Search your papers by meaning, not just words

What you can do

• Every passage is embedded and stored in a local Postgres database with pgvector.
• The built-in model runs locally on ONNX Runtime and downloads once.
• Reading, notes and chat on short papers work before it is downloaded.
• The graph’s similar-content layer and the References ranking use the same index.
• Choose another source in Settings → Search; switching re-embeds in the background (next release).
• A cloud source asks before the first send, with a cost estimate (next release).

How it works

1. Embed

   Each passage becomes 768 numbers describing its meaning.

2. Store

   The vectors live next to your papers in a local database.

3. Match

   A question is embedded the same way, and the closest passages are returned.

Tips

• Ask in full sentences: meaning-based search works better with context than with keywords.
• Keep the built-in model if privacy matters most: nothing leaves your machine.

Works well with

• Ask a paperChat with a paper; every answer cites the passage it used.
• Library graphCitations, authors, topics, notes and similar content, in five views.
• Any modelOllama, Anthropic or any OpenAI-compatible server, picked per question.

Guides

• How to chat with a PDF locally, without uploading it — Ask papers questions with a local model, and check every answer against the page.
• Can you trust AI with research papers? — Invented citations, and a five-step check for any AI answer.
• Running a local LLM for research — Hardware, Ollama, choosing a model, and making it answer well.

Terms explained

• Retrieval-augmented generation (RAG)
• Embeddings
• Semantic search
• Vector database