JJJJJJJJJJJ          AAA   VVVVVVVV           VVVVVVVV
          J:::::::::J         A:::A  V::::::V           V::::::V
          J:::::::::J        A:::::A V::::::V           V::::::V
          JJ:::::::JJ       A:::::::AV::::::V           V::::::V
            J:::::J        A:::::::::AV:::::V           V:::::V 
            J:::::J       A:::::A:::::AV:::::V         V:::::V  
            J:::::J      A:::::A A:::::AV:::::V       V:::::V   
            J:::::j     A:::::A   A:::::AV:::::V     V:::::V    
            J:::::J    A:::::A     A:::::AV:::::V   V:::::V     
JJJJJJJ     J:::::J   A:::::AAAAAAAAA:::::AV:::::V V:::::V      
J:::::J     J:::::J  A:::::::::::::::::::::AV:::::V:::::V       
J::::::J   J::::::J A:::::AAAAAAAAAAAAA:::::AV:::::::::V        
J:::::::JJJ:::::::JA:::::A             A:::::AV:::::::V         
 JJ:::::::::::::JJA:::::A               A:::::AV:::::V          
   JJ:::::::::JJ A:::::A                 A:::::AV:::V           
     JJJJJJJJJ  AAAAAAA                   AAAAAAAVVV            
                                                          
EEEEEEEEEEEEEEEEEEEEEELLLLLLLLLLL             IIIIIIIIII        
E::::::::::::::::::::EL:::::::::L             I::::::::I        
E::::::::::::::::::::EL:::::::::L             I::::::::I        
EE::::::EEEEEEEEE::::ELL:::::::LL             II::::::II        
  E:::::E       EEEEEE  L:::::L                 I::::I          
  E:::::E               L:::::L                 I::::I          
  E::::::EEEEEEEEEE     L:::::L                 I::::I          
  E:::::::::::::::E     L:::::L                 I::::I          
  E:::::::::::::::E     L:::::L                 I::::I          
  E::::::EEEEEEEEEE     L:::::L                 I::::I          
  E:::::E               L:::::L                 I::::I          
  E:::::E       EEEEEE  L:::::L         LLLLLL  I::::I          
EE::::::EEEEEEEE:::::ELL:::::::LLLLLLLLL:::::LII::::::II        
E::::::::::::::::::::EL::::::::::::::::::::::LI::::::::I        
E::::::::::::::::::::EL::::::::::::::::::::::LI::::::::I        
EEEEEEEEEEEEEEEEEEEEEELLLLLLLLLLLLLLLLLLLLLLLLIIIIIIIIII        
                                                   
                                NNNNNNNN        NNNNNNNN 
                                N:::::::N       N::::::N 
                                N::::::::N      N::::::N 
                                N:::::::::N     N::::::N 
                                N::::::::::N    N::::::N 
                                N:::::::::::N   N::::::N 
                                N:::::::N::::N  N::::::N 
                                N::::::N N::::N N::::::N 
                                N::::::N  N::::N:::::::N 
                                N::::::N   N:::::::::::N 
                                N::::::N    N::::::::::N 
                                N::::::N     N:::::::::N 
                                N::::::N      N::::::::N 
                                N::::::N       N:::::::N 
                                N::::::N        N::::::N 
                                NNNNNNNN         NNNNNNN 
        

vector space embeddings visualisation

Lance inspector and TUI for Arrow/Lance datasets.

Javelin is a Rust-based command-line and TUI tool for inspecting datasets stored in the Lance format (and compatible Parquet exports). It focuses on fast, ergonomic exploration of embedding-like matrices, sparse COO data, and 1D vectors using an interactive terminal UI built on ratatui and crossterm. Next step: graph visualisations, please star ⭐ the repo.

Download one of the pre-built packages in Releases.

Compile with cargo:

cargo install javelin-tui

Usage

Quickstart

# generate a toy dataset
javelin generate

# load the dataset in the tui
javelin --filepath ./javelin_test

# Select one of the supported files types
- Dense
- Sparse (COO, like adjacency matrices)
- 1D vectors (i.e. norms)

# Explore the different modes
[Head]
[Sample]
[Display]
[T] Transpose
[q] Exit

Basic CLI

# Show the first 20 rows
javelin --filepath /path/to/dataset.lance head --n 20

# Randomly sample 50 rows, preserving original indices
javelin --filepath /path/to/dataset.lance sample --n 50

# Open full dataset in TUI viewer
javelin --filepath /path/to/dataset.lance display

TUI launcher (default)

# Launcher for a directory of Lance datasets
javelin --filepath /path/to/dir

If you omit the subcommand:

• When filepath is a directory, the launcher scans it for .lance files.
• When filepath is a file, you can still use the launcher to choose a command.

Launcher key bindings

• Up / Down or k / j: Move selection between files.
• Left / Right or h / l: Cycle between commands (Head, Sample, Display, …).
• Enter: Run the selected command on the selected file.
• q / Esc: Exit the launcher.

Features

Single-entry TUI launcher

• javelin --filepath /path/to/dir starts a launcher that:
  • Scans a directory for .lance datasets.
  • Lets you select a file with Up/Down keys.
  • Lets you select a command (Head, Sample, Display, etc.) with Left/Right keys.
  • Launches the corresponding interactive viewer for the chosen file.

If no subcommand is provided, the default is the TUI launcher.

Dense matrix viewer

• Detects Lance “vector” layout and generic dense col_* layouts.
• Shows a scrollable table with:
  • A row index column.
  • Feature columns from col_*.
  • Per-row mean and standard deviation (for multi-column dense layouts).
• Supports:
  • Horizontal scrolling over features.
  • Vertical scrolling over rows.
  • A transposed view (features × samples) toggled via a key.

1D vector viewer

• Specialized UI for LanceLayout::Vector1D data (e.g. eigenvalues, norms).
• No avg/std columns; values are displayed with 12 decimal digits.
• Same navigation shortcuts as the dense viewer.

Sparse COO viewer

• Expects COO data in row, col, value schema.
• Layout:
  • Top: matrix metadata and density.
  • Middle:
    • Triples table with vertical scrolling over (row, col, value) entries.
    • ASCII sparsity map (downsampled for large matrices) that highlights nonzeros.
  • Bottom: diagonals and connectivity summaries (e.g., most connected rows).

Sampling and indexing

• cmd_sample:

  • Randomly selects n distinct row indices.
  • Reads the minimal prefix needed to cover those indices.
  • Uses Arrow take to build a sampled RecordBatch.
  • Adds a row_idx column with the original dataset indices.
  • Opens the sampled batch in the TUI viewer.

• cmd_head:

  • Shows the first n rows in the interactive viewer.

• cmd_stats:

  • Reports dataset row count and schema.
  • Prints per-column structural information.

Storage integration

• Uses a LanceStorage backend to:

  • Load dense matrices from Lance vector datasets.
  • Load sparse COO matrices from Lance triplet datasets.
  • Save dense matrices back as Lance vector datasets via save_dense("raw_input").

• Parquet support:

  • Reads Parquet into Arrow RecordBatches.
  • Detects:
    • Lance-like vector format (FixedSizeList<Float64>), or
    • Wide columnar float format with col_* columns.
  • Converts wide columnar Parquet to a dense matrix and saves it as Lance row‑major “raw_input” using save_dense.

Build from source

Prerequisites

• Rust (stable toolchain and cargo).
• A terminal that supports ANSI escape codes.

Build from source

git clone https://gitlab.com/yera/javelin.git
cd javelin
cargo build --release

The binary will be available at:

target/release/javelin

Interactive viewers

Dense and 1D viewers

Key bindings:

• Up / Down or k / j:
  • Scroll vertically over rows.
• Left / Right or h / l:
  • Scroll horizontally over feature columns (dense) or vector columns (1D).
• H:
  • Jump to the first visible column.
• E:
  • Jump to the last visible column window.
• t:
  • Toggle transpose (N×F ↔ F×N) in dense layouts.
• q / Esc:
  • Exit the viewer.

Behavior:

• Dense layouts show:

  • Synthetic “Row” index column.
  • col_* features.
  • Per-row avg and std computed over all numeric feature columns.

• 1D layouts show:

  • Row index.
  • One or more value columns with 12 decimal digits and no avg/std.

Sparse COO viewer

Key bindings are the same for scrolling:

• Up / Down or k / j: vertical scroll through triples.
• q / Esc: exit.

Panels:

• Metadata: matrix dimensions and density.
• Triples table: index, row, col, value with vertical scrolling.
• Sparsity map: ASCII grid marking nonzeros.
• Structure summary: main diagonal entries and most-connected rows.

Data formats

Dense Lance (vector) format

• Stored as a single FixedSizeList<Float64> column (e.g. vector).
• Reconstructed as a dense matrix in column-major order for computation.
• Displayed in the TUI as:
  • A dense matrix table, or
  • A 1D vector viewer for LanceLayout::Vector1D.

Sparse COO format

• Schema:

row:   UInt32
col:   UInt32
value: Float64

• Matrix dimensions stored in schema metadata (rows, cols, nnz).
• Reconstructed internally as a CSR matrix when needed.

Parquet import

• Vector-like: FixedSizeList<Float64> column(s).

• Wide columnar: multiple Float64 col_* columns.

• Wide columnar data is:

• Interpreted as a dense matrix.

• Saved via save_dense("raw_input") into Lance format.

• Then handled by the same TUI paths as native Lance data.

Typical workflows

Inspect an embedding matrix

javelin --filepath embeddings.lance display

• Scroll across feature dimensions.
• Toggle transpose to view feature-centric slices.
• Inspect per-row avg and std.

Random sample with original indices

javelin --filepath embeddings.lance sample --n 100

• Examine the row_idx column to see which original rows were sampled.

Explore a directory of experiments

javelin --filepath ./experiments

• Pick a dataset with Up/Down.
• Choose Head, Sample, or Display with Left/Right.
• Press Enter to launch the viewer.

Inspect sparse matrices

javelin --filepath graph.lance display

• Use the COO viewer to inspect structure, sparsity pattern, and connectivity.

License

See the LICENSE file in this repository for licensing details. Respect all third-party library licenses when redistributing binaries or integrating Javelin into other systems.