Cognoxium

Pandas for trustworthy AI context.

Cognoxium is a local-first Rust and Python library that turns messages, retrieved documents, and tool results into typed, provenance-aware context. It selects a deterministic context pack under a token budget and records every exclusion, merge, truncation, and policy decision.

Cognoxium sits between application inputs and a model SDK:

retriever / tools / messages
            |
            v
      CognitionFrame
  filter -> dedupe -> rank
            |
            v
  policy + budget -> ContextPack + PackManifest
            |
            v
    provider renderer -> model SDK

It does not retrieve documents, call a model, authorize tool actions, transmit data, or require an account.

Where should I start?

• New to Cognoxium: follow 10 minutes to Cognoxium.

• Coming from pandas: read Cognoxium for pandas users.

• Integrating external data: read Building a CognitionFrame, then Trust boundaries and policies.

• Handling a model limit: read Token profiles and budget packing and Overflow recovery.

• Evaluating security claims: read Security model and limitations and Status and non-binding roadmap.

• Looking up a symbol: use Python API reference.

Learn

• Installation
  • Recommended: uv 0.12.1
  • pip
  • Optional integrations
  • Supported wheels
  • Development checkout
• 10 minutes to Cognoxium
  • Build, transform, and pack context
  • Understand the decision
  • Render only at the SDK boundary
  • Keep the audit record
• Cognoxium for pandas users
  • A practical mapping
  • Familiar operations
  • Deliberate differences from pandas
• Core concepts
  • Context items
  • Frame, pack, and manifest
  • Provenance, trust, and sensitivity
  • Retention
  • Lineage
  • Determinism
• Building a CognitionFrame
  • From Python records
  • Standard fields
  • Payload forms
  • Files and Arrow-compatible input
  • Source plugins
• Filtering, ranking, and deduplication
  • Filtering
  • Ranking
  • Exact deduplication
• Token profiles and budget packing
  • Processing order in the 0.1 series
  • Retention is a contract
  • Defaults
  • Token profiles
  • Inspect the result
  • Required overflow
• Overflow recovery
  • Inspect and recover
  • Recovery choices
  • Truncation constraints
  • Localization
• Trust boundaries and policies
  • Core policy matrix
  • Untrusted is a label, not a block
  • Required violations
  • Custom policies
• Lineage and manifests
  • What a manifest records
  • Transformation lineage
  • Storage guidance
• OpenAI, Anthropic, and text rendering
  • Role mapping in the 0.1 series
  • Inspect the generated shapes
  • Payload mapping
  • Custom renderers
• Security model and limitations
  • What the core enforces
  • What Cognoxium does not guarantee
  • Secret-pattern scope
  • Trust assignments
  • Structured diagnostics
  • Audit storage
• Writing plugins
  • Ranker
  • Policy
  • Source
  • Renderer
  • Python tokenizer fallback
  • Reproducibility contract
• Performance
  • What runs where in the 0.1 series
  • Token cache
  • Benchmark policy
• Recipes
  • Select one retrieved document under a budget
  • Quarantine failed tool results
  • Demote old conversation turns explicitly
  • Persist an auditable frame
  • Recover from budget overflow
• Troubleshooting
  • BudgetExceeded
  • PolicyViolation
  • UnsupportedContent
  • SchemaError
  • PluginError
  • Approximate token counts
  • Missing Arrow, pandas, or Polars
  • Unexpectedly selected untrusted content
  • Different manifests across runs
• Glossary
• Status and non-binding roadmap
  • Implemented in the 0.1 series
  • Current limitations
  • Non-binding directions

Reference and development

• Python API reference
  • Where to look
  • Rust API
• Rust API reference
• Migration guide
  • Before 1.0
  • Diagnostic compatibility
  • 1.0 policy
• Changelog
• Contributing
  • Set up a checkout
  • Run checks
  • Documentation changes
• Architecture
  • Current execution boundary
  • Dependency boundary
  • Extension direction

Stability

Cognoxium 0.1.1 is alpha software. Public diagnostic codes are stable within a minor release, while serialized schemas may change before 1.0 with a documented migration path. The Status and non-binding roadmap distinguishes implemented behavior from non-binding future directions.