Increase Recommendation Systems’ Precision with LLMs, Using Python

2026-06-08 · Source: Towards Data Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Intermediate, medium

Summary

This article details a two-stage recommendation system designed to enhance precision using Large Language Models (LLMs) while managing computational costs. The system first employs a cheap, rule-based distance filter (Stage 1) to narrow down a dataset of approximately 10,000 restaurants across eight cities to 50 geographically closest candidates. Subsequently, a powerful LLM (Stage 2) reranks these 50 candidates based on a user's natural language query, providing 5 to 10 highly precise recommendations with a "fit_score" and explanation. This hybrid approach balances scalability and intelligence, leveraging LLM capabilities only on a pre-filtered, relevant subset to optimize cost and performance.

Key takeaway

For AI Engineers or Machine Learning Engineers designing intelligent recommendation systems, you should adopt a two-stage funnel to balance LLM costs with precision. By first using a cheap, rule-based filter to generate high-recall candidates, you can then apply an LLM for high-precision reranking on a significantly smaller dataset. This strategy ensures your system is both scalable and intelligent, making efficient use of expensive LLM resources.

Key insights

Optimize LLM-powered recommendation systems by combining cheap, high-recall filtering with expensive, high-precision reranking.

Principles

• System design involves trade-offs, often described by the Accuracy-Scale-Time triangle.
• Initial candidate generation should prioritize high recall and low precision.
• Final selection benefits from high precision models applied to a reduced candidate set.

Method

Generate synthetic restaurant data. Implement a rule-based filter to select N_DISTANCE_CANDIDATES (e.g., 50) based on proximity. Use an LLM with structured output (e.g., Pydantic) to rerank these candidates, providing fit scores and reasons.

In practice

• Use simple, rule-based methods for initial data reduction to minimize LLM API calls.
• Apply LLMs for nuanced interpretation and precise ranking on smaller, pre-filtered datasets.
• Enforce structured LLM responses using tools like Pydantic for reliable parsing.

Topics

• Recommendation Systems
• Large Language Models
• System Design
• Hybrid AI Architectures
• Data Filtering
• Pydantic
• OpenAI API

Code references

• PieroPaialungaAI/RestaurantLLM

Best for: AI Engineer, Machine Learning Engineer, AI Student

Related on AIssential

• Semantic Search At LinkedIn, LLM-Driven Autonomous Optimization for Industrial-Scale Recommendation Systems, and More! — Recent advancements in information retrieval focus on LLM-driven optimization, efficient embeddings, and scalable recommendation systems.
• Using GPT-4o-mini as an Entity Resolution Judge: 95% Precision for $0.04 — LLMs can act as cost-effective precision filters in entity resolution pipelines, especially for complex product data.
• Issue #115 - Reranking in your RAG pipeline — Reranking enhances RAG pipeline precision by reordering initially retrieved documents using a Cross-Encoder.
• Why Adding More Context Makes LLMs Less Reliable — Excessive or unstructured context degrades LLM reliability by creating competing signals that models cannot effectively prioritize.
• From Heuristic Selection to Automated Algorithm Design: LLMs Benefit from Strong Priors — Integrating benchmark code examples into prompts significantly enhances LLM-driven algorithm optimization performance and robustness.
• 2025 Year in Review — Integrating LLMs with RecSys can yield powerful, steerable recommendation systems capable of cross-platform suggestions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards Data Science.