# What are the two main ways to integrate retrieval with an LLM (prompting a frozen model with external info versus fine-tuning the model on a corpus), and what are the benefits of each approach?

Source: https://zilliz.com/ai-faq/what-are-the-two-main-ways-to-integrate-retrieval-with-an-llm-prompting-a-frozen-model-with-external-info-versus-finetuning-the-model-on-a-corpus-and-what-are-the-benefits-of-each-approach
Author: zilliz

The two main ways to integrate retrieval with a large language model (LLM) are **prompting a frozen model with external information** and **fine-tuning the model on a specific corpus**. Each method addresses different needs, depending on factors like data freshness, computational resources, and domain specificity. Here’s a breakdown of how they work and their benefits:

**Prompting a frozen model with external information** involves injecting retrieved data (e.g., documents, database entries) directly into the input prompt while keeping the LLM’s parameters unchanged. For example, a question-answering system might fetch relevant text snippets from a knowledge base and include them in the prompt to guide the model’s response. The key benefit is **flexibility**: since the model isn’t modified, you can update or swap the external data source without retraining. This is useful for dynamic or time-sensitive information, like news articles or real-time analytics. It’s also **cost-effective** because it avoids the computational overhead of fine-tuning. Tools like Retrieval-Augmented Generation (RAG) use this approach to combine retrieval and generation efficiently. However, the model’s reliance on prompt context limits its ability to deeply internalize domain-specific patterns.

**Fine-tuning the model on a corpus** involves training the LLM on a dataset specific to a domain (e.g., medical journals, legal contracts) to adapt its weights to the target task. This allows the model to **internalize patterns** from the data, improving accuracy and coherence in specialized contexts. For instance, a model fine-tuned on technical documentation will generate more precise answers in software engineering scenarios compared to a general-purpose model. Fine-tuning also reduces prompt length and inference latency because the model doesn’t require external context to perform well. However, it demands **significant resources** for training and may struggle with data that changes frequently, as updates require retraining. It’s best suited for static, domain-specific tasks where consistency and depth of knowledge are critical.

Choosing between these approaches depends on the use case. Prompting with external data works well for scenarios requiring up-to-date information and minimal upfront investment. Fine-tuning is better for specialized domains where the model must deeply understand terminology and context. Developers often combine both methods—using fine-tuning to adapt the model to a domain and retrieval to inject dynamic data—to balance specialization and flexibility.