Talk With Your Documents Via Docs2KG

Session Track: Data Science

Session Time: 06:00 am - 06:15 am November 7

Session description

Even with a conservative estimate, 80% of enterprise data is stored in unstructured files within data lakes that contain various formats. Traditional search engines can no longer meet the needs for information seeking, especially when the task involves browsing and exploring to gain insights without clear search keywords. Large language models (LLMs) and Retrieval Augmented Generation (RAG) offer a solution to this problem. RAG currently splits and embeds documents, then performs a semantic similarity search to retrieve relevant content. However, this approach can lead to an information cocoon problem. For example, if you want to understand the profit from the last financial year, your question will retrieve the profit report; however, it may not provide information on why the profit increased or decreased, as this information might be in sections before or after the profit section within your report document, and the content there may not have high semantic similarity with your question. In such cases, both the document's layout structure and semantics are equally important when using LLMs to answer exploratory tasks where you may not know exactly what you want to find out. To address this, we developed an open-source package called Docs2KG: https://docs2kg.ai4wa.com/, which will build a multimodal knowledge graph with dual aspects: One aspect represents the structural relationships within the documents, and the other aspect represents the semantic relationships. Using this foundation, we can implement GraphRAG, a method proposed by Microsoft, to achieve better results, especially for exploratory questions.