Frequently asked questions

Comfortable minimum: 16 GB of RAM, a recent CPU (4+ cores), 10 GB of disk space for the index.

Recommended: 32 GB of RAM and an NVIDIA GPU (8 GB VRAM+) for smooth local LLMs (Mistral 7B, Llama 3 8B). Without a GPU, queries still work but are slower (10-30 s per answer).

On Apple Silicon Macs (M1/M2/M3), Metal is used automatically — no NVIDIA GPU needed.

No software limit. In practice, libraries of 5,000 to 20,000 articles are handled with ease. Indexing a typical PDF takes 2 to 8 seconds depending on mode (with or without OCR) and hardware.

For larger volumes (50,000+), cloud mode or a dedicated on-premise deployment is advised.

Yes. RefChat simply points to your Zotero (storage/) or Mendeley storage folder. Metadata is enriched automatically via GROBID and OpenAlex. No extension to install on the Zotero side.

The embedding model (multilingual E5-large) covers around a hundred languages. Best handled: English, French, German, Spanish, Italian, Portuguese, Dutch. You can ask a question in English on a mixed EN/FR corpus — RefChat retrieves the relevant passages in both languages.

Local mode: no data leaves your machine. You can even work offline after installation.

Cloud mode: indexing and LLM on our OVHcloud servers (Roubaix, France). Data encrypted in transit (TLS 1.3), deletion on request, no model training on your content. Documented GDPR compliance.

Yes. On request, we offer a live demo (30 min) and a trial on a sample of your own library, under a confidentiality agreement if needed. Request a demo.

RefChat is specialised for scientific libraries (GROBID parsing, EasyOCR, hybrid dense+BM25+reranking search, automatic query-mode detection). Above all: a 100% local option — your articles, patents or sensitive reports never leave your machine.

General-purpose cloud tools send your documents to servers in the United States and don't offer the same level of source citation or thematic organisation.

Detailed comparison: RefChat vs NotebookLM.