elephant-php-neuron

A Neuron AI FileDataLoader reader for .docx documents, powered by elephant-php.

Drop it into a Neuron RAG pipeline and .docx files become embeddable documents alongside the bundled PDF, HTML and plain-text readers.

Installation

composer require endless-creativity/elephant-php-neuron

Requires PHP 8.2+. No external binaries needed (unlike PdfReader).

Usage

use NeuronAI\RAG\DataLoader\FileDataLoader;
use EndlessCreativity\ElephantPhpNeuron\DocxReader;

$documents = FileDataLoader::for(__DIR__.'/knowledge')
    ->addReader('docx', new DocxReader())
    ->getDocuments();

MyRAG::make()->addDocuments($documents);

Pass a directory and Neuron walks it, picking the right reader per extension; pass a single file to ingest just that one.

Output format

By default the reader returns plain text via Converter::extractRawText() — paragraphs separated by "\n\n", no markup. This is usually what you want for embeddings: less syntactic noise, more semantic signal per token.

If you'd rather preserve headings, lists and links — for example because your splitter or post-processor relies on Markdown structure — request Markdown explicitly through the reader options:

FileDataLoader::for($path)
    ->addReader('docx', new DocxReader())
    ->getDocuments(['format' => DocxReader::FORMAT_MARKDOWN]);

The $options array is forwarded by FileDataLoader to every reader, so the same flag is in effect for the whole loading pass.

Limitations

• Only OOXML .docx is supported. Legacy binary .doc (Word 97–2003) is not handled by elephant-php and therefore not by this reader either.
• Images embedded in the document are dropped during text extraction. This is intentional for RAG — embeddings are text-only.
• Conversion warnings emitted by elephant-php (Result::messages) are currently silenced. If you need them, open an issue.

License

BSD-2-Clause. See LICENSE.