From this article, I hope to explore different areas of using LlamaIndex to form a production-ready LLM application. Let's start with document metadata.
document metadata
Document metadata is data about data. It is descriptive information about the document, such as document title, keywords, abstract, etc. Document metadata enriches nodes with additional information that can be used during retrieval. Let's explore how metadata enables higher retrieval precision in RAG.
Use Cases
Based on one of our previous stories, Analyzing Financial Reports Using LlamaIndex and OpenAI, let's extend the use case to load the US government's financial reports for FY 2022 and FY 2021, and then ask questions about financial reports for specific fiscal years. This use case is slightly different from the articles we mentioned earlier in that instead of comparing and contrasting, we do a direct Q&A on the report for a specific year, which is a very common RAG use case.
We will illustrate query results for the same set of questions in two scenarios: with and without metadata. Let's observe the difference in response.
How to add metadata to documents
You might think that adding metadata to documents must be a daunting task. Fortunately, LlamaIndex recently released the MetadataExtractor module. The metadata extractor uses LLM to extract certain contextual information related to the document, which is stored in each node to better help retrieval and language model to disambiguate similar passages.
In short, Metadata Extractor uses LLM to automatically generate metadata of documents, which helps LLM to achieve more accurate retrieval. outstanding!
The metadata extractor module comes out of the box with a list of extractors:
TitleExtractor: Extracts the title of the document in the context of each node and stores it in the document_title metadata field.
QuestionsAnsweredExtractor: Node-level extractor. Extract a set of questions that the node can answer and store these questions in questions_thi