LinkedIn Unveils Advanced AI-Driven People Search: A New Era in Enterprise AI
This week, LinkedIn introduces its cutting-edge AI-powered people search feature, marking a significant milestone after a prolonged development period. This innovation leverages generative AI to transform how users discover expertise within the platform’s vast network.
From Concept to Reality: The Journey Behind LinkedIn’s AI People Search
Although generative AI has been mainstream since the debut of ChatGPT three years ago, LinkedIn’s rollout of this people search capability arrives six months after its AI-enhanced job search feature. This timeline underscores a critical insight for enterprise technology leaders: integrating generative AI at the scale of over 1.3 billion users is a complex, iterative process demanding meticulous refinement and practical engineering.
Insights from exclusive discussions with LinkedIn’s product and engineering teams reveal the strategic approach taken to develop this feature.
How the AI-Powered People Search Transforms User Queries
Users can now enter natural language questions such as “Who has expertise in cancer treatment?” directly into LinkedIn’s search bar. Unlike the previous keyword-based system, which would only match exact terms like “cancer,” the new AI understands the underlying intent and semantic relationships.
For example, the AI recognizes that “oncology” and “genomics research” are closely related to cancer treatment, enabling it to surface relevant professionals even if their profiles don’t explicitly mention the word “cancer.” This semantic comprehension vastly improves the relevance of search results.
Moreover, the system intelligently balances relevance with accessibility. Instead of listing only top-tier experts who might be distant connections, it prioritizes individuals within your immediate network who can act as valuable intermediaries, enhancing the practical utility of the search.
Strategic Focus: Mastering One Vertical Before Expanding
LinkedIn’s approach reflects a deliberate strategy of focusing on one domain at a time. After the success of its AI job search-which notably increased hiring chances by 10% for candidates without four-year degrees-the company applied the same methodology to the more complex people search challenge.
Erran Berger, VP of Product Engineering, highlights the difference in scale: managing tens of millions of job listings is one challenge, but extending AI-powered search across a network of over a billion members requires a fundamentally different level of engineering rigor.
Building on a Proven Foundation: The Multi-Stage AI Pipeline
The people search product builds upon a robust framework developed for job search. This framework begins with a meticulously curated “golden dataset” of several hundred to a thousand real query-profile pairs, evaluated against a comprehensive 20-30 page product policy document.
To scale training, LinkedIn used this dataset to generate extensive synthetic data via a large foundational model. This synthetic data trained a 7-billion-parameter “Product Policy” model, which excels at judging relevance but is too resource-intensive for real-time use.
Facing challenges in balancing strict policy adherence with user engagement, the team innovated by decomposing the problem. They distilled the large model into a 1.7-billion-parameter teacher model focused on relevance, complemented by additional teacher models predicting user actions like connection requests or job applications. This ensemble approach enabled the final student model to learn nuanced ranking through soft probability scores.
Scaling Retrieval and Ranking for Over a Billion Users
The architecture employs a two-step process: an 8-billion-parameter model performs broad candidate retrieval, followed by a highly compressed student model that fine-tunes ranking. For people search, the student model was aggressively pruned from 440 million to 220 million parameters, achieving the necessary speed for LinkedIn’s massive user base with less than 1% loss in relevance.
Unlike job search, people search required a fundamental shift in infrastructure. The team transitioned from CPU-based indexing to GPU-powered systems to meet the latency demands of a responsive search experience across a billion-plus profiles.
Organizationally, LinkedIn merged expertise from separate job and people search teams, bringing in key leaders who had pioneered the policy-driven distillation approach to ensure a smooth transfer of knowledge and best practices.
Innovative Optimizations: Achieving a Tenfold Throughput Increase
To further enhance efficiency, the team developed an additional LLM trained via reinforcement learning to summarize input context, reducing input size by 20 times with minimal loss of information. This innovation, combined with the compact 220-million-parameter model, resulted in a 10x boost in ranking throughput, enabling seamless service to LinkedIn’s extensive user base.
Prioritizing Practical AI Tools Over Agentic Hype
Erran Berger emphasizes that the true enterprise value lies in refining recommender systems rather than pursuing flashy agentic AI products. The new people search incorporates an intelligent query routing layer powered by LLMs, which pragmatically directs queries either to the semantic AI stack or the traditional keyword-based search, depending on which is more effective.
This design philosophy treats the AI system as a powerful tool to be leveraged by future agents, rather than an autonomous agent itself. Berger notes, “Even the best reasoning model can’t compensate for a weak underlying search engine.”
While LinkedIn plans to eventually integrate agentic capabilities, the current focus remains on perfecting the foundational AI infrastructure. The successful methodologies from job and people search are slated to be extended across other LinkedIn products.
Key Takeaways for Enterprise AI Development
- Adopt a pragmatic approach: Concentrate on mastering one vertical at a time, even if it requires a year or more.
- Develop a repeatable framework: Document policies, establish distillation pipelines, and engage in co-design to create a scalable process.
- Commit to continuous optimization: Major performance gains come after initial deployment through model pruning, distillation, and creative techniques like RL-based summarization.
LinkedIn’s experience demonstrates that sustainable competitive advantage in enterprise AI stems from perfecting the end-to-end pipeline rather than chasing the latest model architectures or agentic trends.
Looking Ahead
As LinkedIn’s AI people search rolls out, it sets a new standard for large-scale, practical AI deployment in professional networking. Enterprises aiming to build their own AI solutions can draw valuable lessons from LinkedIn’s methodical, data-driven, and user-centric approach.
