Neon published a blog post yesterday (August 6), announcing a collaboration with Castform to disclose a training approach for retrieval models.
Note: Neon is a fully managed, serverless PostgreSQL database. It separates compute from storage, supports automatic scaling, scaling down to zero on demand, and Git-like database branching, making it well suited to modern cloud-native applications and AI agents.
Castform primarily provides AI voice agents and model training platforms for enterprises and developers. Through highly realistic voice interactions and automated workflows, it helps businesses handle routine calls, sales follow-ups, and customer support.
The blog post states that a good AI agent needs two capabilities: first, using tools to find the right contextual data; and second, routing and determining which model to use for the search.
Therefore, the Neon team worked with Castform to use reinforcement learning to post-train a 4B-parameter open-source model. On search tasks, the model's accuracy was comparable to, or even higher than, OpenAI's GPT-5.6 Sol, while costing approximately 1/100th as much.
The Neon team noted that the prevailing approach to document search is currently embedded search, which typically converts documents into numerical data and then looks for similar content.
As AI agents have become more widespread, however, search workflows have begun shifting toward agentic search. In these systems, AI models break large questions into multiple smaller ones, autonomously decide what to search for, review the search results, and initiate subsequent searches based on those results until they have gathered the required information.
Neon said agentic search is better suited to handling complex questions, but each search requires calling a high-performance model, which increases latency and cost. According to Neon’s typical request figures, a single search request using GPT-5.6 Sol takes more than 10 seconds and costs approximately $0.03.
Castform chose to train a relatively small open-source model to teach it how to find the required information from specific data. Neon provided the document storage location and search functionality, while Castform trained the model to determine “what to search for.”
During the training process, the model attempts to complete tasks, the system scores the results, and those scores are incorporated into subsequent trials. The evaluation criteria include whether the correct document was found, whether appropriate passages were cited, and whether the final answer was correct.
In Neon's evaluation, the model post-trained by Castform achieved an average evaluation score of 1.447, surpassing GPT-5.6 Sol's 1.369 and the 1.377 score of GPT-5.4, which was used as a control, setting a new record in the evaluation.
In terms of cost, the post-trained model's per-inference cost was $0.000929 (approximately 0.0063 yuan at the current exchange rate); GPT-5.6 Sol's per-inference cost was $0.087338 (approximately 0.59 yuan at the current exchange rate).