VaxPulse Query Corner

• VaxPulse Query Corner is a query-focused component within the VaxPulse ecosystem that leverages retrieval-augmented generation to analyze complex vaccine concerns in near real-time.
• It integrates diverse data from social media, search trends, and clinical guidelines with advanced preprocessing, filtering, and multilingual normalization to ensure reliable evidence synthesis.
• The system supports both public health surveillance and point-of-care clinical advice through a dual-mode architecture that emphasizes dashboard-driven insight and agentic RAG for informed interventions.

VaxPulse Query Corner is a query-focused component in the VaxPulse ecosystem that uses Retrieval Augmented Generation to answer complex queries about public vaccine concerns on online platforms, while also functioning as an interactive portal for dashboard-driven surveillance and feedback. In the published formulations, it is situated within a broader near-real-time monitoring framework that integrates social media, search trends, sentiment analysis, topic modelling, adverse events following immunisation (AEFI), and expert review. A separate technical blueprint further shows how an Agentic RAG vaccination-guideline system can be adapted into a VaxPulse Query Corner feature for clinical question answering, extending the concept from population-level concern monitoring to point-of-care information access (Javed et al., 17 Jul 2025, Dimaguila et al., 6 Jul 2025, Javed et al., 7 Jul 2025, Zeggai et al., 4 Jul 2025).

1. System definition and architectural position

Within the VaxPulse Vaccine Infodemic Risk Assessment Lifecycle (VIRAL), Query Corner is the user-facing interactive portal embedded in a Power BI suite plus a bespoke web app. The surrounding lifecycle comprises data ingestion, preprocessing, machine learning modules, active learning and data augmentation, a risk scoring engine, and finally Query Corner itself, where dynamic dashboards and expert feedback loops are exposed to operational users. The same papers describe VaxPulse more broadly as a modular learning-health-system for continuous, real-time monitoring and mitigation of vaccine-related misinformation, with Query Corner acting as the interface through which analysts, epidemiologists, and stakeholders inspect outputs and return corrections to the model-development cycle (Dimaguila et al., 6 Jul 2025).

A second formulation of Query Corner is explicitly RAG-centric. In that version, the tool addresses complex queries about public vaccine concerns on various online platforms and is intended to aid public health administrators and stakeholders in understanding public concerns and implementing targeted interventions to boost vaccine confidence. The architecture is therefore not limited to passive visualization: it supports prompted question answering, ranked evidence retrieval, summarisation, and concern extraction over a continuously updated corpus of vaccine discourse (Javed et al., 17 Jul 2025).

A third line of work provides a direct adaptation pathway from AI-VaxGuide. That system was built as a multilingual question-answering system over national protocols and WHO recommendations using Agentic RAG, and its technical overview labels the adapted interface an “Agentic RAG Pipeline for VaxPulse Query Corner.” This does not imply that all Query Corner deployments are identical; rather, it establishes a documented design pattern in which document parsing, semantic chunking, hybrid retrieval, an agentic controller, and answer generation are assembled into a vaccination QA feature (Zeggai et al., 4 Jul 2025).

2. Data sources, preprocessing, and multilingual normalization

The data foundation of Query Corner is heterogeneous. Reported inputs include X, Reddit, YouTube, Facebook, Google Trends, and, in the broader VIRAL setting, news media APIs and WHO EARS feeds. Collection is performed through authenticated API clients such as Tweepy for X, PRAW for Reddit, and the Google API Client Library for YouTube, with continuous weekly pulls used to capture evolving conversation. In the clinical-guideline adaptation, the corpus instead consists of structured vaccination documents parsed and chunked for downstream retrieval (Javed et al., 7 Jul 2025, Javed et al., 17 Jul 2025, Dimaguila et al., 6 Jul 2025, Zeggai et al., 4 Jul 2025).

Preprocessing is correspondingly multi-stage. The published pipelines remove URLs, user mentions, special characters, HTML, emojis, and stop-words; tokenize with white-space and punctuation tokenizers or spaCy/NTLK; and drop very short sentences, including the configurable threshold $|d|<3$. Language detection is performed with fastText-langid or fastText, after which non-English comments are batched through an LLM-based translation pipeline, exemplified with GPT-4o, to produce English text for uniform modelling while retaining source-language metadata for stratified analysis. This multilingual stratification is a central feature rather than a post hoc convenience, because it enables per-language cohort analyses such as sentiment trends in Mandarin versus Arabic social media (Javed et al., 7 Jul 2025).

The vaccine-specific filtering stage is also explicit. One implementation uses an ensemble classifier combining fine-tuned RoBERTa-large-MNLI and GPT-4o with prompt engineering to route each comment into “relevant vaccine discussion,” “personal experience,” or “irrelevant,” allowing only relevant posts to advance. Another uses a fine-tuned BERT classifier to distinguish “vaccine” versus “health-mention” versus “other,” followed by GPT-4o prompting to select posts about a particular product such as Shingrix. Social-bot filtering is additionally reported in the Shingrix RAG pipeline, and the broader VIRAL system treats bot detection as a first-class machine-learning module (Javed et al., 7 Jul 2025, Javed et al., 17 Jul 2025, Dimaguila et al., 6 Jul 2025).

Search behavior enters the system through a separate but related signal family. Google’s Vaccine Search Insights index is constructed from vaccine-related query counts $N_{r,t}^{(\mathrm{vac})}$ and total search counts $N_{r,t}^{(\mathrm{total})}$ within region $r$ and week $t$, after a differential-privacy noise mechanism is applied to both quantities. The raw weekly search proportion is

$$P_{r,t} = \frac{N_{r,t}^{(\mathrm{vac})}}{N_{r,t}^{(\mathrm{total})}},$$

and the published description defines the normalized quantity as

$$\mathrm{VSI}_{r,t} \equiv \mathrm{scale}\times P_{r,t}.$$

For longer-term analyses, monthly averages are computed as

$$\overline{\mathrm{VSI}}_{r,m} = \frac{1}{|W_m|}\sum_{t\in W_m}\mathrm{VSI}_{r,t}.$$

These search-derived inputs are not identical to Query Corner’s discourse corpus, but they provide a closely related near-real-time signal about vaccination intent that can be combined with dashboard views and outreach logic (Malahy et al., 2021).

3. Retrieval-augmented generation and agentic reasoning

The public-concern version of Query Corner uses a two-iteration RAG pipeline. Each filtered comment is embedded with OpenAI’s text-embedding-ada-002 into a 1536-dim vector and stored in a LangChain-backed vector store. For a user query with embedding $v_q$ and document embedding $v_d$, similarity is computed by cosine scoring,

$N_{r,t}^{(\mathrm{vac})}$0

Iteration 1 performs broad top-$N_{r,t}^{(\mathrm{vac})}$1 retrieval by semantic similarity, initial ranking with longcontextreorder, and reranking with FlashRank or rank-T5-flan depending on corpus size. Iteration 2 applies contextual compression, reducing each comment to 20% of original length while preserving query-relevant content, followed by rescoring and a final cut retaining the top 50% of compressed, reranked comments. GPT-4o then synthesizes the final answer from the ranked snippets and the requested output format (Javed et al., 17 Jul 2025).

This architecture is explicitly framed as a hallucination-mitigation strategy. The generation prompt requires that all facts be traceable to retrieved snippets; compression and reranking reduce irrelevant noise; and one example Q&A is provided in the system prompt as few-shot guidance for faithful patterns. Evaluation is then performed with the RAGAs framework using answer faithfulness and relevance. In the reported formulation,

$N_{r,t}^{(\mathrm{vac})}$2

where $N_{r,t}^{(\mathrm{vac})}$3 is the generated answer, $N_{r,t}^{(\mathrm{vac})}$4 the reference answer, and $N_{r,t}^{(\mathrm{vac})}$5 the retrieved context (Javed et al., 17 Jul 2025).

The AI-VaxGuide adaptation defines a more explicitly agentic stack. Its block diagram consists of a Document Parser, Semantic Chunking, Embedding Encoder, Vector Store (ChromaDB), Hybrid Retriever, Agentic Controller, Prompt Engine, Vector Tools, and LLM Generator (Gemini 2.0). Dense retrieval uses cosine similarity, sparse retrieval uses BM25, and the published ensemble score is

$N_{r,t}^{(\mathrm{vac})}$6

The retriever uses top k=6 for dense retrieval and top k=2 for BM25, while Gemini generates $N_{r,t}^{(\mathrm{vac})}$7 reformulations for LLM-powered query expansion. Above this retrieval layer, a ReAct-inspired controller decomposes the user query into subtasks, selects section-specific tools, aggregates retrieved chunks, and performs a validation pass before answer emission (Zeggai et al., 4 Jul 2025).

A plausible implication is that Query Corner can support multiple retrieval modes depending on corpus type. In the discourse-monitoring setting, the unit of retrieval is a filtered social-media comment or compressed snippet. In the guideline-QA setting, the unit is a semantically chunked document section with explicit source highlighting. The papers do not collapse these into a single implementation, but they do show a consistent architectural pattern: retrieval, reranking, grounding, and controlled answer synthesis (Javed et al., 17 Jul 2025, Zeggai et al., 4 Jul 2025).

4. Surveillance linkage, early warning, and risk scoring

A defining feature of VaxPulse is the linkage between online discourse and formal surveillance. In the SAEFVIC-enhanced framework, time-series alignment is performed between weekly topic prevalence curves and weekly AEFI report counts. Lead-lag analyses use Pearson correlation, Bayesian updating, and Granger causality tests to determine whether surges in online discussion of a specific AEFI precede confirmed cases in the reporting registry. The reported Pearson statistic is

$N_{r,t}^{(\mathrm{vac})}$8

with $N_{r,t}^{(\mathrm{vac})}$9 denoting weekly topic volume and $N_{r,t}^{(\mathrm{total})}$0 weekly reported AEFI count. The same framework describes cross-correlation analysis for detecting leading indicators in online sentiment topics prior to formal AEFI spikes (Javed et al., 7 Jul 2025).

The VIRAL platform adds a daily risk-scoring layer. For time window $N_{r,t}^{(\mathrm{total})}$1 and region $N_{r,t}^{(\mathrm{total})}$2, the misinformation ratio is $N_{r,t}^{(\mathrm{total})}$3, the bot amplification score is $N_{r,t}^{(\mathrm{total})}$4, and the sentiment shift is $N_{r,t}^{(\mathrm{total})}$5. These are aggregated into a composite Infodemic Risk Index,

$N_{r,t}^{(\mathrm{total})}$6

with weights $N_{r,t}^{(\mathrm{total})}$7. Thresholds are reported as $N_{r,t}^{(\mathrm{total})}$8 for High alert, $N_{r,t}^{(\mathrm{total})}$9 for Moderate risk, and $r$0 for Low risk. Query Corner exposes these outputs through geo-map heat maps, time series, topic word clouds, misinformation explorers, sentiment and topic trend charts, and social-bot influence graphs (Dimaguila et al., 6 Jul 2025).

Search-intent analytics provide a parallel early-warning channel. Weekly VSI was linked to vaccinations three weeks later using the mixed-effects model

$r$1

with a random intercept for state. Reported associations were $r$2 additional first doses per $r$3 per 10-unit increase in VSI for the week of January 11, 2021, $r$4 for the week of March 22, 2021, and $r$5 for the week of August 16, 2021. Early monthly VSI also correlated with later vaccination rates, with state-level correlations reaching $r$6 for April 2021 VSI versus May–June vaccination. This suggests that search-behavior signals can enrich Query Corner’s surveillance layer by identifying communities with high expressed interest but lagging uptake (Malahy et al., 2021).

5. Operational outputs, intervention logic, and empirical performance

The operational logic of VaxPulse includes both dashboards and decision rules. A published “VaxPulse quadrants” framework classifies regions by tertiles of current VSI and cumulative vaccination rate into Q1 (High VSI & High Vax), Q2 (High VSI & Low Vax), Q3 (Low VSI & Low Vax), and Q4 (Low VSI & High Vax). Q2 is interpreted as likely “access barriers,” with possible interventions including mobile clinics, extended hours, and transport vouchers; Q3 is interpreted as likely “engagement barriers,” with possible interventions including community ambassadors, culturally tailored messaging, and local media campaigns. An example VaxPulse-style dashboard flow includes a time-series panel of average VSI and vaccination rates, a map panel by quadrant, a tabular report of top 10 “Q2 hotspots,” automated alerts for movement into Q2 or Q3 for at least two consecutive weeks, and a forecast module

$r$7

These components give Query Corner a prescriptive as well as descriptive role (Malahy et al., 2021).

The Shingrix Query Corner study provides the clearest end-to-end RAG evaluation. From a corpus of 60,935 raw posts spanning January 2018 to October 2024, 35,103 vaccine-specific posts remained after BERT+GPT filtering. Retrieval precision and recall improved from 0.56 and 0.85 in Iteration 1 to 0.66 and 0.91 in Iteration 2. Average answer scores by query type were 0.90 faithfulness and 0.89 relevance for Q&A, 0.90 and 0.91 for topics of discussion, 0.96 and 0.94 for summarise, and 0.88 and 0.82 for public concerns. The abstract foregrounds the same headline result: analysing 35,103 Shingrix social media posts, the tool achieved answer faithfulness 0.96 and relevance 0.94 (Javed et al., 17 Jul 2025).

The SAEFVIC-linked VaxPulse framework reports a different evaluation profile. Its sentiment classifier is CT-BERT V2 fine-tuned on 13,715 annotated social media posts around AEFI, with 93% cross-validated accuracy, and comments below a confidence threshold $r$8 are re-scored by GPT-4o so that the ensemble decision maximizes $r$9. On held-out test sets, sentiment classification reached $t$0. BERTopic topic coherence averaged 0.52 (UMass), outperforming standard LDA at 0.37, and early-warning correlations exceeded Pearson $t$1 at lag $t$2 week for myocarditis posts and AEFI reports. In the women’s vaccine hesitancy case study, the dataset contained 819,388 cleaned comments, of which 705,145 (86.1%) were identified as COVID-19 vaccine-related; sentiments were negative 9.87%, neutral 62.90%, and positive 27.23%; and female-focused topic prevalence was 18% for irregular menstrual cycles, 12% for pregnancy advice, and 7% for miscarriage concerns. Topic volumes peaked at approximately 800 menstrual-concern comments in November 2023 and approximately 1,200 in May 2024, after which SAEFVIC deployed targeted TikTok/Instagram videos (Javed et al., 7 Jul 2025).

The AI-VaxGuide adaptation provides a performance profile for the agentic clinical-QA variant. On the Vaccination-Guide QA Benchmark, AgenticRAG obtained an average score of 0.73, average time 16.72 s, and citation rate 100%; EnhancedRAG obtained 0.43, 5.10 s, and 100%; SimpleRAG obtained 0.03, 20.48 s, and 100%; and the Fine-Tuned baseline obtained 0.00, 0.00 s, and 0. In category-level results, AgenticRAG accuracy was 0.50 on fact-based questions, 1.00 on complex questions, and 0.70 on cross-document questions. The mobile implementation, built with React Native and FastAPI, reports p95 latency of approximately 300 ms in EnhancedRAG mode and approximately 1.1 s in AgenticRAG mode, multilingual input in French, Arabic, and English, inline source highlighting, and per-clinician chat history (Zeggai et al., 4 Jul 2025).

6. Limitations, methodological tensions, and related research

The published limitations are substantial and recur across implementations. The Shingrix RAG system reports latency of approximately 3–5 s due to two-iteration reranking and compression, coverage restricted to English-only posts, and no incorporation of non-textual media such as images or video transcripts. The SAEFVIC framework notes that polarity classification omits nuanced hesitancy subtypes, real-time API rate limits constrain latency, and future work should extend AEFI linkage with geotagged data. The VIRAL platform likewise places Query Corner inside a feedback loop that still depends on human annotation, active learning, and stakeholder correction rather than purely autonomous inference (Javed et al., 17 Jul 2025, Javed et al., 7 Jul 2025, Dimaguila et al., 6 Jul 2025).

A second limitation concerns linguistic and cultural equity. VaxPulse explicitly retains source-language metadata to stratify concerns across ethno-lingual communities, and it emphasizes the need to address non-English languages. Related work on LLM vaccine messaging shows that responses can vary significantly across languages: using the Vaccine Hesitancy Scale, English responses were the most hesitant on average, Spanish the least, with overall VHS means of 1.90 for English, 1.40 for Spanish, and 1.50 for French, and one-way ANOVA yielding $t$3, $t$4. This does not directly evaluate Query Corner, but it indicates that multilingual deployment cannot be reduced to translation alone; response consistency is itself an evaluative target (Joshi et al., 2024).

A third tension concerns training data and benchmark realism. Vax-Culture contributes a 6,373-tweet dataset annotated for vaccine-hesitancy stance, misinformation, criticized entities, supported entities, and intended meaning, with baseline results such as 75.6% accuracy and 70.7 F1 for RoBERTa-large on misinformation detection. This suggests a concrete upstream resource for classifiers used in vaccine-discourse analysis, while also illustrating the difficulty of rare multi-label classes and open-ended meaning generation. A plausible implication is that Query Corner performance depends not only on retrieval and prompting, but also on the quality and granularity of the supervised filters that determine which posts enter the retrieval index in the first place (Zarei et al., 2023).

Taken together, the literature defines VaxPulse Query Corner not as a single fixed application but as a family of grounded vaccine-information interfaces. In one form it is a RAG system over public discourse with measured faithfulness and relevance; in another it is the dashboard layer of a broader infodemic risk-assessment lifecycle; and in a third it is an agentic QA feature adapted from vaccination-guideline retrieval. Across these variants, the common technical commitments are evidence-grounded retrieval, explicit preprocessing and filtering, linkage to surveillance or protocol sources, and operational outputs designed for intervention rather than observation alone (Javed et al., 17 Jul 2025, Dimaguila et al., 6 Jul 2025, Zeggai et al., 4 Jul 2025).