Multi-Label Conformal Prediction for Uncertainty-Aware RAG Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing retrieval augmented generation (RAG) systems lack flexibility in communicating uncertainty about queries, leading to potential hallucinations and incorrect responses due to the lack of uncertainty quantification, especially when returning a fixed number of passages.
Innovation Solution
Implementing conformal prediction to convert heuristic uncertainty into a statistically rigorous notion of uncertainty, allowing for multi-label prediction sets that adjust based on model uncertainty, ensuring a higher certainty level of document relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a fixed number of passages is returned in RAG systems, then the system operates with simple and predictable behavior, but the system cannot communicate uncertainty about queries leading to potential hallucinations and incorrect responses
Solution Approach 1:
The patent changes the parameter of passage selection from a fixed number (k) to a dynamic set based on conformal prediction p-values. Instead of returning a fixed k passages, the system returns a variable number of passages where each passage is included if its conformal p-value exceeds a threshold, allowing the system to communicate uncertainty while maintaining reliability.
Solution Approach 2:
The system incorporates feedback through conformal prediction calibration, where the p-values are computed based on the distribution of scores from the retrieval model. This feedback mechanism allows the system to adjust the set of returned passages based on the uncertainty level indicated by the p-values, creating a self-regulating system that adapts to query complexity.
2Quantity of substance
If multiple documents are retrieved to ensure comprehensive information, then the completeness of information is improved, but the computational burden and processing time increase
Solution Approach 1:
The patent replaces the mechanical approach of retrieving a fixed number of top-k passages with a statistical approach using conformal prediction. Instead of relying solely on the retrieval model's ranking mechanism, the system uses p-value thresholds to determine which passages to return, substituting the mechanical sorting process with a statistical filtering process that is more efficient at identifying truly relevant documents.
Solution Approach 2:
The system applies partial action by returning only those passages that meet the conformal p-value threshold, rather than returning all top-k passages regardless of their relevance. This selective approach ensures that only passages with sufficient confidence are processed, reducing computational burden while maintaining completeness of relevant information.
Data Source
AI summary
Systems and methods for retrieving relevant documents. A computing system obtains, from each document in a corpus of documents, a plurality of chunks corresponding to portions of text. It computes a score for each one of the plurality of chunks in relation to a query. The chunks are reordered according to score. A sum of the highest scores is computed, and a subset of chunks associated with the highest scoring documents are retrieved. A large language model (LLM) may be used to generate response text from the retrieved documents.