Resolve the Personal Hop First: Corpus-Routed Query Trees for Colloquial Questions over Expert Literature

Many questions people actually ask an assistant mix two kinds of information that live in two different places. Consider "can I take my allergy pill before tomorrow's long run, given my asthma has been acting up?" Part of the answer sits in the user's own records: which pill, what dose, what "acting up" meant last week. The rest sits in formal literature written in a vocabulary the user never uses ("cetirizine", "exercise-induced bronchoconstriction", "exacerbation"). Current systems break such questions into sub-questions with a general language model and send every piece to one retriever. The result is a set of colloquial sub-queries ("my allergy pill", "asthma acting up") that match neither the user's records precisely nor the expert corpus at all.

We propose a query decomposer that builds a small tree of sub-questions, labels each leaf by the corpus it must be answered from (personal history or domain literature), and answers personal leaves first. It then substitutes the concrete entities those leaves return (the drug name, the measured value) into the domain leaves. A small rewriter, trained on the domain corpus and conditioned on how far this particular user's vocabulary sits from that corpus, translates whatever colloquial wording remains into the corpus's own terms. The bet is that the most reliable bridge across the vocabulary gap is often the user's own data: once "my allergy pill" becomes "cetirizine 10 mg", the domain lookup is mostly a keyword problem.

Core Claim

The vocabulary gap in mixed personal-and-expert questions is not one gap but two, and they have different fixes. Colloquial references to the user's own things ("my pill", "the device I bought") cannot be translated by any domain model, because the referent is private. They have to be resolved by retrieval over personal history. Colloquial descriptions of general phenomena ("acting up", "feels tight") can be translated by a model that knows the domain's register. A decomposition tree that routes each leaf to the right corpus and passes resolved entities along its dependency edges should beat both flat personalized expansion and user-blind decomposition on these questions. It should do so without hurting ordinary multi-hop questions that have no personal component.

Where This Sits in Prior Work

The decomposition side supplies our method and our main competitor. Question Decomposition for Retrieval-Augmented Generation prompts a language model to emit fact-seeking sub-queries, keeps the original question, pools retrieved passages, and filters them with a cross-encoder reranker, meaning a model that reads query and passage together and scores their relevance. The paper reports large retrieval gains on MultiHop-RAG. It also acknowledges three weaknesses: the number of sub-queries is fixed regardless of how many hops are needed, the reranker depends on being matched to the domain, and the decomposition runs open-loop with no use of intermediate evidence. PLUTO plans sub-queries for tool retrieval and rewrites tool documentation into user language when retrieval fails. It treats the semantic gap globally, for an average user. We keep the tree and the reranker from the first paper. We change two things: what each leaf is retrieved from, and the order in which leaves are resolved.

The personalization side motivates the problem and supplies one component. Personalize Before Retrieve (PBR) expands a query with stylistic pseudo-feedback and PageRank-selected anchors from the user's history, and reports gains over HyDE and ThinkQE on PersonaBench. It produces one flat expanded embedding per query. For a mixed question, that single vector has to point at both the user's records and the expert corpus, which is exactly the conflation we want to avoid. We reuse its anchoring idea, but only on personal leaves. NPA shows that conditioning attention on a user embedding changes which words count as informative. We borrow a much lighter version of that idea: a per-user weight on how much each query term needs rewriting. Learning a user ID embedding would be useless for the cold-start users our benchmark contains.

The domain side supplies the rewriter and a key constraint. BioBART continues pre-training BART on PubMed. It finds that sentence permutation hurts biomedical tasks, and it notes a performance drop on informal patient dialogue (iCliniq). That drop is direct evidence that formal-domain pre-training alone does not close the consumer-to-expert gap. Domain-specific transformer models for query translation translates colloquial e-commerce queries into catalog English with an auxiliary "brand loss" that penalizes changing entity names. We adopt that loss so the rewriter cannot alter entities that came from personal leaves. MonoCoder and A Comparative Study of PEFT Methods for Python Code Generation report that in-domain pre-training matters more than adapter choices. That finding is why we start the rewriter from a domain-pre-trained model rather than a general one with a LoRA adapter. The hand-built synonym ontology of Pham et al. shows the curation cost we hope to avoid.

Research Program

The system has four parts.

Decomposer. A planner model emits a dependency tree whose leaves carry a corpus label (personal or domain) and slots that reference other leaves' answers, for example "interactions of [drug from leaf 1] with exercise-induced bronchoconstriction". Leaf count is not fixed: the planner may emit a single leaf. This addresses the over-decomposition problem noted in the QD+RR paper. We train the planner on a few thousand synthetic trees generated by the benchmark construction below, and we compare it against a zero-shot prompted planner.

Personal leaves. These are resolved by dense retrieval over the user's history with PBR-style anchors. A small extractor then returns typed entities: names, numbers, dates.

Slot filling and rewriting. Entities from personal leaves fill the slots in domain leaves. The rewriter then translates the remaining colloquial spans. It is BioBART continued on consumer-health text such as MEDIQA consumer questions and iCliniq, then fine-tuned with ranking feedback from the domain index in the manner of RaFe: rewrites earn credit when gold passages rise in the ranking. The brand-style preservation loss keeps filled entities exact. The personalization enters through a per-term rewrite weight. For each term, we estimate how often that user's past wording for the concept already retrieved domain documents. A clinician user's terms get low weight and pass through untouched; a layperson's get high weight. This is meant to avoid the degradation that comes from rewriting queries that were already specific.

Reranking. Domain-leaf candidates are reranked by a cross-encoder conditioned on the filled leaf and the original question, following QD+RR.

Benchmark. No existing benchmark in the file requires one personal hop plus one expert hop, so we build one. We pair PersonaBench synthetic users (PersonaBench), extended with injected medication and condition records, with PubMed abstracts drawn from PubMedQA contexts (PubMedQA). Each question is written colloquially by annotators, with gold evidence on both sides. We include a third of single-corpus questions to check that routing does not hurt them.

Baselines

The naive reference point is a single dense retrieval over the union of both corpora using the raw question, with a general reader answering from the top passages. This is the setting in which MultiHop-RAG reports GPT-4 at only 0.56 accuracy with retrieved chunks.

The strongest existing method is QD+RR from Question Decomposition for Retrieval-Augmented Generation. On MultiHop-RAG it reports 87.2% Hits@10 and 0.635 MRR@10, and on HotpotQA 28.1 EM and 35.0 F1. It is the right competitor because it already decomposes and already filters expansion noise. If our routing and resolution add nothing, QD+RR should match us. We run it with the same planner model, the same retriever and reranker checkpoints, the same pooled index over both corpora, the same top-k, and the same number of tuning runs on the dev split.

The second strong baseline is PBR's flat personalized expansion (Personalize Before Retrieve) over the pooled index with the same backbone. The file reports up to a 10% R@5 and N@5 gain over HyDE but no absolute PersonaBench number, so we reproduce it ourselves.

We also include QD+RR plus prompted formal rewriting: the planner is told to phrase every sub-query in expert terminology and is given the user profile in context.

The ablations isolate the key ingredient. They remove, one at a time, corpus routing (all leaves to the pooled index), dependency ordering (leaves resolved in parallel, no slot filling), the trained rewriter (replaced with prompted rewriting), the preservation loss, and the per-user rewrite weight.

Beating QD+RR means a statistically significant gain in joint evidence recall on mixed questions, with no loss on single-corpus questions and on the original MultiHop-RAG test set. The joint metric counts a question as correct only if gold evidence from both corpora is retrieved in the top 10.

Why not just…?

Why not just give a frontier model the user profile in context and let it reason step by step? Linear reasoning chains propagate early mistakes downstream (Cao et al.). PersonaBench reports only about 50% recall on personalized questions even when models are given ground-truth context, so dumping history into the prompt is a weak bridge. We include this as a baseline with the full profile in context, using the same reader. If it matches us on mixed questions, the tree is unnecessary for that model size, and we will report that.

Why not just prompt the decomposer to write sub-queries in medical vocabulary? This is the real threat. A large model may know that "acting up" means exacerbation. Our claim is that prompting cannot resolve private referents. "My allergy pill" has no expert translation without looking at the user's records. Splitting the results by whether a question contains such a referent settles it. If prompted rewriting matches the trained rewriter on questions without private referents, we concede that component and reduce the contribution to routing plus ordering. The small rewriter would then matter only for latency, against the 18.9 s/query retrieval cost reported for QD+RR.

Why not just use a strong general dense retriever and skip rewriting? On ToolBench, PLUTO's baseline Contriever reaches only 31.78% recall on user-to-documentation matching, and ECNU-ICA at TREC 2019 Precision Medicine reports that neural models without domain alignment struggle to beat vector-space baselines. The no-rewriter ablation measures exactly this on our benchmark.

Proposed Experiments

The decisive check is a two-by-two split of mixed questions: whether the question contains a private referent, crossed with whether it contains colloquial descriptions of general phenomena. We predict the following pattern. Dependency ordering accounts for most of the gain on private-referent questions. The rewriter, prompted or trained, accounts for the gain on colloquial-description questions. QD+RR fails mainly in the first cell.

Secondary checks cover four things. The first is entity preservation rate: the fraction of entities from the user's records that appear exactly in domain leaves, with and without the preservation loss. The second is no-regression on MultiHop-RAG against the reported 87.2% Hits@10. The third is downstream answer accuracy on the PubMedQA-derived questions, where 68.1% for multi-phase BioBERT is the reported reference. The fourth is per-question latency. A failure of the first prediction would falsify the core claim: if resolving the personal hop first does not beat QD+RR on private-referent questions, the idea is wrong.