Feder

What is feder

Feder is an javascript tool that built for understanding your embedding vectors, feder visualizes faiss, hnswlib and other anns index files, so that we can have a better understanding how anns work and what are high dimensional vector embeddings.

So far, we are focusing on the Faiss (only ivf_flat) index file and HNSWlib (hnsw) index file, we will cover more index types later.

Feder is written in javascript, and we also provide a python library federpy, which is based on federjs.

NOTE:

• In IPython environment, it supports users to generate the corresponding visualization directly.
• In other environments, it supports outputting visualizations as html files, which can be opened by the user through the browser with web service enabled.

Online demos

• Understanding vector embeddings with Feder by a reverse image search example
• Javascript example (Observable)
• Jupternotebook example (Colab)

How feder works

• Feder ivf layout
• Feder hnsw layout

Wiki

• Usage

HNSW visualization screenshots

IVF_Flat visualization screenshots

Quick Start

Installation

Use npm or yarn.

yarn install @zilliz/feder

Material Preparation

Make sure that you have built an index and dumped the index file by Faiss or HNSWlib.

Init Feder

Specifying the dom container that you want to show the visualizations.

import { Feder } from '@zilliz/feder';

const feder = new Feder({
  filePath: 'faiss_file', // file path
  source: 'faiss', // faiss | hnswlib
  domSelector: '#container', // attach dom to render
  viewParams: {}, // optional
});

Visualize the index structure.

• HNSW - Feder will show the top-3 levels of the hnsw-tree.
• IVF_Flat - Feder will show all the clusters.

feder.overview();

Explore the search process.

Set search parameters (optional) and Specify the query vector.

feder
  .setSearchParams({
    k: 8, // hnsw, ivf_flat
    ef: 100, // hnsw (ef_search)
    nprobe: 8, // ivf_flat
  })
  .search(target_vector);

Examples

We prepare a simple case, which is the visualizations of the hnsw and ivf_flat with 17,000+ vectors that embedded from VOC 2012).

Only need enable a web service.

git clone [email protected]:zilliztech/feder.git
cd test
python -m http.server

Then open http://localhost:8000/

It will show 4 visualizations:

• hnsw overview
• hnsw search view
• ivf_flat overview
• ivf_flat search view

Pipeline - explore a new dataset with feder

Step 1. Dataset preparation

Put all images to test/data/images/. (example dataset VOC 2012)

You can also generate random vectors without embedding for index building and skip to step 3.

Step 2. Generate embedding vectors

Recommend to use towhee, one line of code to generating embedding vectors!

We have the encoded vectors ready for you.

Step 3. Build an index and dump it.

You can use faiss or hnswlib to build the index.

(*Detailed procedures please refer to their tutorials.)

Referring to test/data/gen_hnswlib_index_*.py or test/data/gen_faiss_index_*.py

Or we have the index file ready for you.

Step 4. Init Feder.

import { Feder } from '@zilliz/feder';
import * as d3 from 'd3';

const domSelector = '#container';
const filePath = [index_file_path];

const mediaCallback = (rowId) => mediaUrl;

const feder = new Feder({
  filePath,
  source: 'hnswlib',
  domSelector,
  viewParams: {
    mediaType: 'img',
    mediaCallback,
  },
});

If use the random_data, no need to specify the mediaType.

import { Feder } from '@zilliz/feder';
import * as d3 from 'd3';

const domSelector = '#container';
const filePath = [index_file_path];

const feder = new Feder({
  filePath,
  source: 'hnswlib',
  domSelector,
});

Step 5. Explore the index!

Visualize the overview

feder.overview();

or visualize the search process.

feder.search(target_vector[, targetMediaUrl]);

or randomly select an vector as the target to visualize the search process.

feder.searchRandTestVec();

More cases refer to the test/test.js

Blogs

• Visualize Your Approximate Nearest Neighbor Search with Feder
• Visualize Reverse Image Search with Feder

Roadmap

We're still in the early stages, we will support more types of anns index, and more unstructured data viewer, stay tuned.

Acknowledgments

• faiss
• hnswlib
• d3