Skip to content

Bibby AI: An Editor-Native Agentic Platform for Academic Research, Writing, and Publishing

🕒 Published (v1): 2026-07-03 00:00 UTC · Source: HuggingFace · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Bibby AI is a cloud LaTeX editor that unifies literature discovery, reference management, writing, and venue formatting into a single agentic platform. By owning the full document model and compilation pipeline, its agents can perform compile-verified structural edits and retrieval-grounded citation insertion as first-class operations. The platform serves 5,000+ researchers across 50+ universities and models ~7.6 hours of monthly time savings per user.

Problem

Academic writing is fragmented across disconnected tools (search engine → reference manager → LaTeX editor → format converter → submission portal), forcing repeated context switches, format conversions, and manual copy-paste steps. LLM assistants bolted onto individual stages (e.g., browser extensions on Overleaf) cannot access full document state, revision history, or the compilation pipeline, making their edits unverifiable and limiting them to text suggestions rather than structural operations.

Method

Bibby AI is built as a standalone cloud LaTeX editor (replacing Overleaf, not extending it) where the editor, server-side compiler, reference graph, and agents share one source of truth: the project store. Four integrated layers:

  1. Editor/Compilation Core: Browser-based LaTeX environment with server-side pdflatex in isolated containers. Every agent-proposed edit is applied to a shadow copy, compiled, and rejected or auto-repaired before being presented as a diff — converting "plausible but broken LaTeX" into an internal retry loop.

  2. Ingestion Pipelines: PDF → LaTeX (layout-aware reconstruction), DOCX → LaTeX (style-to-sectioning mapping), and handwritten math → LaTeX (OCR-based recognition), each terminating in a compile-validated live project.

  3. Retrieval and Citation Layer: Semantic/metadata search over Semantic Scholar and OpenAlex, with direct BibTeX insertion. Augmented with a translational impact signal per paper computed by joining the scholarly record against USPTO PatentsView and the Marx–Fuegi front-page patent citation corpus — surfacing downstream technological use at citation-choice time.

  4. Agent Layer: Task-scoped agents operating on the document's abstract syntax representation. Single-shot agents handle low-latency revisions; workflow agents coordinate retrieval, generation, and compilation validation for literature triage, full-document review, and one-click venue retargeting (preamble/sectioning/bibliography style rewrite, compile-verified).

Key Contributions

  • Editor-native agentic architecture where every structural edit is validated by the compiler before surfacing to the user, eliminating the synchronization/patch-conflict problems of plugin approaches.
  • Ingestion pipelines (PDF/DOCX/handwriting → compilable LaTeX) with in-platform compile checking, removing the largest onboarding cost of LaTeX workflows.
  • Translational-impact retrieval signal derived from patent-to-paper citations (PatentsView + Marx–Fuegi corpus), unavailable in any incumbent writing platform.
  • Deployment report and workflow-level time-savings evaluation framework instantiated with production telemetry; platform live at 5,000+ researchers, 50+ universities.

Results

Time savings modeled via \(T(w) = \sum_{i=1}^{n_w}(t_i^{\text{task}} + t_i^{\text{switch}} + t_i^{\text{repair}})\) and \(E[S_{\text{month}}] = \sum_{w \in W} f_w S(w)\):

  • Citation search → cited paragraph: baseline 25 min → 6 min (saves 19 min/instance × 8×/month = 152 min)
  • DOCX → LaTeX project: 90 min → 8 min (saves 82 min × 1×/month)
  • Venue retargeting: 180 min → 20 min (saves 160 min × 0.5×/month)
  • Handwritten math → section: 30 min → 4 min (saves 26 min × 2×/month)
  • Compile-error debugging: 20 min → 5 min (saves 15 min × 6×/month)
  • Total modeled saving: \(E[S_{\text{month}}] \approx 456\ \text{min} \approx 7.6\ \text{h}\) per researcher per month
  • Aggregate: ~38,000 researcher-hours recovered monthly across the user base
  • Note: these are modeled estimates from onboarding interviews; production telemetry validation is ongoing and not yet reported.

Limitations

  • Time-saving numbers are modeled estimates from onboarding interviews and task decomposition, not yet validated against production telemetry.
  • Patent citation signal undercounts science–technology linkage (in-text citations and non-patented uses invisible); coverage is USPTO-centric, excluding international patent offices.
  • Retrieval-grounded drafting can select real-but-suboptimal references; automated claim–citation entailment checking is planned but not implemented.
  • Editor-nativeness requires users to migrate from Overleaf, making migration friction the primary growth constraint.
  • No controlled user study or ablation against baselines reported; capability comparisons (Table 1) are self-assessed.

Relevance to Agentic AI / LLM Agents

Bibby AI is a concrete existence proof that grounding agents in a platform's own state representation (document AST + live compiler) enables qualitatively different capabilities than plugin-based LLM assistants — specifically, compile-verified edits and multi-step workflow execution that are architecturally impossible from outside the editor. The compile-as-validator pattern (shadow-copy edit → compile → reject/repair → diff) is a transferable design for any domain where a deterministic checker can replace model self-evaluation. The paper also illustrates task-scoped agent decomposition (single-shot vs. workflow agents) in a production system, and demonstrates how retrieval augmentation can be enriched with domain-specific, non-textual signals (patent citations) to produce value no general-purpose RAG system offers.