Python Runtime
for AI-Protocol.
Developer-friendly, protocol-driven AI client. Full async support, Pydantic v2 type safety, production-grade telemetry, and smart model routing. Python 3.10+.
pip install ai-lib-python Key Features
Async-First Design
Native async/await throughout. httpx-powered transport with connection pooling. CancellableStream for graceful stream termination.
Pydantic v2 Types
Full type hints with Pydantic v2 validation. ProtocolManifest, Message, ContentBlock, StreamingEvent — all type-safe with runtime validation.
Resilient Executor
ResilientExecutor combines backpressure, rate limiting, circuit breaker, and retry with exponential backoff. PreflightChecker gates requests before execution.
Model Routing
ModelManager with ModelArray for intelligent model selection. Strategies include round-robin, weighted, cost-based, and quality-based routing.
Production Telemetry
MetricsCollector with Prometheus export. Distributed tracing via OpenTelemetry. Structured logging. Health monitoring. User feedback collection.
Structured Output
JSON mode configuration, schema generation from Pydantic models, output validation. Guardrails with content filters and PII detection.
Pythonic & Intuitive
A clean, Pythonic API that feels natural. Create a client from a model identifier, build requests with a fluent builder, and iterate over unified streaming events.
The same code works across all providers. Switch from OpenAI to Anthropic to DeepSeek by changing one string — the protocol manifest handles the rest.
from ai_lib_python import AiClient
# Works with ANY provider
client = await AiClient.create(
"anthropic/claude-3-5-sonnet"
)
# Streaming chat
async for event in client.chat() \
.user("Explain AI-Protocol") \
.temperature(0.7) \
.max_tokens(1000) \
.stream():
if event.is_content_delta:
print(event.as_content_delta.text,
end="")
# Non-streaming with stats
response, stats = await client.chat() \
.user("Summarize this") \
.execute_with_stats()
print(f"Tokens: {stats.total_tokens}")
print(f"Latency: {stats.latency_ms}ms")Internal Architecture
Same layered architecture as the Rust runtime — protocol-driven, operator-based, resilient by default.
Module Overview
client/
AiClient, AiClientBuilder, ChatRequestBuilder, ChatResponse, CallStats, CancelToken, CancellableStream.
protocol/
ProtocolLoader (local/env/GitHub), Pydantic manifest models, JSON Schema validator, hot-reload support.
pipeline/
Decoder (SSE, JSON Lines, Anthropic SSE), Selector (JSONPath), Accumulator, FanOut, EventMapper (protocol-driven, default, Anthropic).
resilience/
ResilientExecutor, RetryPolicy, RateLimiter, CircuitBreaker, Backpressure, FallbackChain, SignalsSnapshot, PreflightChecker.
routing/
ModelManager, ModelArray, selection strategies (round-robin, weighted, cost-based, quality-based). Pre-configured catalogs for major providers.
telemetry/
MetricsCollector (Prometheus), Tracer (OpenTelemetry), structured Logger, HealthChecker, user feedback collection.
embeddings/ + tokens/
EmbeddingClient with vector operations. TokenCounter (tiktoken integration) and cost estimation for usage tracking.
plugins/ + structured/
Plugin base class, PluginRegistry, HookManager, middleware chain. JSON mode config, SchemaGenerator, OutputValidator.