QUID

VIT-AP Capstone · Work in progress

QUID
Queries Unmasked by Iterative Diffusion

Dense retrieval fails when short queries miss domain vocabulary. QUID expands queries with masked text diffusion (LLaDA) so expansions stay semantically anchored. Months 1-3 established the core method and BEIR evidence. Month 4 is underway: we are now exploring an agentic query router that chooses when expansion should fire - and which tool to call.

Month 4 in progress Preprint coming soon Code: github.com/Zhreyu/quid Capstone under Dr. G. Muneeswari
Thesis
Written report
Core QUID method, BEIR results, anchoring analysis, and the Month 4 agentic roadmap.
Weeks
14 detailed logs
Five-page weekly reports from foundation through the current router pilots.
Paper
Draft PDF
Full working draft: BEIR results, anchoring analysis, and Month 4 agentic router pilots. Preprint coming soon.
Code
Zhreyu/quid
Public repository for the ongoing implementation.
System

How QUID retrieves

flowchart LR Q[User query] --> D[LLaDA masked diffusion] D --> E[Expanded query] E --> B[BGE-M3 embed] B --> R[Dense retrieve] R --> O[Ranked documents]
Month 4 · now

Agentic AI expansion & query router

QUID helps most when the bottleneck is a vocabulary gap (medical / finance). On science-claim style queries, always expanding can be the wrong move. That motivates the next layer we are building: treat expansion methods as tools, and let an agent decide whether and when to call them.

flowchart TD Q[Query] --> F[Observe features + confidence] F --> P{Plan tool} P -->|vanilla| V[Skip expansion] P -->|QUID| U[Diffusion expand] P -->|HyDE| H[Hypothetical doc] V --> R[Retrieve] U --> R H --> R R --> C{Critique confidence} C -->|OK| A[Accept] C -->|Weak| T[Retry other tool] T --> R
Designing
Tool set
vanilla (skip), QUID (diffusion expand), HyDE-style expand, critique, optional retry.
Trying now
Routing signals
Query length, domain cues, question form, and unsupervised retrieval confidence.
Early signal
Selective beats blind
Small pilots suggest always-expand is brittle; selective tool use looks more stable. Full study still running.
Next
Month 4 plan
Larger BEIR slices, better critique, learned router toward oracle labels, preprint polish.

We are deliberately not freezing agentic numbers as final claims yet. The serious claim for Month 4 is the research question: can agentic tool use decide when diffusion expansion helps - and recover when it does not?

Capstone team

Students
22MIC7013Yuvraj Jha
22MIC7014Lavanya Gupta
22MIC7168Amardiya S. Mujeeb
22MIC7084Shreyas S

Done under the guidance of
Dr. G. Muneeswari, Professor (Grade 2), Head of the Department of Data Science and Engineering (DSE), School of Computer Science and Engineering (SCOPE), VIT-AP University

School of Computer Science and Engineering (SCOPE) · VIT-AP University

Capstone project · Done under the guidance of Dr. G. Muneeswari, Professor (Grade 2), Head of the Department of Data Science and Engineering (DSE), School of Computer Science and Engineering (SCOPE), VIT-AP University. Code: github.com/Zhreyu/quid · Site: zhreyu.github.io/quid · Status: work in progress.