Selenite: Browser-Based Online Sensemaking

Updated 7 February 2026

Selenite is a browser-integrated sensemaking scaffold that leverages LLMs to automatically generate diverse, high-coverage overviews without requiring extensive user input.
It integrates modular components like retrieval, LLM prompting, criteria extraction, semantic clustering, and adaptive personalization to enhance online evaluation and navigation.
Empirical studies show Selenite reduces task completion time by 36% and boosts valid criteria identification, highlighting its practical impact on streamlined sensemaking.

Selenite is a browser-integrated system for scaffolding online sensemaking, designed to address the cold-start and incompleteness challenges of existing sensemaking tools by harnessing LLMs as reasoning engines and context-adaptive assistants. Unlike prior approaches that depend on extensive prior user input or manual structuring, Selenite automatically produces high-coverage, criterion-driven overviews of complex information spaces, supports adaptive navigation and personalization, and incorporates algorithmic summarization and content clustering. Selenite thus represents a highly instrumented, LLM-augmented ‘in-situ’ sensemaking scaffold that enables users to systematically compare options, coordinate criteria for evaluation, and externalize their cognitive processes within the browser context (Liu et al., 2023).

1. Theoretical Foundations and Motivations

Selenite arises from tensions in traditional online sensemaking frameworks. Early work on stigmergic social annotation (&&&1&&&) and browser-based working memory tools (e.g., Fuse (Kuznetsov et al., 2022)) established that indirect coordination via markers (explicit and implicit) and lightweight project externalization facilitates emergent filtering, interpretation, and group sensemaking. However, these models depend on either platform-centric control (opaque algorithms, slow knowledge accumulation) or place excessive manual overhead on individual users for organization and annotation. Selenite’s introduction of LLMs as dynamic overviews and extractive engines mitigates the “cold-start problem,” jumpstarting sensemaking tasks even in the absence of prior user data, and shifts the focus from user-generated markers to automated, coverage-oriented semantic extraction and adaptive scaffolding (Liu et al., 2023).

2. System Architecture and Pipeline

Selenite is architected as a modular browser sidebar integrating five core components: retrieval, LLM prompting, criteria/option extraction, adaptation, and UI feedback. The pipeline operates as follows:

Topic Recognition: On visiting a page, Selenite infers the information topic via LLM semantic clustering, using page titles and snippets.
Overview Generation: Through template prompts, the LLM enumerates high-relevance criteria (“List the top N criteria people consider...”) and iteratively self-refines the set for diversity and coverage, employing semantic similarity to deduplicate.
Option Extraction: A zero-shot NLI model (e.g., BART-mnli) detects mentions of relevant options, which are then annotated in situ.
Local Grounding: Paragraphs are labeled with corresponding criteria and sentiment using NLI classification, enabling fine-grained navigation.
Progress Feedback and Navigation: User dwell time, clicks, and edits are logged to update importance weights for criteria (e.g., $w_i \leftarrow w_i + \eta \cdot \delta$ where $\delta=1$ if dwell $\geq2$ s).
Personalization: User modifications (pinning, reordering) persist across sessions and inform adaptive refinement, blending global and user-specific salience [ $W_i = \beta G_i + (1-\beta) P_i$ ].

This architecture supports both global grounding (coverage and relevance across a topic) and local grounding (evidence and sentiment at the paragraph or snippet level), unifying strategic overview with tactical reading assistance (Liu et al., 2023).

3. Algorithms and Workflow

Selenite’s core automated functions rely on LLM-driven prompt engineering, semantic similarity metrics, and multi-stage selection strategies:

Prompt Templates: The system uses chain-of-thought LLM prompts to elicit extensive, diverse, and relevant criteria sets, complemented by user-correctable actions.
Self-Refine Loop: Iteratively requests additional, non-redundant criteria until coverage plateaus.
Deduplication and Diversity Optimization: Relevance is calculated from LLM scores; diversity as $D(c, c') = 1 - \text{cosine}(\text{Embed}(c), \text{Embed}(c'))$ . Subset maximization optimizes relevance-diversity tradeoff.
Option and Criteria Grounding: In-document option mentions are highlighted using NLI classifiers. At the paragraph level, criteria and sentiment are assigned to create annotated evidence trails.
Personalization: Importance weighting dynamically blends LLM-derived (global) and interaction-logged (personal) relevance.

The pseudo-algorithm for core overview generation is:

def GenerateOverview(topic, pageContent):
    criteriaList = LLMQuery("List top 10 criteria for [topic]…")
    for iteration in range(R):
        newCriteria = LLMQuery("List 5 more diverse criteria not in [criteriaList]…")
        criteriaList = merge(criteriaList, newCriteria)
    criteriaList = deduplicate(criteriaList)
    optionsList = ExtractOptions(pageContent)  # LLM-based Named Entity style
    return criteriaList, optionsList

(Liu et al., 2023)

4. Interface and User Interaction Patterns

Selenite’s interface manifests as a persistent browser sidebar with several affordances:

Global Criteria/Option List: Displays sortable and expandable names, descriptions, and derived importance scores.
Paragraph and Option Highlighting: In-page highlights (border tags, overlays) show where evidence for each criterion resides. Hovering enables granular analysis (criterion/sentiment segmentation).
Progress Summaries and Suggestions: End-of-page analysis summarizes covered criteria, suggests additional exploration axes (“consider searching for [c3, c4…]”), and surfaces one-click navigation to evidence instances.
Personalized Editing: Users may pin, hide, reorder, or add criteria/options, with persistent user profiles guiding future adaptation.
Navigation Shortcuts: Criterion-wise next/previous controls jump to evidence in sequence, supporting systematic review.
Case Study Visualizations: Example layouts highlight the interplay between overview, option extraction, and in-document analysis.

This interface model is designed to be minimally intrusive yet workflow-integrated, providing real-time, adaptive scaffolding for online sensemaking (Liu et al., 2023).

5. Empirical Evaluation and Impact

Three empirical studies establish Selenite’s efficacy and usability:

Intrinsic Evaluation: On ten curated topics, Selenite’s composite criteria recall reached 95%, with 80% precision against human-generated ground truth.
Usability Study (N=12):
- Selenite reduced task completion time by 36% relative to the Unakite baseline (840 vs. 1,319 seconds, $F(1,20)=102.5$ , $p<.01$ ).
- Users identified 90.4% more valid criteria (12.93 vs. 6.79, $p<.01$ ) and performed 6× more additional searches, with 5× more new criteria explored.
- NASA TLX scores indicated significant reductions in mental demand (3.03 vs. 6.43, $p<.05$ ) and subjective workload.
Open-ended Case Study (N=8): Selenite exposed criteria not anticipated in the seed set (e.g., “Licensing,” “Motion Transfer”), indicating adaptive headroom for exploration.

Qualitative themes from these studies include increased comprehensibility, improved navigation, and enhanced time savings. However, Selenite’s automated overviews also introduce user trust calibration challenges and possible over-reliance risks (Liu et al., 2023).

6. Relationships to Prior Work and Comparative Systems

Selenite generalizes and extends functionality previously seen in systems such as Fuse (Kuznetsov et al., 2022) and stigmergic social annotation frameworks (Tamari et al., 2022):

From Manual Collection to Semantic Scaffolding: Fuse’s low-cost clipping, flexible nesting, and visual compression are retained, but Selenite adds automatic clustering (using cosine similarity and temporal weighting on embeddings), task templates, and transformer-based summarization for reduced manual effort.
Table: Affordance Comparison (extract, see (Kuznetsov et al., 2022)):

Feature	Fuse	Selenite
In-situ sidebar integration	✔️	✔️
Manual clipper	✔️	✔️ + adaptive selection
Flexible nesting	✔️	✔️ + auto-suggested clusters
Visual compression	✔️	✔️ + density-aware layouts
Automation of grouping	—	✔️ via semantic scoring
Summarization	—	✔️ transformer-based
Task templates	—	✔️

Relation to Stigmergic/OSA Models: Selenite diverges from distributed marker-centric approaches, opting for automated, LLM-based knowledge extraction over stigmergic consensus annotations. It does not implement decentralized pod storage or explicit marker interoperability but offers alternative scaffolding through LLM world knowledge and dynamic interaction adaptation (Tamari et al., 2022).
Structured Sensemaking Scaffolds: Compared to SnuggleSense (Xiao et al., 27 Apr 2025), which targets trauma-informed guidance for online harm survivors, Selenite emphasizes information-rich, criterion-adaptive navigation and coverage, using LLMs rather than peer-driven recommendation and restorative action planning.

7. Limitations and Open Research Challenges

Selenite’s limitations include:

LLM Hallucination and Knowledge Gaps: For rare or specialized domains, LLM-based summaries may be biased or incomplete. The use of retrieval-augmented generation (RAG) is a proposed mitigation.
Coverage Calibration and Confidence: The system lacks robust uncertainty modeling—there is a need for adaptive self-refine strategies and coverage indicators.
Dynamic Content Handling: Real-time context extraction from video, interactive elements, or non-text modalities remains open.
Over-Reliance and User Learning: Prolonged use may foster dependency on generated overviews, potentially reducing deep engagement.
Scalability and Deployment: Moderate computational overhead (embedding, summarization) may limit real-time performance on resource-constrained devices. Large-scale, longitudinal deployments are needed to assess long-term productivity effects and social impact (Liu et al., 2023).

This suggests future work will involve deeper study of adaptive scaffolds combining strengths of decentralized marker approaches and LLM reasoning, development of robust user trust calibration mechanisms, and exploration of group sensemaking workflows that interleave automated and stigmergic annotation.

Selenite embodies a next-generation paradigm for browser-centered online sensemaking, synthesizing LLM-powered overview generation, adaptive interface scaffolding, and in-situ evidence navigation. Its empirical validation establishes a new baseline for the integration of automated reasoning, human feedback loops, and semi-structured information organization in epistemic environments (Liu et al., 2023, Tamari et al., 2022, Kuznetsov et al., 2022, Xiao et al., 27 Apr 2025).