A fast, local, and fully offline semantic search CLI & TUI for your terminal.
Index folders of codebases, documentation, or notes, and search them instantly using natural language.
Zero cloud APIs. Zero costs. 100% local and private.
Sift is a personal hobby project built for developers who want a lightweight, local alternative to cloud vector databases. By loading HuggingFace's BGE-small-en-v1.5 embeddings model locally via ONNX Runtime (CGo) and pairing it with a from-scratch HNSW vector graph index implemented in pure Go, Sift delivers semantic file search in less than 1ms traversal latency—completely offline.
- Offline Local Embeddings: BGE-small-en-v1.5 embeddings computed locally on your CPU via native ONNX Runtime inference.
- High-Speed HNSW Index: A highly optimized Hierarchical Navigable Small World (HNSW) graph implemented from scratch in Go (
M=16,ef=50), featuring O(1) visited bitsets and introsort routines for minimal allocations. - Hybrid Keyword Boosting: Intersects dense vector search scores with sparse keyword matches to improve search relevance for exact matches and short queries.
- Semantic Paragraph Chunking: Intelligent boundary chunking based on lines, markdown paragraphs (
\n\n), and code blocks instead of mechanical word splits. - Debounced TUI: A sleek interactive terminal interface built with BubbleTea featuring fuzzy instant search, spinner indicators, Vim navigation, and direct editor integrations.
- Incremental Watching: Multi-directory
fsnotifyfile watcher that debounces writes and dynamically updates the index on modification or creation. - No DB Dependencies: The complete index fits in a lightweight local folder (
.sift/) featuring a custom flat binary graph layout and a JSON metadata skip-cache.
Sift is built on a modular pipeline designed to be simple, clean, and self-contained:
Components
──────────
cmd/sift/ Cobra CLI subcommands (root, index, search, watch, tui, stats, clear, rebuild, bench, version)
internal/config non-global .sift.toml configuration parsing
internal/chunker streaming word-window text splitter, binary sniff
internal/embed ONNX session + tokenizer, EmbedDocs / EmbedQuery
internal/hnsw from-scratch HNSW graph + binary serialiser
internal/index ties chunker → embedder → HNSW, flush / load
internal/watcher fsnotify recursive dir watcher with debounce
internal/tui BubbleTea TUI (spinner · icons · vim nav · statusbar)
- Go 1.21+
make&curl- GCC (for native CGo tokenizer bindings)
Clone, fetch dependency assets, and build the binary:
git clone https://github.com/tejas242/sift
cd sift
# 1. Download native ONNX Runtime shared library (lib/onnxruntime.so)
make download-ort
# 2. Fetch the BGE-small-en-v1.5 model and config files
make download-model
# 3. Compile the production binary
make buildSift provides a simple command-line interface for indexing, searching, and managing your files.
# Index a directory recursively (creates a local .sift/ index folder)
./sift index ./docs
# Perform a quick semantic search — prints top-10 ranked chunks
./sift search "how does HNSW handle graph persistence"
# Get results formatted in JSON for integration with other shell tools (like jq)
./sift search --json "asymmetric retrieval prefix"
# Limit result pool size
./sift search --top-k 5 "vector dimensions"
# Quiet execution (suppress verbose logs from ONNX model loading)
./sift -q stats
# Launch the interactive BubbleTea TUI
./sift tui
# Monitor directory recursively and update the index in real-time
./sift watch ./docs
# Wipe your index and rebuild completely from scratch
./sift rebuild ./docs
# Check index file statistics and size
./sift stats
# Wipe index and remove index files
./sift clearSift parses a .sift.toml file in the current working directory to save your setup:
model-dir = "./models"
ort-lib = "./lib/onnxruntime.so"
threads = 0 # 0 = auto-detect optimal CPU core threads
max-file-kb = 512 # skip indexing files larger than 512KBWhen running ./sift tui, you enter a fully interactive terminal application:
| Key | Action |
|---|---|
Type anything |
Re-searches the index in real-time (debounced at 300ms) |
↑ / ↓ or k / j |
Navigate through search results |
Enter |
Open the selected file in your $EDITOR directly at the exact line number |
Ctrl+I |
Toggle index diagnostic statistics pane |
Esc |
Back to search view |
Ctrl+C / Ctrl+Q |
Exit Sift |
HNSW builds a hierarchical structure of navigable small world graphs. Nodes are inserted with an exponentially decaying layer probability. High layers act as a fast highway network containing a sparse subset of points, while layer 0 contains all points.
During a query:
- Search starts at the top layer, finding the local minimum.
- Descent goes down layer-by-layer, using the previous minimum as the entry point for the next.
- On Layer 0, a bounded beam search of size
efSearchis executed to collect the exact nearest neighbors.
We tune our pure-Go HNSW implementation for highly accurate retrieval bounds:
| Parameter | Value | Description |
|---|---|---|
M |
16 |
Maximum bidirectional links per node. |
efConstruction |
200 |
Size of dynamic candidate list built during insertion. |
efSearch |
50 |
Beam width size evaluated during a search query. |
Run locally on modest consumer hardware (AMD Ryzen 3 3250U @ 2.6GHz, CPU-only):
| Metric | Result |
|---|---|
| Recall@10 (1,000 vectors, BGE-small) | 90.6% |
| Graph Insertion Latency | ~3.6ms / vector |
| Graph Search Latency (ef=50) | < 0.7ms / query |
| Embedding Speed (BGE on CPU) | ~30-80ms / text chunk |
Even though this is a personal hobby project, I wanted to keep the engineering standards high! We've set up automated workflows:
- Build & Unit Tests (
ci.yml): Automatically compiles the code, checks for formatting/linting issues viago vet, and runs unit tests model-free using a mock embedder interface. - Continuous Benchmarking (
bench.yml): Tracks our customBenchmarkHNSWInsertandBenchmarkHNSWSearchexecution speeds over commits usinggithub-action-benchmarkto ensure no changes introduce performance regressions.
Sift is open-source software released under the MIT License.