More than 20 million listings pass through Avito's automated systems daily, setting a new benchmark for scaling content control on major online platforms. At such volumes, traditional manual moderation methods and cumbersome infrastructure loops begin to fail under the weight of operating expenses and human error. Developers are seeking ways to transfer verification logic into language models without compromising unit economics and service stability.
As Ruslan Sakaev, DS Team Lead at Avito, noted, the company processes over 20 million listings daily, with more than 95% of checks already automated. The remaining stream is directed to additional verification, handled by several hundred moderators. Manual review of massive data arrays runs up against the limits of human perception: rulebooks reach 200 pages, and a new employee takes about six months to reach a productive plateau in quality. Attempts to stuff such volumes of rule texts entirely into a language model's context inevitably lead to drops in processing speed, attention degradation, and rising inference costs. Under a load of millions of checks per day, every extra hundred tokens in the context window or heavy search loop like RAG multiplies across traffic volumes, turning into substantial compute expenses. This is precisely why engineers count every millisecond of latency and every cent of infrastructure costs, balancing filtering quality against strict unit-economics constraints.
Automating rules via compact prompts
Avito's team consciously abandoned RAG and full rulebooks in context due to risks of model attention degradation and redundant infrastructure overhead. Instead of building a separate search system, the engineers translated the rulebooks into compact structured prompts, turning them into an algorithmically optimizable hyperparameter.
We considered three options: RAG, the complete rulebook in context, and a prompt assembled from the rulebook.
This approach maintains low request costs and predictable processing speeds without deploying complex auxiliary loops. Verification rules are formulated in a strict format where the title, description, and photographs are fed as inputs, and the model returns a definitive answer without additional explanations.
Architecture selection for different violation types
Legal risks and security requirements preclude sending parts of the data to external cloud services, keeping the data loop strictly within the company's internal infrastructure. Due to the volatility of rules and the continuous emergence of new circumvention schemes, constantly fine-tuning models for every single change proves inefficient. Therefore, when mass-migrating verifications to LLMs, Avito's team takes the out-of-the-box 35B model Qwen3.5-35B-A3B as their primary working tool.
For complex violations where baseline performance falls short, the company employs fine-tuning of its proprietary Avision-8B model, which demonstrates an F1 score of 76%. The company continues testing new architectures on its tasks, evaluating their applicability for production workloads. The results indicate that a hybrid approach dividing tasks between ready-to-use mid-sized models and specialized fine-tuned versions maintains a strict balance between accuracy and compute costs.
An analysis of such cases demonstrates that the key success factor when handling million-scale data streams is abandoning universal heavy solutions in favor of pragmatic engineering. When dealing with tens of millions of daily transactions, inference economics dictates strict rules: the architecture must pay for itself on every single request, and any auxiliary modules like vector databases or search engines are justified only when their cost is fully covered by a reduction in error rates.
What came before
According to the company, Avito automated systems already process over 95% of all checks without routing data through third-party cloud providers. Can Avito keep moderation accurate with off-the-shelf Qwen3.5-35B-A3B and compressed prompts when fraud schemes evolve faster than multi-page rulebooks can be rewritten?