readme.md

metadata

title: ConjunctionReservoir Document Chat
emoji: 🧠
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
short_description: Chat with docs via sentence-level retrieval
tags:
  - rag
  - retrieval
  - nlp
  - neuroscience
  - document-qa

ConjunctionReservoir Document Chat

Github is at: https://github.com/anttiluode/conjunctionreservoir

Upload any .txt or .pdf document and chat with it.

What makes this different from standard RAG:

Instead of asking "do query terms appear somewhere in this chunk?", ConjunctionReservoir asks "do query terms appear in the same sentence?"

This is grounded in auditory neuroscience:

• Norman-Haignere et al. (2025): auditory cortex integration windows are time-yoked (~80ms fixed clocks)
• NMDA receptor logic: hard AND gate — both inputs must arrive simultaneously
• Vollan et al. (2025): coverage-maximizing theta sweep for exploration

Benchmark: 100% Rank-1 rate on conjunction queries vs 60% for BM25 and SweepBrain.

Usage

1. Upload a .txt or .pdf, or paste text directly
2. Ask questions — works best for queries requiring two concepts together
3. Adjust the conjunction threshold slider to tune precision vs recall
4. Use :coverage, :summary, :threshold N commands in chat

No dependencies beyond NumPy for retrieval. Generation via HuggingFace Inference API (free).