OverCite: Streamlining LaTeX Citations

• OverCite is an open-source, lightweight citation tool that uses contextual search and rule-based ranking to reduce drafting interruptions.
• It integrates directly into editors such as Overleaf and VS Code by combining local sentence context with ADS/SciX query results.
• Its deterministic design avoids LLMs and focuses solely on precise in-editor citation insertion rather than comprehensive bibliography management.

Searching arXiv for the specified paper and closely related citation-attribution work. OverCite is an open-source, lightweight citation tool for adding LaTeX citations without leaving the editor. It is designed to remove a specific drafting interruption: an author stops writing, searches for a paper already in mind, copies its BibTeX entry into the project bibliography, renames the cite key, and returns to the manuscript. In Overleaf, OverCite uses rough citation placeholders such as \citep{Perlmutter1999} together with local sentence context to query ADS/SciX-indexed literature, rank likely matches, and insert the selected reference; a companion VS Code extension provides the same functionality for local LaTeX projects (Shariat, 14 Apr 2026).

1. Problem setting and design target

The system is built for the case in which the intended paper is already roughly known, but the conventional citation workflow is still inefficient. The paper describes that workflow as involving four tasks: identify the intended paper, insert its BibTeX entry into the project bibliography, update the citation key in the manuscript, and do all of this inside the editor. OverCite therefore targets the in-editor act of citation insertion rather than bibliography management in the broader sense.

Its core idea is narrow and explicit. OverCite uses a rough citation placeholder in the manuscript, the local sentence context around that citation, and the ADS/SciX literature database to search and rank candidate matches. The implementation is deterministic and does not use a LLM. Instead, it relies on typed citation cues, local sentence context, and rule-based ranking of ADS/SciX results. This design makes the tool specialized rather than general-purpose: it focuses on reducing interruption during drafting, not on open-ended literature exploration or automated prose generation.

The corpus accessed through ADS/SciX includes astronomy, physics, computer science, mathematics, biology, and all indexed arXiv e-prints. The intended utility is therefore not restricted to astronomy, even though ADS has strong historical associations with that domain. The paper explicitly presents OverCite as useful across a broad range of scientific disciplines.

2. Operational workflow inside the editor

The workflow described in the paper is editor-centric and cursor-driven. The user places the cursor inside a citation command in the editor, triggers OverCite from within the editor, and the active citation token is parsed. That token, together with the surrounding sentence context, is used to search ADS/SciX. Candidate papers then appear in a popup, the user selects the intended record, and OverCite updates the citation key in the manuscript while inserting the corresponding BibTeX entry into the project bibliography.

The figure caption condenses this interaction into a single loop: a partial citation key in the manuscript is interpreted together with surrounding sentence context, ADS/SciX matches are shown in a popup, and the user selects a record to insert its BibTeX entry directly into the project bibliography. The example figure shows a partial citation key highlighted in yellow, surrounding sentence context in purple, and a popup with ADS/SciX matches on the right; the intended citation is \citet{Shariat2025}.

A notable feature of this workflow is that the citation placeholder is not merely a temporary string. It acts as a search cue that can encode approximate author, year, initials, title fragments, or collaboration names. The sentence context then functions as an additional constraint during ranking. This suggests a model of citation insertion in which partially specified intent, rather than a finished BibTeX key, is the primary user input.

3. Search modes, query forms, and ranking behavior

OverCite supports three search modes. In Contextual search, the typed token is combined with local sentence context. In Simple search, sentence context is ignored and the system searches typed author/year information alone, sorted by citation count. In ADS-query mode, the typed token is sent directly to ADS/SciX as a literal query (Shariat, 14 Apr 2026).

The paper gives a concrete range of supported query forms:

\citep{Perlmutter1999} \citep{Shariat25} \citep{Smith25} \citep{SmithJ25} \citep{Abbott} \citep{Hawking1975} \citep{Astropy Collaboration} \citep{Gaia Collaboration2021} \citep{title:"emcee"} \citep{author:"schlegel" maps of dust} \citet{Shariat2025}

These examples indicate that the token can be minimal or relatively constrained. Disambiguation cases such as \citep{Smith25} can be narrowed with initials, as in \citep{SmithJ25}. The year may be omitted, as in \citep{Abbott}, or fully specified, as in \citep{Hawking1975}. Collaboration names are also supported.

Ranking is rule-based. Candidate papers are ranked using the citation token and the surrounding sentence context, with the context acting as an additional check when many possible matches exist, especially for common surnames such as \citep{Smith25}. In simple search, results are sorted by citation count. Non–first-author surname matches are supported, but they generally rank below first-author matches, and first initials can be used to help filter matches. The resulting behavior is a deterministic disambiguation scheme rather than a generative or probabilistic citation assistant.

4. Editor integration and implementation architecture

OverCite operates in two editor environments: an Overleaf browser extension and a VS Code extension for local LaTeX projects. The public distribution channels listed in the paper are a Chrome extension, a Firefox Add-on, and a VS Code extension. This dual deployment model covers both browser-based collaborative writing and local LaTeX workflows.

The paper characterizes the tool as lightweight for several concrete reasons. It is focused narrowly on citation insertion during drafting; it uses the existing ADS/SciX APIs rather than maintaining a heavy backend; each user authenticates with their own ADS/SciX token; API usage is therefore distributed per user rather than routed through a shared OverCite backend; and the implementation is deterministic and rule-based, not LLM-based. ADS currently allows 5000 regular requests per day for most request types, with higher limits available by request.

Project-level variability is handled through settings. The tool supports preferred citation-key style, bibliography entry order, and selection of the target bibliography file when multiple .bib files are present. The current public implementation is stated to be v0.1.3. These details indicate that the system is meant to adapt to heterogeneous LaTeX project conventions rather than impose a single bibliography workflow.

5. Scope, limitations, and intended use

The paper is explicit about what OverCite is not. It is not meant to replace reference managers such as Zotero or Mendeley, which are described as tools for collecting references, organizing bibliographies, and maintaining libraries. OverCite instead focuses on the in-editor act of locating a known paper and inserting it.

It is also not intended for broader literature discovery, exploration, or review of unfamiliar work. The paper says it is best suited for querying known papers during writing. A common misconception would therefore be to interpret OverCite as either a general reference-management system or a literature-review engine; the text rejects both characterizations. Its proper scope is narrower and more operational: authoring-time retrieval and insertion of a citation already envisaged by the writer.

The disciplinary scope is broader than the interface might initially imply because ADS/SciX includes astronomy, physics, computer science, mathematics, biology, and all indexed arXiv e-prints. Future versions could extend the framework to more editors and browsers, writing environments beyond LaTeX, alternative bibliography styles, improved search and ranking, and additional literature databases. This suggests an expandable framework, but the paper presents the current system as a focused drafting aid rather than a universal citation platform.

6. Position within citation, attribution, and bibliographic infrastructure research

OverCite belongs to a different layer of the citation stack from systems that generate or evaluate evidence-grounded citations. FullCite, for example, studies structured inline citation generation in which each claim is linked to both a source document and a supporting evidence span, and its central finding is that document identification is often easier than precise evidence-span localization (Yeginbergen et al., 5 Jun 2026). FORCEBENCH similarly argues that a citation can be present and topically relevant yet still under-warrant the strength of the attached claim, formalizing this as evidence-force calibration (Qian et al., 27 May 2026). OverCite does not attempt claim-level support verification or span-level grounding; it addresses the earlier problem of inserting the intended bibliographic record into a LaTeX document.

It is also distinct from open citation-data infrastructures. OpenCitations is an independent, not-for-profit open e-infrastructure for publishing open bibliographic and citation data using Semantic Web and Linked Data technologies, aligned with the UNESCO founding principles of Open Science, the I4OC recommendations, and the FAIR data principles (Giambattista et al., 2022). COCI, the OpenCitations Index of Crossref open DOI-to-DOI citations, treats citations as first-class data entities and exposes them through RDF, SPARQL, REST, bulk downloads, and HTTP content negotiation (Heibi et al., 2019). OverCite, by contrast, is not a citation index; it is an editor-integrated interface for querying ADS/SciX and inserting BibTeX entries during drafting.

A further contrast appears with citation parsing research. “Citation Parsing and Analysis with LLMs” evaluates open-weight LLMs for converting plaintext citations into structured JATS fields and reports strong performance relative to Crossref search and GROBID (Sarin et al., 21 May 2025). OverCite takes the opposite architectural path: its paper emphasizes deterministic, rule-based behavior and explicitly excludes LLMs from the core search-and-ranking loop. This suggests complementary roles rather than overlap. Parsing work addresses the transformation of existing references into structured metadata, while OverCite addresses author-time insertion of references before that parsing problem arises.

Taken together, these comparisons place OverCite at the authoring interface: it sits between the writer’s rough citation intent and the project bibliography. Its novelty is not claim verification, citation-network construction, or reference parsing at corpus scale, but the reduction of a highly localized interruption in LaTeX drafting.