@aerograph/adapter-langchain
This adapter bridges LangChain.js workflows into the AeroGraph.
Deterministic Mapping
LangChain's complex callback hierarchy is deterministically mapped to the minimal Flight Recorder contracts as follows:
| LangChain Callback | AFR Event Kind | Metadata Mapped |
|---|---|---|
handleLLMStart |
prompt |
messages/prompts, model.name, parentSpanId, runId -> spanId |
handleLLMEnd |
response |
generations, streaming metrics, usage tokens, durationMs, runId -> spanId |
handleLLMError |
error |
error message, durationMs, runId -> spanId |
handleToolStart |
tool_call |
input string/JSON, runId -> spanId |
handleToolEnd |
tool_result |
output, durationMs, runId -> spanId |
handleToolError |
error |
error message, durationMs, runId -> spanId |
handleChainStart / handleChainEnd |
note |
Detects LangGraph node boundaries (kind: "langgraph_node"), capturing state_before (Inputs) and state_update (Outputs) |
Installation
npm install @aerograph/adapter-langchain @aerograph/sdk(Requires Node.js >= 18.18.0)
Quick Start
The adapter provides a callback handler that you inject into your LangChain invocations.
import { FlightRecorder } from "@aerograph/sdk";
import { AeroGraphCallbackHandler } from "@aerograph/adapter-langchain";
import { ChatOpenAI } from "@langchain/openai";
const recorder = new FlightRecorder({
endpoint: "http://localhost:4317",
actor: { id: "my-langchain-agent" }
});
const handler = new AeroGraphCallbackHandler(recorder);
const model = new ChatOpenAI({ modelName: "gpt-4" });
// The handler automatically maps LangChain callbacks to AeroGraph events
await model.invoke("Hello, how are you?", {
callbacks: [handler]
});Supported Features
- LLM Call Tracking (Prompts & Responses)
- Streaming Telemetry (TTFT, tokens/sec)
- RAG Retriever Document Payloads
- Tool Calls & Results
- LangGraph State Snapshots
License
Apache-2.0
| handleAgentAction| (ignored) | Caught by tool/llm events |
For Phase 1 MVP, we focus strictly on LLMs and Tools plus lightweight chain boundary notes to keep the graph comprehensible.