LinkedIn has officially launched its highly anticipated AI-powered people search feature, marking a significant evolution from traditional keyword-based search methods to a more sophisticated, intent-driven discovery process. This rollout comes three years after the introduction of ChatGPT and just six months following LinkedIn’s launch of its AI job search functionality. For enterprise leaders and AI practitioners, this timeline underscores a critical lesson: deploying generative AI in real-world enterprise environments is fraught with challenges, particularly when scaling to accommodate a user base of 1.3 billion.
The new people search feature allows users to input natural language queries directly into LinkedIn’s search bar. For instance, a user can ask, “Who is knowledgeable about curing cancer?” In contrast to the previous keyword-based search system, which would have struggled to provide relevant results by only identifying profiles that mentioned “cancer,” the new AI system understands the semantic meaning behind the query. It recognizes that “cancer” is related to terms like “oncology” and “genomics research,” enabling it to surface a more relevant list of professionals, including oncology experts and researchers, even if their profiles do not explicitly mention “cancer.”
This shift from a rigid keyword search to a nuanced understanding of user intent represents a significant advancement in how LinkedIn connects users with potential contacts. The AI-powered system not only prioritizes relevance but also considers the usefulness of the connections. Instead of merely presenting the top oncologist globally—who may be an unreachable third-degree connection—the system weighs the relevance of first-degree connections who might serve as valuable intermediaries to those experts. This dual focus on relevance and usefulness enhances the overall user experience, making it easier for individuals to find the right contacts within their professional networks.
The development of this AI-powered people search is rooted in a comprehensive “cookbook” approach that LinkedIn has meticulously crafted over time. This replicable, multi-stage pipeline emphasizes distillation, co-design, and relentless optimization. LinkedIn’s engineering team recognized early on that attempting to build a unified system for all of its products was overly ambitious and ultimately stalled progress. Instead, they focused on perfecting one vertical before applying the same principles to another. The success of the AI job search feature, which reportedly increased the hiring likelihood of job seekers without a four-year degree by 10%, provided a solid foundation for the new people search initiative.
To tackle the complexities of scaling the people search feature, LinkedIn’s engineering team began with a “golden dataset” comprising a few hundred to a thousand real query-profile pairs. These pairs were meticulously scored against a detailed product policy document spanning 20 to 30 pages. This small yet high-quality dataset served as the basis for generating a massive volume of synthetic training data using a large foundation model. The synthetic data was instrumental in training a 7-billion-parameter “Product Policy” model, which acted as a high-fidelity judge of relevance. Although this model was too slow for live production, it proved invaluable for teaching smaller models.
However, the team encountered significant challenges during the initial phases of model training. For six to nine months, they struggled to develop a single model capable of balancing strict adherence to policy (relevance) with user engagement signals. The breakthrough came when the team realized the need to decompose the problem. They distilled the 7-billion-parameter policy model into a 1.7-billion-parameter teacher model focused solely on relevance. This teacher model was then paired with separate models trained to predict specific member actions, such as job applications for the jobs product or connecting and following for people search. This “multi-teacher” ensemble produced soft probability scores that the final student model learned to mimic through KL divergence loss.
The resulting architecture operates as a two-stage pipeline. The first stage involves a larger 8-billion-parameter model responsible for broad retrieval, casting a wide net to pull candidates from the extensive user graph. The second stage features a highly distilled student model that takes over for fine-grained ranking. While the job search product successfully deployed a 600-million-parameter student model, the new people search product required even more aggressive compression. The engineering team pruned their new student model from 440 million parameters down to just 220 million, achieving the necessary speed for 1.3 billion users while maintaining less than 1% relevance loss.
Transitioning to the people search feature necessitated a fundamental architectural shift. The previous retrieval stack was built on CPUs, but to meet the new scale and latency demands for a “snappy” search experience, the team migrated its indexing to GPU-based infrastructure. This foundational change was essential to handle billions of records efficiently.
Organizationally, LinkedIn benefited from a collaborative approach. Initially, two separate teams—one focused on job search and the other on people search—were working in parallel to solve the respective problems. However, once the job search team achieved its breakthrough using the policy-driven distillation method, leadership intervened to bring over the architects of the job search success. Product lead Rohan Rajiv and engineering lead Wenjing Zhang were tasked with transplanting their “cookbook” directly to the new domain of people search.
With the retrieval problem addressed, the team turned its attention to the challenges of ranking and efficiency. This is where the cookbook was adapted with new, aggressive optimization techniques. One of the most significant optimizations involved input size. To enhance the model’s performance, the team trained another large language model (LLM) specifically designed to summarize the input context. This summarizer model was able to reduce the model’s input size by an impressive 20-fold with minimal information loss. The combined effect of the 220-million-parameter model and the 20x input reduction resulted in a remarkable 10x increase in ranking throughput, allowing the team to serve the model efficiently to its vast user base.
Throughout the development process, Erran Berger, VP of Product Engineering at LinkedIn, emphasized the importance of pragmatism over hype. He articulated a clear vision that the real value for enterprises lies in perfecting recommender systems rather than chasing the allure of “agentic” AI products. Berger refrained from discussing specific models used for the searches, suggesting that the choice of models should be based on efficiency for the task at hand.
The new AI-powered people search embodies Berger’s philosophy that optimizing the recommender system should take precedence. The architecture includes an intelligent query routing layer powered by LLMs, which pragmatically determines whether a user’s query—such as “trust expert”—should be directed to the new semantic, natural-language stack or the older, reliable lexical search. This complex system is designed to function as a tool that future agents will utilize, rather than being an agent itself.
As the people search feature becomes available, Berger hinted at the possibility of offering agents to leverage this functionality in the future, although he did not provide specific timelines. Furthermore, he indicated that the successful recipe developed for both job and people search would be extended across LinkedIn’s other products.
For enterprises looking to build their own AI roadmaps, LinkedIn’s journey offers several key takeaways:
1. **Be Pragmatic**: Focus on winning one vertical at a time, even if it takes considerable time and effort. Trying to tackle multiple challenges simultaneously can lead to stagnation and inefficiency.
2. **Codify Your Process**: Transform successful initiatives into repeatable processes. This includes developing comprehensive policy documents, establishing distillation pipelines, and engaging in co-design efforts.
3. **Optimize Relentlessly**: The most significant gains often come after the initial model deployment. Prioritize pruning, distillation, and innovative optimizations, such as employing reinforcement learning to enhance model performance.
LinkedIn’s experience illustrates that for real-world enterprise AI, the emphasis should not be solely on specific models or the latest trends in agentic systems. Instead, the strategic advantage lies in mastering the underlying pipeline—the “AI-native” cookbook of co-design, distillation, and rigorous optimization.
As the landscape of AI continues to evolve, LinkedIn’s approach serves as a masterclass in scaling generative AI effectively in the enterprise context. By focusing on pragmatic solutions, leveraging existing successes, and continuously optimizing their systems, LinkedIn has positioned itself as a leader in the integration of AI technologies within professional networking. The implications of this launch extend beyond LinkedIn, offering valuable insights for organizations across various sectors seeking to harness the power of AI to enhance their operations and user experiences.