DEV Community: Andrew Markhai

AI-ассистент для документации из wiki Яндекса с использованием RAG и LangChain

Andrew Markhai — Sun, 26 Oct 2025 16:40:33 +0000

Собрал небольшой прототип чат-ассистента, который умеет отвечать на вопросы по внутренней документации.

Под капотом — классическая схема Retrieval-Augmented Generation (RAG): данные хранятся в векторной базе, а при каждом запросе к OpenAI подмешивается контекст из релевантных документов.

Проект заточен под wiki Яндекса, но если заменить парсер, можно использовать для любой другой базы знаний.

Как работает:

Парсинг документации — скрипт вытягивает нужные страницы из wiki и сохраняет их в .md.
Ингест — Markdown-файлы превращаются в векторы (через Sentence-Transformers или OpenAI embeddings) и индексируются в FAISS.
RAG-агент — при запросе ищет релевантные куски текста, добавляет их в промпт и отправляет в OpenAI API.
Всё это обёрнуто в простой CLI-интерфейс, можно общаться с ботом прямо из терминала.

Что умеет:

Отвечает на вопросы по документации
Находит и цитирует источники (https://wiki.yandex.ru/...)
Поддерживает русский язык (вопросы и ответы)

Как запустить расписывать здесь не буду, есть подробное README.md в репозитории.

Полезность сомнительная, но для экспериментов с RAG и векторными базами — пойдет.

Ну и, конечно, почти весь код написал ChatGPT, я только немного подкрутил под свои нужды.

PHP Traceroute

Andrew Markhai — Wed, 13 Aug 2025 18:46:10 +0000

A simple traceroute implementation in PHP, inspired by the article How Does Traceroute Work.
Result here: php-traceroute

How It Works
The idea is straightforward — send ICMP packets with increasing TTL values until:

We receive a reply from the target IP, or
The TTL reaches its maximum.

Currently, the script supports two parameters:

host (required) — target hostname or IP
max-hops (optional) — maximum TTL, defaults to 256

Known Issues & Open Questions

Timeout at TTL=2
There’s a noticeable delay after sending a packet with TTL=2. In Wireshark, I see no incoming data until an NBNS packet with Registration NB SMB_NSCHECK appears. Cause unknown.
Why UDP in standard traceroute?
Regular traceroute sends UDP packets, but UDP datagrams don’t carry TTL info themselves. How does TTL handling work in that case?

Notes & Implementation Details

Packet inspection Used Wireshark to confirm packet similarity with standard traceroute. Initially tried capturing only Docker container traffic to avoid noise, but getprotobyname() failed inside the container (missing /etc/protocols). Instead, I ran everything locally and filtered traffic with:

udp or icmp

Socket options
All socket_set_option parameters were found via Google. Network programming in PHP isn’t very common, so examples are scarce.
CLI handling
symfony/console is optional — I just didn’t want to deal with raw input/output.
Simplified packet count
Only one ICMP packet is sent per hop (instead of the usual three), so the output shows a single time in milliseconds.

Example Usage

php traceroute.php example.com --max-hops=30

The script works, so feel free to try it out for fun — just don’t expect production-grade traceroute magic.

How Symfony Flex Works

Andrew Markhai — Sun, 23 Mar 2025 15:46:10 +0000

What is it?

Symfony Flex is a Composer plugin from Symfony that handles the configuration of installed packages. It adds configuration files, updates .gitignore and .env, registers bundles, and more.

For example, when installing PHPUnit, Flex will:

Add a default configuration file phpunit.dist.xml
Create tests/bootstrap.php
Add /phpunit.xml and /.phpunit.cache/ to .gitignore
Create .env.test with environment variables for tests

Where does Flex get its instructions?

In Flex terminology, these instructions are called recipes. They are stored in a GitHub repository. You can check out the recipes for the current version of PHPUnit here.

Is this a complete list of recipes?

No. The symfony/recipes repository is maintained by Symfony itself, while community-contributed recipes are stored in a separate repository. Anyone can contribute there.

To allow Flex to apply community recipes, you need to add the following to composer.json:

"extra": {
    "symfony": {
        "allow-contrib": true
    }
}

How does it work under the hood?

When initialized, Flex loads links to all current recipes in Downloader::index. By default, it fetches them from two repositories described above:

private const DEFAULT_ENDPOINTS = [
'https://raw.githubusercontent.com/symfony/recipes/flex/main/index.json',
'https://raw.githubusercontent.com/symfony/recipes-contrib/flex/main/index.json',
];

However, you can add your own repositories by specifying them in composer.json under extra.symfony.endpoint.

Since Flex is a Composer plugin, it listens for various events:

$events = [
    PackageEvents::POST_PACKAGE_UPDATE => 'enableThanksReminder',
    PackageEvents::POST_PACKAGE_INSTALL => 'recordFlexInstall',
    PackageEvents::POST_PACKAGE_UNINSTALL => 'record',
    InstallerEvents::PRE_OPERATIONS_EXEC => 'recordOperations',
    PluginEvents::PRE_POOL_CREATE => 'truncatePackages',
    ScriptEvents::POST_CREATE_PROJECT_CMD => 'configureProject',
    ScriptEvents::POST_INSTALL_CMD => 'install',
    ScriptEvents::PRE_UPDATE_CMD => 'configureInstaller',
    ScriptEvents::POST_UPDATE_CMD => 'update',
    'auto-scripts' => 'executeAutoScripts',
];

During install and update commands, the Flex::install() method runs. It checks whether there are recipes for the current package by looking at both the package name and its version.

Recipe versioning is handled using subdirectories inside the package directory. For example, looking at PHPUnit’s recipes, you’ll see folders for versions 10.0, 9.6, 9.3, and 4.7. Flex selects the highest available recipe version that is less than or equal to the installed package version.

For example:

PHPUnit 12 will use the 10.0 recipe
PHPUnit 9.9 will use the 9.6 recipe

If a recipe is found, the instructions in its manifest are executed using various configurators. The console will display a message like:

Symfony operations: 1 recipe (13c8b05b36f03f13c596b688cb97d0da)
  - Configuring phpunit/phpunit (>=10.0): From github.com/symfony/recipes:main

Some files have been created and/or updated to configure your new packages.

Useful commands

To check the current state of recipes:

composer symfony:recipes

If a package has an updated recipe (marked as update available), you can update it with:

composer recipes:update vendor/package