Amni-Prism

Shared knowledge atlas built on PRISM-TEX (.ptex). Every fact gets a unique nonce — contribute knowledge, it deduplicates automatically by content hash, and the atlas grows. NDJSON manifests are append-only and git-merge-friendly, so multiple contributors never collide.

Status: alpha (0.1.x). API may shift before 1.0.

Install

pip install amni-prism

For local development:

git clone https://github.com/Amnibro/Amni-Prism
cd Amni-Prism
pip install -e .

Quick Start

import prism

prism.contribute_text('./codex', 'The speed of light is 299792458 m/s', domain='physics')

result = prism.query_text('./codex', 'The speed of light is 299792458 m/s')

hits = prism.search_keyword('./codex', 'speed of light')

CLI

prism contribute "The speed of light is 299792458 m/s" -d physics

prism search "speed of light"

prism propose "Water boils at 100C at sea level" -d physics --confidence 0.8

prism pending
prism verify

prism scrape --file article.txt --source "https://en.wikipedia.org/wiki/Light"

prism stats

How It Works

1. Nonces — Every word/line/block gets a unique nonce in finite-field space. Same content → same nonce → automatic dedup.
2. PRISM-TEX Storage — Knowledge stored as nonce-addressed maps. Vocab in NLX format, hierarchical tiers in HNA format.
3. Content-Hash Dedup — SHA-256 over normalized content ensures no duplicates across contributors merging different codexes.
4. Two-Tier Verification — Small models propose (~50 tokens), large models verify (~20 tokens). ~70 tokens per verified fact.
5. NDJSON Manifests — Append-only JSON-lines manifests are git-merge-friendly. No merge conflicts when two contributors add facts to the same domain.

Domains

25 built-in domains: code, math, science, history, language, music, art, philosophy, psychology, economics, politics, law, medicine, engineering, technology, education, sports, entertainment, food, travel, nature, religion, mythology, literature, general.

Why use this

• Knowledge that survives model swaps. The atlas is a pile of facts addressed by content, not a fine-tune of a specific model. Swap your LLM tomorrow, keep your atlas.
• Federation without coordination. Two contributors building independent codexes can prism merge their work and get a deduplicated union, with no central server.
• Built for the Adam runtime. Amni-Prism is the federated layer of Amni-Ai (Adam). The [federated] extra in Adam pulls this in directly.

Source Allowlist

Scraping is gated to license-compatible sources (Wikipedia, arXiv, RFC, Python docs, MDN, HuggingFace). See prism.scrape.ALLOWED_SOURCES.

Contributing

Fork, clone, pip install -e ., add facts via the CLI or Python API, and open a PR with your codex export. NDJSON manifests merge cleanly under git.

License

MIT — see LICENSE.