cognoxium 0.1.0

A trustworthy context DataFrame for AI applications, powered by Rust.

Cognoxium

日本語 · Documentation · Security

Cognoxium is a trustworthy context DataFrame for AI applications, powered by Rust.

AI applications ingest messages, retrieved documents, tool results, and external resources from trust domains that do not naturally fit into a single safe prompt. Cognoxium treats those inputs as typed data with provenance, sensitivity, retention, token cost, and lineage—then compiles them into a deterministic, auditable context pack.

Project status: alpha. The Python API is usable today and automatically selects the Rust extension when installed. Serialization compatibility is not guaranteed before 1.0.

Install

pip install cognoxium

Optional interoperability dependencies are separate:

pip install "cognoxium[arrow,pandas,polars,tokenizers]"

Quickstart

import cognoxium as cx

records = [
    {"id": "policy", "payload": "Never disclose credentials.", "role": "system",
     "sources": ["app://policy"], "trust": "trusted", "retention": "required"},
    {"id": "result", "payload": "The build passed.", "sources": ["tool://ci"]},
]
pack = (
    cx.CognitionFrame.from_records(records).dedupe().rank("build status")
    .pack(budget=512, token_profile=cx.profiles.approximate(),
          boundary=cx.Boundary.external("openai"))
)
print(pack.to_openai())

Why Cognoxium?

• DataFrame ergonomics: lazy filtering, projection, joins, grouping, ranking, and deterministic deduplication.
• Budget-aware packing: required, preferred, and optional retention contracts with explicit overflow behavior.
• Trust boundaries: provenance, quarantine, sensitivity labels, expiry, and secret-aware external packing.
• Auditable lineage: every exclusion, merge, truncation, and token decision is captured in a manifest.
• Provider-neutral output: render the same pack for OpenAI, Anthropic, or plain text.
• Local by default: no network calls, API keys, model calls, accounts, or telemetry in the core library.

Security model

Cognoxium enforces deterministic information-flow rules. It does not claim to solve prompt injection or determine whether arbitrary content is truthful. Injection detection is a risk signal; trust labels, provenance, least privilege, and application-side authorization remain necessary.

Development

python -m unittest discover -s tests -v
cargo test --workspace

Build the extension with maturin:

maturin develop

See CONTRIBUTING.md for the full workflow.

License

Licensed under the Apache License 2.0.