---
title: Quality-Controllable Retrieval
url: https://www.emergentmind.com/topics/quality-controllable-retrieval
type: topic
---

# Quality-Controllable Retrieval

Searching arXiv for recent papers on quality-controllable retrieval and related controllable retrieval frameworks.
Quality-controllable retrieval denotes retrieval systems in which retrieval behavior is explicitly steered toward a specified target quality profile, typically through a control function, a conditional variable, or an inference-time parameter. In the controllable learning formalization for information retrieval, the requirement space is written as $\mathcal{T}=\{\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}},\mathbf{s}_{\mathrm{tgt}}\}\in\Gamma$, and a control function $h$ maps a base learner $f$ and the requirement description and context to a controlled learner $f_{\mathcal{T}}$, with the crucial property that new requirements can be handled without retraining the base model [2407.06083]. Within retrieval, “quality” is not a single universal metric: the target may be precision, diversity, recall, retrievability, watch time, correctness, efficiency, security, maintainability, relevance, aesthetics, factual consistency, or an explicit balance between accuracy and cost [2407.06083] [2303.11648] [2412.13844] [2506.11066] [2602.21175] [2502.12145].

## 1. Formalization and conceptual scope

The survey formulation of controllable learning places quality-controllable retrieval within a broader IR framework. Let $\mathcal{X}$ denote inputs such as queries or user profiles, let $\mathcal{Y}$ denote outputs such as ranked lists, and let $\Gamma$ denote task requirements. The controlled learner is written as
$$
f_{\mathcal{T}} = h\bigl(f,\mathbf{s}_{\mathrm{desc}},\mathbf{s}_{\mathrm{ctx}}\bigr),
$$
where $\mathbf{s}_{\mathrm{desc}}$ is the machine-readable description of the target, $\mathbf{s}_{\mathrm{ctx}}$ is auxiliary context, and $\mathbf{s}_{\mathrm{tgt}}$ is the desired numeric or semantic target [2407.06083]. In this formulation, quality is simply one dimension of $\mathbf{s}_{\mathrm{tgt}}$; for example, it may be “precision@10 $\ge 0.75$,” or a weighted balance between precision and diversity expressed by a control weight $\alpha\in[0,1]$ [2407.06083].

The literature represented by recent systems shows that “quality” is instantiated in multiple non-equivalent ways. In controllable video-to-music retrieval, it is the trade-off between broad audiovisual correspondence and label-driven retrieval specificity, controlled by the inference-time mixing weight $\alpha$ in
$$
z^M=(1-\alpha)\,z_{ss}^M+\alpha\,z_{sup}^M
$$
[2412.05831]. In dense search, it may mean improving long-tail content discovery by controlling the narrow-versus-broad intent of generated queries through $p_\theta(q\mid e,c)$, where $c\in\{\mathrm{narrow},\mathrm{broad}\}$ [2303.11648]. In recommendation, it can be the ability to steer retrieval toward watch-time objectives by conditioning the user tower on a watch-time variable $c$ [2412.13844]. In RAG, it may refer to balancing coverage, ranking quality, and factual consistency under entropy and variance constraints [2510.27054], or balancing answer accuracy against retrieval cost through an interpolation parameter $\alpha$ over retrieval strategies [2502.12145].

This suggests that quality-controllable retrieval is better understood as a family of mechanisms rather than a single architecture. What unifies the family is the presence of an explicit control interface and an observable response in retrieval outputs or retrieval-conditioned generation.

## 2. Control targets and loci of intervention

The survey distinguishes several control dimensions relevant to retrieval. Under “What is controllable?”, it identifies Multi-Objective Control, Historical-Behavior Control, and Controllable Environmental Adaptation [2407.06083]. Multi-Objective Control treats the target as a vector of objectives $\{J_1,J_2,\dots,J_M\}$, such as precision, diversity, and latency, with a description vector $\boldsymbol\lambda=(\lambda_1,\dots,\lambda_M)$ satisfying $\sum\lambda_i=1$ [2407.06083]. Historical-Behavior Control allows user-editable profiles or historical interaction manipulation to alter downstream quality. Environmental adaptation injects tags such as time-of-day or device context so that quality dimensions such as precision or freshness change with the environment [2407.06083].

The same survey also organizes control by where it is applied: pre-processing, in-processing, and post-processing [2407.06083]. Pre-processing changes the input before retrieval. CtrlQGen is a canonical case: an entity is serialized as text, a control token indicating narrow or broad intent is prepended, and a T5-base model is trained to generate intent-conditioned synthetic queries that later reshape retriever training data or user query suggestions [2303.11648]. The QCQC framework for text-to-image retrieval is another pre-processing design: a short underspecified query $Q$ is completed into $Q'$ by a language model conditioned on discretized relevance and aesthetic levels, after which a frozen VLM retrieves images using cosine similarity between $g_{\text{text}}(Q')$ and image embeddings [2602.21175].

In-processing changes the retrieval model or routing dynamics themselves. Control-MVR learns separate self-supervised and supervised branches for audio and video, and mixes them at inference using $\alpha$ [2412.05831]. CRM inserts a scalar or vector condition $c$ into the user tower of a two-tower retriever, making the scoring function conditional:
$$
s(u,i\mid c)=\langle u^c,v_i\rangle
$$
[2412.13844]. Flare-Aug interpolates the weights of two external classifiers,
$$
W_\alpha=(1-\alpha)W_{coc}+\alpha W_{roc},
$$
to decide whether a query should use no retrieval, single-step retrieval, or multi-step retrieval [2502.12145]. Multi-granular confidence-constrained RAG dynamically routes across multiple memory granularities and prunes low-confidence paths when entropy or variance thresholds are exceeded [2510.27054].

Post-processing alters ranking or retrieval outputs after a base model has produced candidates. The survey gives MMR as a classic example:
$$
d_i=\arg\max_{d\in R\setminus S}\bigl[\lambda\,\mathrm{sim}(q,d)-(1-\lambda)\max_{d'\in S}\mathrm{sim}(d,d')\bigr]
$$
for trading relevance against redundancy [2407.06083]. In code retrieval, CoQuIR describes multi-headed scoring in deployment: compute semantic score $s_s$ and quality score $s_q$, then combine them as $\text{final\_score}=\alpha\cdot s_s+(1-\alpha)\cdot s_q$ [2506.11066]. In generative text-music retrieval, GD-Retriever applies post-hoc control through negative prompting and DDIM inversion in the latent retrieval space rather than modifying the pretrained teacher encoder [2506.17886].

## 3. Mechanisms of controllability

A major class of methods uses explicit interpolation between competing retrieval objectives. Control-MVR trains with self-supervised InfoNCE losses and supervised SupCon losses in both audio-to-video and video-to-audio directions, sums the four losses equally with $\lambda=1$ in experiments, and then exposes a single inference-time knob $\alpha\in[0,1]$ to trade off broad audiovisual alignment against label information [2412.05831]. The same paper reports a smooth trade-off: at $\alpha=0.0$ the model attains best self-supervised behavior with $R@10\approx 9.8\%$ and $P@10\approx 37\%$; at $\alpha=1.0$ it reaches $R@10\approx 5.1\%$ and $P@10\approx 43\%$; intermediate values give $R@10\approx 10.4\%$ at $\alpha\approx 0.4$ and $P@10\approx 43.3\%$ at $\alpha\approx 0.8$ [2412.05831].

A second class uses conditional generation to reshape the effective query. CtrlQGen learns $p_\theta(q\mid e,c)$, where the condition $c$ specifies narrow or broad intent, and uses weak labeling to create triplets $(e,c,q)$ without manual intent annotations [2303.11648]. The generated queries are then used either to retrain dense retrievers or to suggest broader user queries, reducing retrievability bias in Tracks, Podcasts, and Books [2303.11648]. QCQC follows the same broad pattern but with different control semantics: it discretizes image relevance and aesthetics into levels such as Low, Medium, and High using percentile thresholds $p_1=33$ and $p_2=66$, fine-tunes a language model on captions conditioned by these levels, and then retrieves with a frozen OpenCLIP model [2602.21175].

A third class uses conditional features inside the retrieval model. CRM extends the two-tower paradigm by inserting a watch-time condition $c$ into the user tower. In the DNN version, the conditional user representation is $u^c=\mathrm{CRM\_User\_Tower}(\mathrm{UserFea},c)$; in the transformer variant, cumulative watch-time-to-go tokens are fed into a Decision-Transformer-style block before producing $u^c$ [2412.13844]. The system uses a time-division multiplexing strategy at serving time, switching between $c=\max(w_{n-31\dots n})$ and $c=\mathrm{avg}(w_{n-31\dots n})$ with probability $p$, thereby steering the aggregate trade-off among watch time, video count, and interactions [2412.13844].

A fourth class uses routing, uncertainty, and closed-loop control. The multi-granular confidence-constrained RAG framework builds a hierarchical memory $M^{(\ell)}$ across levels such as token, passage, document, and temporal or global memory, computes per-level routing weights
$$
a^\ell=\frac{\exp(\max_i S_i^\ell)}{\sum_{k=1}^L \exp(\max_j S_j^k)},
$$
fuses retrieved contexts, and regularizes generation with generation loss, entropy, and variance terms [2510.27054]. During inference, if entropy or variance exceeds a threshold, the framework removes the lowest-confidence retrieval path and renormalizes the weights [2510.27054]. MemR$^3$ implements a different closed-loop controller: a router selects among `retrieve`, `reflect`, and `answer`, while a global evidence-gap tracker maintains accumulated evidence $\mathcal{E}_k$ and unresolved gaps $\mathcal{G}_k$; deterministic constraints enforce an iteration budget and cap consecutive reflections [2512.20237].

A fifth class uses generative latent-space manipulation. GD-Retriever generates a “ghost” audio embedding sequence in an audio-only latent space from text conditioning, retrieves nearest audio neighbors to the generated latent, and supports control through classifier-free guidance, negative prompting, and DDIM inversion [2506.17886]. Its negative prompting update explicitly repels an unwanted attribute embedding:
$$
\tilde x_{t-1}^{\mathrm{NP}}=(1+w)\,G_\theta(x_t,t,z_T^{q+})-w\,G_\theta(x_t,t,z_T^{q-})
$$
[2506.17886]. This allows post-hoc manipulation of retrieval behavior without retraining the frozen audio teacher encoder.

## 4. Representative systems and application domains

The recent literature spans search, recommendation, cross-modal retrieval, code retrieval, and retrieval-augmented generation. The following systems illustrate the range of control variables and target quality notions.

| System | Control variable | Controlled quality dimension |
|---|---|---|
| Control-MVR | $\alpha\in[0,1]$ | self-supervised vs. label-supervised retrieval |
| CtrlQGen | $c\in\{\text{narrow},\text{broad}\}$ | retrievability and discovery |
| CRM | watch-time condition $c$ | watch time vs. video count vs. interactions |
| Flare-Aug | $\alpha\in[0,1]$ | accuracy vs. retrieval cost |
| QCQC | $(l_{rel},l_{aes})$ | relevance and aesthetics |
| CoQuIR-style retriever | $\alpha$ or dimension weights | correctness, efficiency, security, maintainability |

In search, retrievability is treated as a measurable property of the catalog. CtrlQGen adopts the retrievability score
$$
r(e)=\sum_{q\in Q} w_q\,\mathbf{1}\bigl(\mathrm{rank}(q,e)\le k\bigr)
$$
with $k=100$ in its experiments, and summarizes retrievability bias via the Gini coefficient $G$ [2303.11648]. On Tracks, a CtrlQGen-trained Bi-Encoder reduces Gini from $0.856$ to $0.693$ while increasing $R@100$ from $0.142$ to $0.333$, and moves **9%** of entities from zero to nonzero retrievability; broad-query suggestion lowers Gini by up to $11\%$ and brings **11%** more entities to $r(e)>0$ [2303.11648]. These results situate quality control not only as re-ranking quality, but as control over the distribution of exposure across a catalog.

In recommendation, CRM treats retrieval quality as alignment between retrieval and downstream regression-oriented business targets. Offline, the transformer variant reaches **42 s** average time per video view versus next best $\approx 39$ s; online A/B tests on Kuaishou report significant gains including **Video Watch Time: +0.372% / +0.457%**, **Total App Usage Time: +0.175% / +0.196%**, and **Follows: +0.447% / +0.828%**, with all gains significant at $p<0.05$ [2412.13844]. Here, controllability is embedded directly into the retrieval score rather than delegated to a later ranking stage.

In code retrieval, CoQuIR makes quality dimensions explicit and benchmarkable. It contains **42,725** natural-language–code queries, **134,907** candidate snippets, and **11** programming languages, and evaluates correctness, efficiency, security, and maintainability through Pairwise Preference Accuracy and Margin-based Ranking Score [2506.11066]. Off-the-shelf models often achieve only $PPA\approx 0.5$–$0.6$ and $MRS\approx 0$–$0.1$ on many tasks, while quality-contrastive tuning improves PPA by **20–30 pp** and MRS by **10–25 pp** without sacrificing semantic relevance by more than about $\pm 1$–$2$ pp in nDCG@10/MRR [2506.11066]. The benchmark thereby shifts quality control from an informal notion to an evaluable retrieval property.

In text-to-image retrieval, QCQC addresses short and underspecified queries by conditioning query completion on discrete relevance and aesthetic levels before retrieval with a frozen OpenCLIP ViT-H-14 model [2602.21175]. On MS-COCO in the High-High condition, it improves relevance from **0.33** to **0.39** and aesthetics from **4.82** to **5.02** compared to the best fine-tuned unguided LLM, and five-level discretization still preserves smooth control [2602.21175]. The paper also reports that off-the-shelf LLMs without fine-tuning often degrade retrieval quality and that post-retrieval reranking by aesthetics cannot jointly optimize relevance and aesthetics as effectively [2602.21175].

In retrieval-augmented generation and memory systems, quality control frequently targets reliability rather than only ranking quality. The multi-granular confidence-constrained RAG framework reports on CISI that its method attains **77.8 / 92.0 / 90.0 / 0.72** on QA Accuracy, Recall@5, NDCG@5, and factual consistency, compared with **69.3 / 80.0 / 78.0 / 0.65** for Self-RAG [2510.27054]. MemR$^3$, evaluated on LoCoMo, reaches **86.75%** overall LLM-as-a-Judge score with a RAG backbone and GPT-4.1-mini, versus **79.46%** for vanilla RAG, and improves both RAG and Zep backends while adding only **10–20%** extra retrieval tokens relative to vanilla RAG [2512.20237].

## 5. Evaluation methodology

Evaluation in quality-controllable retrieval is necessarily heterogeneous because the target quality dimension changes with the application. The survey lists standard single-objective IR metrics such as Precision@k, Recall@k, NDCG@k, Hit Rate, Coverage, $\alpha$-NDCG, and ERR-IA, and multi-objective metrics such as Hypervolume and correlation between control parameter $\alpha$ and an observed metric $s$ via Pearson or Spearman correlation [2407.06083]. It also prescribes an evaluation protocol in which one synthesizes a held-out requirement, generates the controlled model, measures target metrics, plots metric values against the control variable, and computes correlation or regret relative to ideal requirements [2407.06083].

Several systems add task-specific evaluation constructs. CtrlQGen uses retrievability score, Retrievability Gini, nonzero retrievability percentage, and $R@100$ [2303.11648]. Control-MVR distinguishes self-supervised retrieval, evaluated by Recall@K and MRR, from genre-supervised retrieval, evaluated by Precision@K and macro-averaged MRR over **11** genres [2412.05831]. CoQuIR introduces Pairwise Preference Accuracy, whose random baseline is **0.5** and ideal is **1.0**, and Margin-based Ranking Score, where negative values indicate pathological preference for low-quality code [2506.11066]. QCQC reports average relevance and average aesthetics over top-5 retrieved images [2602.21175]. Flare-Aug uses answer accuracy and retrieval cost
$$
C=\frac{1}{N}\sum_{i=1}^N \#\text{retrieve calls for }q_i
$$
to characterize the accuracy–cost frontier [2502.12145].

A recurring empirical criterion is monotonicity or smoothness of the response curve. Control-MVR reports a smooth trade-off over $\alpha$ [2412.05831]. Flare-Aug reports that both expected retrieval cost $C(\alpha)$ and accuracy $A(\alpha)$ are approximately monotonic in $\alpha$, and on Flan-T5 XL shows validation points such as $\alpha=0.0$ with accuracy **0.388** at **1.3** average steps, $\alpha=0.4$ with **0.439** at **2.2** steps, and $\alpha=1.0$ with **0.441** at **3.0** steps [2502.12145]. QCQC reports a clear monotonic trend across low-to-medium-to-high quality conditions [2602.21175]. This emphasis on response curves is central: a controllable system is not merely high-performing, but predictably steerable.

## 6. Limitations, misconceptions, and open directions

A common misconception is that controllability is equivalent to a simple post-retrieval score adjustment. The surveyed systems contradict this. Control may be implemented by rule-based post-processing, Pareto optimization, hypernetworks, controllable query generation, conditional towers, uncertainty-based path pruning, or agentic retrieve–reflect–answer loops [2407.06083] [2303.11648] [2412.13844] [2510.27054] [2512.20237]. Another misconception is that natural-language prompting alone suffices. CoQuIR finds that pre-tuned instruction retrievers are largely insensitive, with **<1 pp** change in PPA/MRS, to explicit prompts such as “retrieve buggy” versus “retrieve correct,” indicating that specialized quality signals must be injected during training [2506.11066]. QCQC likewise reports that off-the-shelf LLMs without fine-tuning often degrade retrieval quality [2602.21175].

The literature also shows that more retrieval is not automatically better. Flare-Aug is motivated by the fact that over-retrieving simple queries is wasteful and that users may require different points on the accuracy–cost spectrum [2502.12145]. MemR$^3$ explicitly prevents both under-retrieval and over-retrieval by maintaining evidence and gap sets and enforcing budgets such as $n_{\max}$ and $n_{cap}$ [2512.20237]. The confidence-constrained RAG framework states that in highly heterogeneous memories factuality degrades unless confidence constraints are tightened, which can in turn reduce coverage [2510.27054]. These results indicate that controllability often mediates irreducible trade-offs rather than eliminating them.

Limitations recur across domains. The survey identifies balancing difficulty, absence of standardized evaluation, task description design, and real-time or online scalability as central challenges [2407.06083]. Multi-granular RAG notes that computational overhead grows with the number of levels $L$ and repeated pruning, and that $\lambda_1$, $\lambda_2$, and confidence thresholds require careful tuning [2510.27054]. MemR$^3$ does not use explicit gradient-based training for its router or tracker, relying instead on prompt engineering and rule-based overrides [2512.20237]. GD-Retriever reports that diffusion sampling is significantly heavier than a single forward pass of a contrastive encoder and inherits distributional biases of its teacher encoders under domain shift [2506.17886]. QCQC reports that cross-dataset fine-tuning demonstrates the importance of dataset-specific adaptation for relevance control [2602.21175].

The research agenda identified by the survey includes theoretical analyses of controllable learners, cost-efficient control, LLM-empowered controllable learning, multi-task and multi-scenario switching, and shared benchmarks and metrics [2407.06083]. The concrete systems surveyed here add more specific directions: extending discrete intent control to a continuous intent spectrum in CtrlQGen [2303.11648], enriching retrieval with multi-dimensional conditions in CRM [2412.13844], expanding multi-granular memory to cross-modal settings [2510.27054], and using active learning or multi-objective optimization for code-quality retrieval [2506.11066]. Taken together, these directions frame quality-controllable retrieval as an increasingly general design principle for retrieval systems that must expose explicit, interpretable, and tunable behavior under changing user requirements and deployment constraints.

Source: https://www.emergentmind.com/topics/quality-controllable-retrieval