In the HR vertical, standard search algorithms are practically useless: simply surfacing text with matching keywords isn't enough. According to Vadim Vakhrushev, Senior DS Engineer at Avito, the system must navigate a constant conflict of interest, balancing candidate relevance with the demands of employers who pay for promotion and the overall "fairness" of search results. To prevent 1.5 million job postings from devolving into a digital junkyard, engineers implemented a three-tier ranking pyramid—ranging from broad candidate generation to a final "blender" that shapes the user's feed.
Funnel mechanics: From rules to neural networks
The process is built on Multi-Task Learning (MTL). At the first stage (L1), "candidate generators" work to bypass rigid filters. For example, when someone searches for remote work in a remote location like Vorkuta, the system recognizes the lack of local offers and pulls vacancies from across the entire country. The ClickPred model, despite its name, does not optimize for empty clicks; it targets conversion actions like calls and messages. This is critical for blue-collar sectors, where users often ignore long benefit lists and prefer to contact the employer immediately. Textual integrity is managed by the RelPred model, trained on LLM-generated markup, while the "item_age" coefficient prevents top results from stagnating by giving new ads a chance to gain their first views.
Business logic vs. clean code
Semantic search and the inclusion of adjacent professions are more than just tech trends; they are an attempt to retain users who often don't know the official title of their own position. Stripping away the marketing jargon, this is a classic TCO and funnel optimization task: making a system with millions of positions run with minimal latency without burning through advertisers' budgets.
In our view, the use of LLM-based markup due to the "noisy" nature of user data is an ironic marker of our time: we have fully entered an era where one neural network explains to another what a human being was actually trying to find. It is the progress we deserve.
The bottom line
The system prioritizes high-intent actions (calls/messages) over clicks to suit blue-collar hiring habits. Multi-stage ranking allows the platform to balance paid promotion with organic relevance. Semantic matching bridges the gap between informal user queries and formal job descriptions.