Iris Search

Iris Search is a hybrid RAG system, that allows you to intake any form of content and turn it into text and natural language searchable content.

Features

• Natural Language Search (Vector Search)
• Direct text search (FTS5 Text Search)
• The entire database is a single macOS package with the extension: irisdb
  • Easily transferrable between computers, just drag and drop a single file.

Supported Formats

• PDF
• TXT (Generic for all plaintext formats)
• HTML
• XML
• OPML

Supported Languages

• English

Installation

Swift Package Manage

Add this package to your Package.swift:

.package(url: "https://github.com/impel-intelligence/IrisSearch", from: "2.2.0")

Building

swift build

Testing

Iris Search uses swift testing for all tests.

swift test

Without Network

swift test --skip network

Without any files from git LFS

swift test --skip lfs

Embedding Models

Some default embeddings are provided in the regular IrisSearch package. More are provided in the Embedders package, which links to swift-embeddings.

CoreML Embedder

The CoreML embedder allows you to use a pre-compiled (.mlmodelc) file as an embedding model. This model must have the following inputs input_ids (int32), attention_mask (int32), token_type_ids (int32). These are based on the input into BERT models.

CoreML Model Configuration

To configure the output of a CoreML model you need to provide a config.json file alongside your .mlmodelc file. You must also provide the vocab.txt for the model you are working with.

The config file should have the same structure and types as this json object:

  {
    "tokenizerClass": "<tokenizer_class>",
    "maximumInputCharactersPerWord": int,
    "cleanText": true | false,
    "handleChineseCharacters": true | false,
    "stripAccents": true | false | null,
    "lowercase": true | false | null,
    "searchPrefix": "<search_prefix>" | null,
    "dimensions": int
}

1