Revolutionizing AI Responses: How QPLANNER Enhances Precision in Large Language Models

Traditional retrieval-augmented generation (RAG) systems often produce lengthy and irrelevant content, failing to meet specific user needs.

This article introduces a new study titled "Learning to Explore and Select for Coverage-Conditioned Retrieval-Augmented Generation". This study aimes to address the challenge of handling specific user information requests with large language models (LLMs).

The motivation behind this paper is to improve RAG systems' performance by constructing and utilizing QTREE to generate customized query outlines that adhere to coverage-conditioned queries (C2), thus better satisfying diverse user interests.

Figure 1 compares traditional RAG systems with the proposed QPLANNER system.

Figure 1: Unlike the existing RAG systems that simply rewrite or decompose the given query, QPLANNER, trained with QTREE, first explores possible information coverage. A subset of queries is selected as an outline to retrieve relevant passages and generate long-form response. By leveraging retrieved passages, we prompt LLM to generate the final response tailored to user interests. Source: QTREE.

Solution

Overview

The paper proposes a 7B language model named QPLANNER, designed to generate customized query outlines that meet the coverage-conditioned queries (C2). QPLANNER is trained using QTREE, which consists of 10K sets of information-seeking queries, to address specific information needs across different topics.

Figure 2 illustrates the overall framework of QPLANNER, including the construction of QTREE and the training process of QPLANNER.

Figure 2: The overview of the framework. [Step 1] Base query (qbase) is decomposed into queries with diverse viewpoints (QTREE), preceded by generating coverage query (qcov). [Step 2] After C2 candidate outlines are extracted, an LLM evaluates each outline and selects the best-scored one. [Step 3] Utilizing this dataset, QPLANNER is trained to sequentially generate its own QTREE and preferred outline by taking the C2 query as an input. Source: QTREE.

Detailed Process

The detailed process of the solution is as follows:

1. QTREE Construction: Initially, base queries (qbase) are collected from datasets such as ASQA, Longform, and ExpertQA, and decomposed into hierarchical query trees (QTREE). Each QTREE contains 39 queries representing various perspectives of the main topic.

2. Generating Coverage Queries (qcov): Based on the user's background knowledge and intent operations (INCLUSION or EXCLUSION), a query is randomly selected from QTREE to generate qcov.

3. Parsing Candidate Outlines: LLM sequentially extracts candidate outlines from QTREE, ensuring that the queries in the outline are interconnected and non-overlapping.

4. Evaluating Outline Quality: LLM evaluates the quality of candidate outlines, scoring them based on how well they align with C2 queries.

5. Training QPLANNER: QPLANNER is trained through supervised fine-tuning (SFT) and direct preference optimization (DPO) to generate customized outlines that meet user requirements.

Evaluation

The paper evaluates QPLANNER's performance through both automatic and human assessments.

• In automatic evaluation, LLM scores the outlines, demonstrating that QPLANNER significantly outperforms random baselines and models trained with only SFT.

• In human evaluation, participants compared outlines and responses generated by SFT-QPLANNER and DPO-QPLANNER, showing that DPO-QPLANNER produced more user-aligned outlines and responses.

Figure 3 shows the evaluation results of QPLANNER's performance in both search queries and content drafts.

Figure 3: Human evaluation results. Source: QTREE.

Overall, Figure 3 highlights two significant advantages of QPLANNER: its generated queries are more effective in retrieving relevant documents, and the content generated based on these queries is more aligned with user needs, thereby improving overall user satisfaction.

Case Study

Below is a case study demonstrating the effectiveness of QPLANNER in handling complex queries.

For example, in response to the C2 query, "Describe the film The Woman Hunt, avoiding discussions of reviews or audience feedback," QPLANNER generated a well-structured and compliant outline, leading to a final long-form response that better met the user's expectations.

Figure 4: Example of parsed outline. Source: QTREE

Conclusion and Insights

This article provides a comprehensive overview of the innovative QPLANNER framework, which significantly improves the performance of retrieval-augmented generation systems in coverage-conditioned scenarios.

By leveraging hierarchical query sets (QTREE) and a well-trained language model (QPLANNER), the framework effectively addresses the challenges of generating long-form responses tailored to specific user needs.

However, from my perspective, it has some limitations:

1. Challenges in Complex Scenarios: Although QPLANNER performs well in many situations, it may still struggle to fully meet user needs when the requests involve very complex or ambiguous domains.

2. Dependence on Training Data Quality: The construction of QTREE and the training of QPLANNER heavily rely on the quality and diversity of the datasets. If the training data is insufficient or biased, the results may be suboptimal.