WisPaper: AI Academic Retrieval Platform

• WisPaper is an AI-driven academic retrieval platform that integrates search, organization, and recommendation into a unified, closed-loop workflow.
• The platform’s modular design, featuring Scholar Search, Library, and AI Feeds, enables both rapid keyword searches and deep, agentic exploration of literature.
• Utilizing advanced methods like BM25 scoring and semantic matching, WisPaper reduces search latency and improves systematic literature curation.

WisPaper is an AI-driven academic retrieval and literature management platform designed to address the challenges posed by the exponential growth of scientific publications. By integrating efficient search, customizable organization, and continuous recommendation within a single environment, WisPaper aims to reduce manual overhead for researchers in academia and industry, offering support across multiple languages and disciplines (Ju et al., 7 Dec 2025).

1. System Overview and Architecture

WisPaper is architected as an end-to-end closed-loop platform comprising three tightly coupled modules: Scholar Search, Library, and AI Feeds. The core of the system is a unified paper database and metadata store, guaranteeing seamless data flow and interoperability across modules. This architecture operationalizes a “Knowledge Workflow” in which literature discovery, systematic curation, and frontier tracking are interdependent rather than siloed.

• Scholar Search initiates the workflow, offering dual-mode retrieval for both rapid and deep academic exploration.
• Library provides a hierarchical, customizable environment for systematic organization and detailed active reading.
• AI Feeds autonomously delivers newly published, semantically relevant works based on adaptive models of a user's interests.

This closed feedback loop is designed so that users search, curate, and receive updated recommendations in a continuous cycle, each action refining subsequent retrievals and recommendations (Ju et al., 7 Dec 2025).

2. Scholar Search: Retrieval Paradigms

WisPaper’s Scholar Search module implements two distinct paradigms:

a) Quick Search

Quick Search employs a high-throughput inverted-index system such as Elasticsearch with BM25 scoring: $$\mathit{score}_{\mathrm{BM25}}(q,d) = \sum_{t\in q} \mathrm{IDF}(t)\,\frac{f(t,d)\,(k_1+1)} {f(t,d) + k_1(1 - b + b\,|d|/\overline{d})}$$ This enables retrieval with sub-20 ms latency, surfacing key metadata (title, authors, venue, date, citation count, abstract snippet) and supporting user interaction features including like/dislike, one-click library import, and query re-execution from history.

b) Deep Search (Agentic Exploration)

Deep Search is driven by an autonomous agent “WisModel” orchestrating a two-stage workflow:

1. Query Understanding / Criteria Generation: Free-form natural language queries are parsed into Boolean sub-queries and a weighted set of explicit criteria $\{(c_i,w_i)\}_{i=1}^m$ with $\sum_i w_i=1$. 2. Paper Validation: Each candidate document $d$ is scored according to programmed judgments %%%%3%%%%, and aggregated:

   $$\mathrm{Score}(d) = \sum_{i=1}^m w_i \mathbf{1}[a_i = \mathrm{support}] + \frac{1}{2}\sum_{i=1}^m w_i \mathbf{1}[a_i = \mathrm{somewhat\_support}]$$

Users can refine criteria and rerun the agent until the desired corpus is obtained (Ju et al., 7 Dec 2025).

3. Library: Hierarchical Knowledge Base and Embedded Tools

The Library module transforms WisPaper into a full-scale knowledge management system characterized by:

• Hierarchical Taxonomies: Unlimited nested folders mirror domain ontologies or user project structures. Cross-library or subtree-constrained search is supported.
• Automated Metadata Extraction: On PDF upload or direct import, title, authors, affiliations, abstract, venue, date, DOI, and references are parsed via PDF structure analysis.
• Embedding-Driven Search: Abstracts and notes are encoded using models such as OpenAI’s text-embedding-3-large into $e(d) \in \mathbb{R}^n$. Semantic similarity for queries is computed by cosine:

  $$\text{sim}(q,d) = \frac{e(q)\cdot e(d)}{\|e(q)\|\|e(d)\|}$$

• Integrated Reading and Export: Features include structured summaries, community/public annotations, full-text search on highlights/notes, and one-click export to BibTeX, RIS, EndNote, as well as two-way sync with Zotero/Mendeley (Ju et al., 7 Dec 2025).

4. AI Feeds: Recommendation and Frontier Tracking

AI Feeds consists of a multistage, feedback-driven recommendation engine:

• 1. Category-Level Filtering: User-selected broad categories (e.g., arXiv cs.CL) determine the ingestion of all new papers $P_{\text{new}}$ since last query.
• 2. Semantic Matching: User profile vector

  $u = \frac{1}{|L|}\sum_{d\in L} e(d)$

  is constructed from Library. Each new paper $p$ receives relevance score

  $s(u,p) = \frac{u\cdot e(p)}{\|u\|\|e(p)\|}$

  with the highest-$k$ papers recommended.

• 3. Implicit Feedback Loop: Libraries additions and dismissals drive a pairwise learning-to-rank update:

  $$\mathcal{L}(\theta) = -\sum_{(p^+, p^-)} \ln \sigma(s_\theta(u,p^+) - s_\theta(u,p^-))$$

  where $\sigma$ is the sigmoid function. This optimizes the relevance function over time based on user preferences (Ju et al., 7 Dec 2025).

5. Multilingual and Multidisciplinary Adaptation

WisPaper provides full cross-lingual corpus and interface support. WisModel is trained and evaluated using 2,777 authentic queries in English and Chinese across 10 disciplines (CS, Medicine, Biology, Physics, Economics, Social Sciences). Criteria generation and prompt templates are language-agnostic, enabling domain-general adaptation. The user interface is localized for input and metadata display in supported languages (Ju et al., 7 Dec 2025).

6. Performance Benchmarks

Extensive evaluation described in (Ju et al., 7 Dec 2025) demonstrates the superiority of WisModel (Deep Search agent) over established LLMs and retrieval agents, specifically:

• Semantic Similarity (Criteria Generation): 94.8% (cosine), ROUGE-L 67.7%, BLEU 39.8%, all higher than GPT-4o, GPT-5.1, Qwen-Max, DeepSeek.
• Paper–Criteria Matching: 93.70% overall accuracy with breakdown: support 94.38%, somewhat_support 91.82%, reject 94.54%, insufficient_information 90.64%. Next best system reaches only 73.23%.
• Retrieval Latency (Quick Search): <20 ms.

7. Comparison with and Differentiation from Existing Tools

WisPaper is distinguished from other academic tools by its closed-loop integration:

Tool	Keyword Search	Knowledge Base	AI Feeds/Recommendation	Agentic Search/Validation	Closed-loop Workflow
WisPaper	Yes	Yes	Yes	Yes	Yes
Google/Semantic Sch.	Yes	No	No	No	No
Zotero/Mendeley	No	Yes	No	No	No
Connected Papers	No	No	No	No	No
ResearchArena, etc.	No	No	No	Yes	No

WisPaper is unique among surveyed tools in unifying rapid keyword retrieval, deep agent reasoning, systematic knowledge management, and automated research frontier tracking in a single platform (Ju et al., 7 Dec 2025).

8. User Interface and Workflow Features

WisPaper adopts a single-pane environment, enabling seamless transition between search, organization, and feeds without context switching. Core features include: like/dislike feedback, one-click library import, collaborative refinement of Deep Search criteria, AI summaries, community and private annotation, full-text search in notes/highlights, and a plugin ecosystem with Zotero/Mendeley sync and standard citation export for manuscript preparation (Ju et al., 7 Dec 2025).