Client¶

The Client class (alias for CRPOrchestrator) is the main entry point for the CRP protocol.

Constructor¶

from crp import Client

client = Client(
    provider="openai",      # Provider name or LLMProvider instance
    config=None,            # Optional dict of configuration overrides
    llm=None,               # Optional pre-configured LLM provider instance
    model="gpt-4o",         # Model name (passed to provider)
    **init_kwargs,          # Additional provider-specific kwargs
)

Parameters¶

Parameter	Type	Default	Description
`provider`	`str \\| LLMProvider`	Required	Provider identifier or instance
`config`	`dict \\| None`	`None`	Configuration overrides
`llm`	`LLMProvider \\| None`	`None`	Pre-configured provider (overrides `provider`)
`model`	`str \\| None`	`None`	Model name for the provider
`**init_kwargs`	`Any`	—	Additional provider-specific arguments

Provider Strings¶

String	Provider Class
`"openai"`	`OpenAIAdapter`
`"anthropic"`	`AnthropicAdapter`
`"ollama"`	`OllamaAdapter`
`"llamacpp"`	`LlamaCppAdapter`

Configuration Options¶

config = {
    "max_continuations": 10,          # Max continuation windows
    "extraction_stages": [1,2,3,4,5], # Which stages to run (1-6)
    "ckf_enabled": True,              # Enable cross-session knowledge
    "ckf_path": "./my_kb",            # CKF storage path
    "ckf_max_size_mb": 500,           # CKF size limit
}

Core Methods¶

ingest()¶

Ingest raw text into the session's warm state:

result = client.ingest(
    raw_text="Kubernetes is an open-source container orchestration platform...",
    source_label="k8s-docs",
)

Returns: ExtractionResult

Field	Type	Description
`facts_extracted`	`int`	Number of facts extracted
`source_label`	`str`	Source identifier
`fact_ids`	`list[str]`	IDs of extracted facts

session_status()¶

Get current session status:

status = client.session_status()
print(f"Session: {status.session_id}")
print(f"Windows: {status.windows_completed}")
print(f"Facts:   {status.facts_in_warm_state}")
print(f"Cost:    ${status.total_cost:.4f}")

Returns: SessionStatus

Field	Type	Description
`session_id`	`str`	Session identifier
`windows_completed`	`int`	Total windows dispatched
`total_input_tokens`	`int`	Cumulative input tokens
`total_output_tokens`	`int`	Cumulative output tokens
`facts_in_warm_state`	`int`	Facts in warm store
`overhead_ratio`	`float`	CRP overhead as fraction of total time
`remaining_budget`	`int \\| None`	Remaining token budget
`total_cost`	`float`	Estimated cost in USD

preview_envelope()¶

Preview envelope construction without dispatching:

preview = client.preview_envelope(
    system_prompt="You are a technical writer.",
    task_input="Explain Kubernetes networking.",
)
print(f"Saturation: {preview.saturation:.1%}")
print(f"Facts:      {preview.facts_included}/{preview.facts_available}")

Returns: EnvelopePreview

Field	Type	Description
`total_tokens`	`int`	Total context window use
`envelope_tokens`	`int`	Tokens used by envelope
`generation_reserve`	`int`	Tokens reserved for generation
`facts_included`	`int`	Facts packed into envelope
`facts_available`	`int`	Total facts in warm state
`saturation`	`float`	Envelope fill ratio (0.0–1.0+)

Compliance Properties¶

All compliance components are accessible as public properties:

Property	Type	Purpose
`client.risk_classifier`	`RiskClassifier`	EU AI Act risk assessment
`client.human_oversight`	`HumanOversightController`	Human oversight levels
`client.compliance_audit`	`ComplianceAuditTrail`	HMAC audit trail
`client.pii_scanner`	`PIIScanner`	PII detection
`client.consent_manager`	`ConsentManager`	GDPR consent tracking
`client.processing_records`	`ProcessingRecordKeeper`	GDPR Art. 30 records
`client.retention_manager`	`RetentionManager`	Data retention
`client.compliance_reporter`	`ComplianceReporter`	Report generation
`client.lineage_tracker`	`DataLineageTracker`	Data lineage

See Compliance API for detailed documentation.

Lifecycle¶

from crp import Client

# Create
client = Client(provider="openai", model="gpt-4o")

# Use
client.ingest(text, "source")
output, report = client.dispatch(system_prompt, task_input)
status = client.session_status()

# Close
client.close()

Session cost

The first dispatch() or ingest() call triggers model loading (GLiNER, embeddings). This takes ~10–15 seconds. Subsequent calls are fast.