How to align with user preference in a RAG system?

Current embedding-based RAG systems primarily rely on semantic similarity. Given a document and a query, the system usually retrieves multiple sections that appear semantically relevant. However, in domain-specific applications, such as financial analysis or legal research, users often have domain-specific preferences for which parts of a document to consult first. These preferences are typically driven by experience about where answers are typically found or which sections are considered more trustworthy sources of information.

For example:

- When querying about financial performance metrics (e.g., earnings adjustments), experienced analysts typically look first at the Management’s Discussion and Analysis (MD&A) section or related financial statement footnotes.

- For questions about company risks, they usually prioritize the Risk Factors section before turning to broader disclosures.

These expert-driven navigation patterns are difficult to capture using embedding-based RAG alone. Fine-tuning embedding models to reflect such preferences is possible, but it tends to be costly and resource-intensive.

An alternative approach is to incorporate reasoning-based retrieval, which mimics how humans find information. For example, when reading a long document, a human typically starts by reviewing the table of contents to determine which sections to read first, based on the context of the query and preference. Similarly, one can build an LLM agent that analyzes the "table of contents" and then navigates through the document according to expert preferences. This can be achieved by using few-shot prompting, where the system learns from sample user preference examples provided in the prompt, allowing it to prioritize sections based on the user’s needs.

To support this paradigm, we developed an open-sourced tool called PageIndex. It can transform any long documents into an LLM-friendly "table-of-contents" tree index, which is ready for the LLM agents to navigate. With PageIndex, you can easily build RAG agents that align with user preferences and domain logic.

Would love any feedback, particularly thoughts on reasoning-based RAG or other potential applications of PageIndex.