NeuronCite

<p align="center"> <img src="crates/neuroncite-web/assets/splash.png" alt="NeuronCite" width="400" /> </p> <p align="center"> <code>Rust / Single Binary / Local-First / MCP-Native / Ollama-Powered</code> </p> <h3 align="center">Your Documents. Your Machine. Your Answers.</h3>

---

NeuronCite is an enterprise-grade semantic search engine that transforms your document library into an instantly searchable knowledge base -- and autonomously verifies every citation in your academic papers using local LLMs. No cloud. No API keys. Your data stays on your machine.

Written in Rust, it ships as a single binary (CPU) or minimal bundle (GPU) for Windows, macOS, and Linux. After a one-time download of models and runtime dependencies, the system operates fully offline, including in air-gapped and classified environments. A built-in MCP server with 43 tools lets AI assistants like Claude control the full workflow -- indexing, searching, verifying citations, annotating PDFs -- directly from the chat interface.

---

Why NeuronCite

Privacy by Design -- All document processing runs locally. No user data leaves the machine, no telemetry, no cloud accounts. No internet connection required after initial setup (embedding models, pdfium, and optional OCR/Ollama binaries). Network-dependent features (DOI resolution, HTML crawling, citation source fetching) are unavailable in air-gapped mode. Supports air-gapped and classified environments once all dependencies are provisioned.

Autonomous Citation Verification -- Feed a LaTeX paper and its bibliography. NeuronCite indexes the cited PDFs, runs a local LLM via Ollama, and verifies every \cite{} command against actual source material. Each citation receives a verdict, confidence score, and correction suggestions.

Enterprise-Grade Architecture -- 16 Rust crates with clear separation of concerns. 43 MCP tools, 34 REST API endpoints, a browser-based GUI with 7 tabs, a Python client library, and 11 CLI commands. CPU-only builds compile into a single executable that runs without Docker, Kubernetes, or external infrastructure. Citation verification additionally requires a running Ollama instance.

---

Quick Start

1. Download the binary for your platform from the Releases page.

2. Double-click the binary. The application opens in a native window (WebView2 on Windows, WebKit on macOS/Linux) -- no terminal, no configuration required. Linux GUI builds require libwebkit2gtk-4.1-0 at runtime (see Installation for details).

3. Select a directory of PDFs in the Indexing tab, choose an embedding model, and click Start. The first run downloads the selected model (50 MB--1 GB depending on model size).

Terminal alternative: Run neuroncite from the command line, or index directly with neuroncite index --directory ./papers.

Docker alternative:

docker run --gpus all -p 3030:3030 \
  -v neuroncite-data:/data/Documents/NeuronCite \
  ghcr.io/ff-tec/neuroncite:latest

---

Features

Hybrid Search

Combines three retrieval algorithms for high-precision document search:

• HNSW vector similarity -- Approximate nearest neighbor search over dense embeddings for semantic matching
• BM25 keyword matching -- SQLite FTS5 full-text index for exact term matching
• Reciprocal Rank Fusion -- Merges and normalizes scores from both retrieval methods
• Cross-encoder reranking -- Optional second-stage scoring with a cross-encoder model for precision-critical queries
• Sub-chunk refinement -- Configurable divisors for finer-grained passage retrieval within chunks

Document Indexing

• PDF extraction with three backends: pdf-extract (pure Rust, default), pdfium (multi-column layout support), and Tesseract OCR (image-heavy pages)
• HTML crawling with BFS traversal, depth limiting, rate limiting, domain filtering, regex URL patterns, and sitemap parsing
• Four chunking strategies: page-based, word-window (fixed word count with overlap), token-window (subword tokens), and sentence-based (respects citation boundaries)
• Eight embedding models from 33M to 335M parameters (384 to 4096 dimensions), downloaded from HuggingFace on first use
• GPU acceleration via ONNX Runtime with CUDA 12.4, DirectML, CoreML, and ROCm execution providers; CPU fallback on all platforms

Citation Verification

NeuronCite parses LaTeX papers for citation commands (\cite, \citep, \citet, \autocite, and variants), resolves them against BibTeX entries, and verifies each claim against indexed source PDFs using a local LLM via Ollama.

Five-stage verification pipeline:

1. Parse LaTeX, extract citation commands, resolve cite-keys against BibTeX
2. Match bibliographic entries to PDFs via Jaro-Winkler similarity (threshold 0.80)
3. Group citations into batches for parallel processing, preserving co-citations and section context
4. Verify each batch through the LLM with a minimum of 2 search queries per citation, including cross-corpus search for alternative sources
5. Export 6 output files: annotation CSV, citation CSV/XLSX, corrections JSON, citation report JSON, full detail JSON, and annotated PDFs

Seven verdict types:

| Verdict | Meaning | |---------|---------| | supported | Claim explicitly found in cited source with page numbers and passages | | partial | Source supports part of the claim; other parts absent or unsupported | | unsupported | PDF found and read but no passage supports the claim | | not_found | Source PDF not in indexed corpus | | wrong_source | Claim verifiable but evidence found in a different source | | unverifiable | Future projections, subjective statements, or insufficiently specific claims | | peripheral_match | Text found only in non-substantive sections (TOC, bibliography, appendix) |

Each verdict includes a confidence score (0.00--1.00), critical/warning flags for contradictions and temporal mismatches, and correction suggestions with types (rephrase, add context, replace citation).

PDF Annotation

Highlights text passages in PDFs with color-coded annotations based on verification verdicts. Uses a 5-stage text matching pipeline:

1. Exact byte-level match
2. Normalized match with whitespace collapsing
3. Fuzzy character-level match (string distance)
4. Fallback extraction via multi-backend PDF text pipeline
5. OCR fallback for scanned pages

Accepts annotation input as CSV or JSON. Supports comment annotations with popup text alongside highlights.

Web Frontend (7 Tabs)

The SolidJS single-page application is compiled into the binary via rust-embed and served at http://localhost:3030.

| Tab | Function | |-----|----------| | Sources | BibTeX management, web crawling (BFS with regex patterns, sitemap parsing), DOI resolution (Unpaywall, Semantic Scholar, OpenAlex), metadata extraction | | Indexing | Directory selection, embedding model and chunking strategy configuration, real-time progress tracking, session management | | Search | Multi-session hybrid search with vector/BM25/reranking toggles, sub-chunk refinement, grouped and flat result views, export as Markdown/BibTeX/CSL-JSON/RIS | | Citations | LaTeX file selection with auto-detection of .bib files, Ollama model selection with connection test, verification mode presets (quick, balanced, thorough), live results with expandable verdicts | | Annotations | CSV/JSON annotation input, source PDF directory selection, 5-stage text location pipeline, color configuration, per-quote progress tracking | | Models | Embedding model catalog with download/activate controls, cross-encoder reranker management, Ollama LLM catalog, GPU/CUDA system info, model diagnostics | | Settings | FTS5 index optimization, HNSW vector index rebuild, database reset, dependency detection (pdfium, Tesseract, Poppler), MCP server registration, real-time log streaming via SSE |

MCP Server (43 Tools)

The Model Context Protocol server exposes 43 tools for AI agent integration via JSON-RPC 2.0 over stdio, organized in 8 categories:

| Category | Tools | Purpose | |----------|-------|---------| | Session Management | 5 | List, delete, update, diff, discover index sessions | | Indexing | 4 | Index directories, add files, reindex, preview chunks | | Search & Retrieval | 8 | Search, batch search, multi-session search, compare, text search, content retrieval, export | | File & Chunk Inspection | 4 | List files, inspect chunks, compare files, quality reports | | Citation Verification | 8 | Create jobs, add claims, submit, check status, fetch rows, export, retry, fetch sources | | Annotation | 4 | Annotate PDFs, check status, inspect annotations, remove annotations | | Web Sources | 3 | Fetch HTML, crawl websites, check crawl status | | System | 3 | List models, health check, log streaming |

REST API (34 Endpoints)

Axum-based HTTP server with OpenAPI specification (utoipa), bearer token authentication with constant-time comparison, per-IP rate limiting, CORS, and Server-Sent Events for real-time progress streaming. Endpoints cover all functionality: sessions, indexing, search, citation, annotation, models, and system health.

Python Client

Typed access to all REST endpoints with subprocess server management.

from neuroncite import NeuronCiteClient

client = NeuronCiteClient()  # default: http://127.0.0.1:3030
results = client.search(session_id="...", query="capital asset pricing model")
for r in results:
    print(f"  [{r.score:.2f}] {r.citation}")

30+ typed methods covering search, indexing, citation, annotation, and model management. See clients/python/README.md for the full API reference.

---

Installation

Pre-built Binaries

Download the binary for your platform from the Releases page. Each release includes SHA-256 checksums for verification.

| Platform | Architecture | GUI | Artifact | |----------|-------------|-----|----------| | Windows | x86_64 | Native window (WebView2) | neuroncite-windows-x64.exe | | Linux | x86_64 | Native window (WebKit2GTK) | neuroncite-linux-x64 | | Linux | x86_64 | Browser-only (headless) | neuroncite-linux-x64-server | | Linux | ARM64 | Native window (WebKit2GTK) | neuroncite-linux-arm64 | | Linux | ARM64 | Browser-only (headless) | neuroncite-linux-arm64-server | | macOS | ARM64 (Apple Silicon) | Native window (WebKit) | neuroncite-macos-arm64 | | macOS | x86_64 (Intel) | Native window (WebKit) | neuroncite-macos-x64 |

Linux GUI variants require libwebkit2gtk-4.1-0 at runtime. Server variants have zero runtime dependencies beyond glibc.

Docker

Four image variants are published to the GitHub Container Registry. See docker/README.md for compose profiles, environment variables, and local build instructions.

# NVIDIA GPU (default)
docker run --gpus all -p 3030:3030 \
  -v neuroncite-data:/data/Documents/NeuronCite \
  ghcr.io/ff-tec/neuroncite:latest

# AMD ROCm
docker run --device=/dev/kfd --device=/dev/dri -p 3030:3030 \
  -v neuroncite-data:/data/Documents/NeuronCite \
  ghcr.io/ff-tec/neuroncite:latest-rocm

# CPU only
docker run -p 3030:3030 \
  -v neuroncite-data:/data/Documents/NeuronCite \
  ghcr.io/ff-tec/neuroncite:latest-cpu

| Variant | Base Image | GPU | Tag | |---------|-----------|-----|-----| | NVIDIA | CUDA 12.4 + cuDNN, Ubuntu 22.04 | CUDA | latest / <version>-nvidia | | ROCm | ROCm 6.4, Ubuntu 22.04 | ROCm | latest-rocm / `<ver