Tree-sitter chunking + line-window fallback	0.2.0
.gitignore / .skygrepignore hygiene	0.2.0
Incremental indexing (mtime-based)	0.2.0
Stale row cleanup	0.2.0
Watch mode	0.2.0
Hybrid lexical + semantic ranking	0.2.0
Stable JSON output	0.2.0
Local answer mode	0.2.0
Local agentic decomposition	0.2.0
Cross-encoder rerank	0.3.0
Asymmetric query/document embedding prefixes	0.3.0
HyDE query rewriting	0.3.0
Multi-resolution retrieval	0.3.0
Lexical prefilter (ripgrep first stage)	0.3.0
File-rank (one chunk per file)	0.3.0
Daemon mode	0.3.0
Quantisation / device knobs	0.3.0
Confidence-gated cascade	0.3.0 (default in 0.4.0)
Bare-form invocation skygrep "<q>"	0.4.0
Per-project auto-index	0.4.0
skygrep doctor health check	0.4.0
Ripgrep fallback for first query	0.4.1
Symbol-aware indexing (L2)	0.5.0
doc2query enrichment (L3, opt-in via skygrep enrich)	0.5.0
File-export PageRank tiebreaker (L4)	0.5.0
Cascade file-mean cosine corpus-wide	0.5.1
Smaller default HyDE model + keep_alive=-1	0.6.0

Development

git clone https://github.com/danielchen26/skylakegrep.git
cd skylakegrep
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[rerank]"

.venv/bin/pytest -q tests/
.venv/bin/python benchmarks/agent_context_benchmark.py --top-k 10 --summary-only

To reproduce the public OSS benchmark numbers, follow the instructions in docs/parity-benchmarks.md.

License

PolyForm Noncommercial 1.0.0 — see LICENSE.

Personal, academic, research, hobby, and any other non-commercial use is fully permitted, including modification and redistribution. Commercial use is NOT permitted under this license. To obtain a commercial license, open a GitHub issue titled "Commercial license inquiry" or email [e-mail].

Acknowledgments

• Ollama for the local embedding and generation runtime.
• tree-sitter for syntax-aware parsing.
• ripgrep for the lexical prefilter stage.
• Mixedbread for the open-source mxbai-rerank-*-v2 cross-encoder family.
• nomic-embed-text for the embedding model.
• Click, NumPy, and SQLite for the core runtime dependencies.