{"id":27966,"date":"2024-04-25T12:00:00","date_gmt":"2024-04-25T10:00:00","guid":{"rendered":"https:\/\/golem.ai\/?p=27966"},"modified":"2024-09-17T11:59:15","modified_gmt":"2024-09-17T09:59:15","slug":"ia-rag-llm","status":"publish","type":"post","link":"https:\/\/miralia.ai\/en\/blog\/ia-rag-llm","title":{"rendered":"Is RAG enough?\u00a0"},"content":{"rendered":"

Retrieval-Augmented Generation (RAG) models combine the capabilities of the LLMs<\/a> (Large Language Models) with the extraction of information from a database or external corpus to answer questions or generate text. This approach overcomes some of the limitations of LLMs, notably in terms of accuracy of information, relevance of answers, updating of knowledge and limiting the \"hallucinations\" that LLMs can have when answering a question where they have no training data. In recent times, RAGs have often been presented as THE solution to these possible shortcomings of LLMs. But is it really enough?\n\nTranslated with DeepL.com (free version) <\/p>\n\n\n\n

What is RAG? <\/strong><\/h2>\n\n\n\n

RAG captivated the Generative AI developer community following the publication of the article entitled \" Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks<\/a> \", written by Patrick Lewis and his team at Facebook AI Research in 2020. This approach has been rapidly adopted by many researchers, both in academia and industry, because of its potential to significantly enrich the capabilities of generative AI systems. <\/p>\n\n\n\n

RAG est l’acronyme de \u00ab Retrieval Augmented Generation \u00bb. Cette approche fusionne des m\u00e9thodes d’extraction d’informations et de g\u00e9n\u00e9ration de contenus par l’intelligence artificielle. Les techniques d’extraction sont int\u00e9ressantes pour r\u00e9cup\u00e9rer des donn\u00e9es de diverses sources en ligne telles que des articles ou des bases de donn\u00e9es, mais elles se limitent \u00e0 reproduire des informations d\u00e9j\u00e0 existantes sans ajouter de nouveaut\u00e9s . \u00c0 l’oppos\u00e9, les mod\u00e8les d\u2019IA g\u00e9n\u00e9ratives sont capables de cr\u00e9er des contenus nouveaux et contextuellement adapt\u00e9s, bien qu’ils puissent parfois manquer de pr\u00e9cision. Ainsi, le mod\u00e8le RAG est n\u00e9 de l’ambition d\u2019allier les avantages de ces deux mondes : il utilise l’extraction pour identifier les informations les plus pertinentes dans les sources disponibles, puis le mod\u00e8le de g\u00e9n\u00e9ration transforme ces \u00e9l\u00e9ments en r\u00e9ponses compl\u00e8tes et pertinentes, surmontant ainsi les limites de chaque approche prise isol\u00e9ment. Dans le cadre d’un syst\u00e8me RAG, l’extraction cible les donn\u00e9es n\u00e9cessaires, tandis que la g\u00e9n\u00e9ration les reformule en une r\u00e9ponse claire et pr\u00e9cise, adapt\u00e9e \u00e0 la demande.\u00a0<\/p>\n\n\n\n

Les diff\u00e9rents avantages de la RAG<\/strong><\/h2>\n\n\n\n

Up-to-date information<\/strong><\/h3>\n\n\n\n

LLMs are trained on large datasets which, once the training process is complete, are no longer updated. This means that, even if an LLM includes information up to a certain date, any new information or events occurring after that date are not incorporated into the model. On the other hand, RAG models can consult continuously or periodically updated external databases or corpora to provide current information. For example, if a user asks a question about the latest advances in a specific scientific field, a RAG model can retrieve and integrate the results of research published after the last LLM model update, ensuring that the answer reflects the current state of knowledge.<\/p>\n\n\n\n

Data accuracy<\/strong><\/h3>\n\n\n\n

An LLM can generate responses based on patterns learned during training, which can lead to responses that generalists<\/strong> or inaccurate<\/strong> for questions requiring specialized or detailed knowledge<\/strong>. RAG models, by retrieving specific information from a reference corpus, can provide much more accurate answers. For example, if a question concerns demographic statistics for a specific region, a RAG model can retrieve this data directly from reliable sources, rather than generating an answer based on estimates or generalizations. This ability to access detailed and specific information enables RAGs to outperform LLMs in terms of data accuracy. However, this solution does not completely eliminate the risk of hallucinations in an answer.<\/p>\n\n\n\n

Bias management<\/strong><\/h3>\n\n\n\n

All datasets contain biases, whether due to data selection, collection method, or the inherent biases of the dataset's creators. LLMs, being trained on large datasets, can incorporate and perpetuate these biases in their responses. RAG models, based on carefully selected and diversified sources of information, can help to alleviate this problem<\/strong>. For example, by selecting sources that have been identified as having different or opposing biases, or by including sources specifically intended to represent under-represented perspectives, a RAG model can produce responses that are more balanced and less biased. That said, managing bias requires constant vigilance and regular evaluation of information sources to ensure that they remain representative and balanced.<\/p>\n\n\n\n

The various limits of RAG   <\/strong><\/h2>\n\n\n\n

Source selection and relevance<\/strong><\/h3>\n\n\n\n

The quality of the answers provided by a RAG model depends heavily on the selection of information sources to which it has access. Finding, selecting and maintaining a set of reliable, up-to-date and representative sources can be complex. What's more, there's also the risk of the model retrieving information from sources that aren't entirely relevant to the question posed, which can lead to inaccurate or irrelevant answers. In this case, a great deal of indexing and orchestration work is required to make this approach viable on a professional level.<\/p>\n\n\n\n

Managing misinformation and bias<\/strong><\/h3>\n\n\n\n

Although RAGs can potentially reduce the bias present in responses by diversifying their sources, they are not immune to retrieving and propagating biased or false information. The presence of misinformation in external sources can lead to the generation of responses that perpetuate errors or prejudices. Source selection must therefore be carried out with care to minimize this risk.<\/p>\n\n\n\n

Complex reasoning skills<\/strong><\/h3>\n\n\n\n

Even if RAGs improve the relevance and timeliness of the information provided, they don't necessarily solve all the challenges of complex reasoning and deep contextual understanding that language models can encounter. As a result, there is always a risk of hallucination despite RAG. This is because the LLM can sometimes fail to find a certain word in the RAG database. Or link too many answers, which generally makes generative AIs less attentive to the essentials.<\/p>\n\n\n\n

Answer integrity<\/strong><\/h3>\n\n\n\n

Integrating retrieved information into generated responses presents a challenge in terms of ensuring that responses remain coherent and logically integrated. It can be difficult to ensure that retrieved information aligns perfectly with the rest of the response or with the LLM model, which can sometimes lead to hallucinations.<\/p>\n\n\n\n

Solutions <\/strong><\/h2>\n\n\n\n

If there are any inconveniences solutions<\/a> emerge to go beyond the limits set out above. For example: <\/p>\n\n\n\n