Skip to content

integrations.md

Latest commit

 

History

History
77 lines (56 loc) · 2.73 KB

integrations.md

File metadata and controls

77 lines (56 loc) · 2.73 KB
Raw

Integrations

The core SDK has only httpx and pydantic runtime dependencies. That is the recommended integration path: call explicit product methods from your agent framework and route on TRACE's decision.

from latence import Latence

trace = Latence()
decision = trace.grounding.rag(
    query=user_question,
    response_text=agent_answer,
    raw_context=retrieved_context,
)

This direct shape works cleanly in LangGraph nodes, LangChain callbacks, LlamaIndex postprocessors, n8n HTTP nodes, Cursor tools, Claude Code hooks, Codex workflows, and custom pipelines.

Integration Principle

Integrations are inheritance layers over the SDK, not separate products.

Every adapter must extract framework-native inputs, call one of the shared SDK product paths, and attach TRACE metadata back to the framework object. Adapters must not implement scoring, thresholds, retries, redaction semantics, memory rules, or policy decisions.

Shared adapter behavior lives in latence.integrations._trace so framework wrappers automatically inherit new runtime fields, SDK fixes, and response metadata as the core product evolves.

Optional Helpers

Optional adapter modules live under latence.integrations:

Framework Module Install
LangChain latence.integrations.langchain pip install "latence[langchain]"
LangGraph latence.integrations.langgraph pip install "latence[langgraph]"
LlamaIndex latence.integrations.llama_index pip install "latence[llama_index]"
OpenAI latence.integrations.openai pip install "latence[openai]"
Haystack latence.integrations.haystack pip install "latence[haystack]"
CrewAI latence.integrations.crewai pip install "latence[crewai]"
AutoGen latence.integrations.autogen pip install "latence[autogen]"
Pydantic AI latence.integrations.pydantic_ai pip install "latence[pydantic_ai]"

Install everything only for integration development:

pip install "latence[all]"

Integration Contract

Every adapter should do the same four things:

  1. Extract the user query or task.
  2. Extract the candidate model response.
  3. Extract retrieved context, code context, or tool evidence.
  4. Call client.grounding.rag(...) through the shared integration helper.

Keep TRACE decisions explicit in your app state. That makes routing, logs, reviews, retries, and replay tests straightforward.

Phase 5 Demos

The free checkpoint demos live in examples/phase5/:

  • LibreChat/OpenRouter proxy prototype.
  • Native SDK, LangChain, and LlamaIndex RAG comparison.
  • LangGraph coding-agent review/retry/pass route.
  • Importable n8n HTTP-node workflows.

See Phase 5 Integration Checkpoint.