Castform and Neon Use RL Post-Training to Beat Frontier Models in Retrieval

Castform and Neon have demonstrated a method to outperform frontier models, such as gpt-5.6-sol, in complex retrieval tasks by utilizing RL post-trained open-weights models. This approach allows enterprises to deploy specialized agents that maintain high performance while drastically reducing the cost and latency associated with general-purpose AI.

The system combines Castform's RL post-training pipeline with Neon's Lakebase Search infrastructure. According to the Neon Blog, small open-weights models used in this framework are approximately 100x cheaper than their frontier counterparts. For comparison, a typical multi-turn search request using gpt-5.6-sol takes over 10 seconds and costs roughly $0.03 end-to-end. By converting proprietary corporate corpora into synthetic question-answer datasets, Castform manages the RL loop to teach smaller models how to effectively navigate specific search tools.

The Shift to Agentic Retrieval

The industry is transitioning from simple one-shot Retrieval-Augmented Generation (RAG) pipelines toward "agentic retrieval." In this paradigm, models do not simply retrieve a single set of documents; they plan and execute multiple search loops to solve complex, multi-step problems. While this iterative process improves accuracy, it significantly increases the financial and temporal costs when using expensive, closed-source frontier models. This has created a critical demand for smaller, specialized models capable of handling these loops without the overhead of a general-purpose giant.

Implications for Enterprise AI

This development signals a broader shift toward specialized intelligence for enterprise tasks rather than a reliance on general intelligence. By utilizing Neon's Lakebase Search—which provides hybrid search via lakebase_text and lakebase_vector during both training and inference—companies can turn their own databases into training assets.

"Most teams' best training data is just sitting in their databases," says Ying Hang Seah, cofounder of Castform. Seah notes that the difficulty has historically been turning raw data into usable training sets and maintaining the infrastructure for agents to search and mutate data at scale. By integrating Castform with Neon, the economic barrier to deploying high-performance AI agents drops significantly, reducing the industry's dependence on a few closed-source providers.

Future Outlook

As the ability to post-train small models on proprietary data matures, the focus will likely shift toward the scalability of these specialized agents. The primary metric for success will move from raw model size to retrieval efficiency and the quality of the synthetic training loops. Observers will be watching to see if this specialized approach can maintain its edge as frontier models continue to evolve their own native retrieval capabilities.