Quotely — AI Citation Autocomplete for VS Code

Quotely is a VS Code extension that suggests relevant citations as you write academic papers in LaTeX or Markdown — powered entirely by your local document collection, with no data sent to the cloud.

What is Quotely?

When writing a scientific article, finding the right source at the right moment is the main friction point. Quotely solves this:

• You write a sentence like "The algorithm converges under convexity assumptions..."
• Press Cmd+Shift+C (or type \cite{)
• Quotely searches your corpus and suggests the most relevant paper
• Accept → the \cite{Key} and full BibTeX entry are inserted automatically

All processing is local: your documents never leave your machine.

Features

• Inline citation suggestions — triggered by \cite{ or [@ (Pandoc), or manually with Cmd+Shift+C
• Semantic search — finds conceptually related papers even without exact keyword match
• Auto-bibliography — accepted citations are appended to a BibTeX block at the end of your file
• Auto-indexing — drop any file into data/ and it is parsed and indexed in the background
• File decorations — green ✓ badge on indexed files in the VS Code Explorer
• Click to open — every suggestion shows a link to open the source document
• Folder indexing — Cmd+Shift+P → "Quotely: Indexer un dossier" to add any folder to the index
• Supports 14 formats — PDF, LaTeX, Word, Markdown, PowerPoint, Excel, CSV, images (OCR), notebooks

Quick Start

Option A — VS Code Marketplace (recommended)

1. Search "Quotely" in the VS Code Extensions panel and click Install
2. On first activation, Quotely will offer to install the backend automatically (~300 MB, one-time)
3. Drop your PDFs/articles into ~/.quotely/data/papers/ and start writing

Option B — Manual setup

git clone https://github.com/ThomasbSAX/quotely.git
cd quotely

# Install backend + models (macOS/Linux)
bash setup.sh

Then install the extension from extension/quotely-1.2.0.vsix.

Requirements: Python 3.8+, macOS or Linux (Windows support coming soon).

Usage Guide

1. Add your documents

Drop PDF, DOCX, TEX, Markdown, or any other supported file into the data/ folder. Sub-folders are supported — organize however you like:

data/
├── papers/
│   ├── smith2023convexity.pdf
│   └── jones2022learning.tex
└── mybiblio/
    ├── chapter1/
    └── chapter2/

Files are indexed automatically within seconds of being added. No restart needed.

2. Get citation suggestions

Place your cursor after \cite{ in a LaTeX file (or [@ in Markdown) and press Cmd+Shift+C:

The gradient descent algorithm converges under smoothness conditions \cite{
%                                                                        ↑ Cmd+Shift+C here

Quotely analyzes the last 3 paragraphs and shows a ranked list of relevant papers. Select one → the key is inserted and the BibTeX entry is appended to your bibliography section.

You can also trigger suggestions at the end of any line (≥ 40 characters) — Quotely will insert a full \cite{Key}.

3. Search across your corpus

Cmd+Shift+F opens a search bar with two modes:

4. Commands

Performance

Benchmarks on Apple M-series CPU, corpus of 300 articles (7975 chunks), models warm.

Latency per query

Cold start (first query after VS Code launch): ~15–20 s while models load into RAM.

Indexing speed

Corpus scaling

Resource Usage

RAM (backend running)

Runs in the background — VS Code itself uses ~500 MB additionally. On machines with 8 GB RAM or more, this is comfortable.

Disk

Total first install: ~1 GB disk space.

CPU

• Indexing: uses ~100% of one core for a few seconds per document (embedding batch)
• Suggestions: ~200 ms peak, then idle
• Idle: < 0.5% CPU (file watcher + FastAPI waiting)

Architecture

User query (last 3 paragraphs of their document)
        │
        ▼
  Query cleaning               ← strip LaTeX commands, \cite{}, \begin{}…
        │
        ▼
  Bi-encoder                   ← all-MiniLM-L6-v2 (384-dim embedding)
        │
        ▼
  ChromaDB HNSW search         ← cosine similarity, n×10 candidate chunks
        │
        ▼
  Deduplication + context      ← top-2 chunks per paper concatenated
        │
        ▼
  Cross-encoder reranker       ← mmarco-mMiniLMv2-L12-H384-v1 (multilingual)
        │
        ▼
  Top-n results                ← sigmoid-normalized scores [0, 1]

Retrieval improvements

• Title-prefix embeddings: every chunk is embedded as "Titre: {title}\n{text}" so the title is always part of the vector representation — retrieval works even if the query only mentions the title
• Dedicated title chunk: a title-only entry is stored per paper for direct title matching
• Author inference: extracted from folder structure (e.g. DocumentsHarbulot/ → author = Harbulot) and from document text patterns ("Par Prénom Nom")

Models Used

Bi-encoder — sentence-transformers/all-MiniLM-L6-v2

Cross-encoder — cross-encoder/mmarco-mMiniLMv2-L12-H384-v1

Vector database — ChromaDB 0.5.3

• Embedded, no server required
• HNSW index with cosine metric
• Persisted locally in data/db/

Supported Formats

Settings

{
  "quotely.backendUrl": "http://127.0.0.1:7331",  // backend port
  "quotely.contextParagraphs": 3,                  // paragraphs used as query context
  "quotely.maxSuggestions": 5,                     // number of candidates shown
  "quotely.triggerOnCite": true,                   // auto-trigger on \cite{
  "quotely.projectPath": "/path/to/quotely"        // set by setup.sh or auto-setup
}

License

MIT — free to use, modify, and distribute.