Abstract
In general, traditional recommendation frameworks may rely on lagging and sparse data profiles. This reliance may limit the contextual relevance of the recommendations and may degrade cold start performance. The disclosed technology provides a unified generative and agentic recommendation framework (an agentic recommendation framework) that addresses these limitations. The agentic recommendation framework may use multimodal perception to process non-textual cues and to extract granular stylistic preferences of a user of a television application. Non-textual cues may be non-textual sensory inputs that may include, but are not limited to, images, video, and audio inputs.
The agentic recommendation framework may use multimodal perception to process the non-textual cues and to extract granular stylistic preferences of a user of a television application. The multimodal perception may process and interpret the non-textual sensory cues using Large Language Models (LLMs) to capture detailed, nuanced stylistic preferences that traditional recommendation models that use text-based metadata may not capture. Real-time contextual data, such as fluctuating prices and local events, may be injected into the generation pipeline using retrieval-augmented generation (RAG).
Furthermore, generative reasoning with semantic identifiers may be utilized to predict subsequent items in a sequence of items. A multi-agent framework included in the agentic recommendation framework may orchestrate the generation of candidate recommendations, the semantic re-ranking of the recommended candidates, and an output critique for the candidate recommendations. A continuous feedback loop may be established without requiring manual model retraining. Candidates for recommendations may be continuously updated with zero-latency data accuracy. The disclosed technology may deliver deeply personalized and explainable recommendations to a user by utilizing autonomous agents for continuous pipeline improvement, minimizing or reducing manual model retraining while recommendation relevance and transparency are increased.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Nayak, Shravan, "Unified Generative and Agentic Recommendation Framework", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10386