LlamaIndex

AxonPushLlamaIndexHandler is a set of plain instance methods you call at the boundaries you care about. It does not auto-register with LlamaIndex’s Settings or callbackManager, you wire each hook into the code that already orchestrates retrieval and generation.

Tested against llamaindex@^0.8.

Install

npm install @axonpush/sdk llamaindex

Wire up the hooks

import { AxonPush, AxonPushLlamaIndexHandler } from "@axonpush/sdk";

const client = new AxonPush();
const handler = new AxonPushLlamaIndexHandler({
  client,
  channelId: process.env.AXONPUSH_CHANNEL_ID!,
  agentId: "rag-pipeline",
});

handler.onQueryStart(question);

handler.onRetrieverStart(question);
const nodes = await retriever.retrieve(question);
handler.onRetrieverEnd(nodes.length);

handler.onLLMStart("gpt-4o", 1);
const response = await index.asQueryEngine().query(question);
handler.onLLMEnd(response);

handler.onQueryEnd(response);

For streaming, call handler.onLLMStream(token) per token.

agentId defaults to "llamaindex".

Method reference

Method	Identifier	axonpush eventType	Payload
onLLMStart(model?, promptCount?)	llm.start	agent.start	model, prompt_count
onLLMEnd(output?)	llm.end	agent.end	truncated output
onLLMStream(token)	llm.token	agent.llm.token	token
onEmbeddingStart(model?, textCount?)	embedding.start	agent.tool_call.start	model, text_count
onEmbeddingEnd(embeddingCount?)	embedding.end	agent.tool_call.end	embedding_count
onRetrieverStart(query)	retriever.query	agent.tool_call.start	query
onRetrieverEnd(nodeCount?)	retriever.result	agent.tool_call.end	node_count
onQueryStart(query)	query.start	agent.start	query
onQueryEnd(response?)	query.end	agent.end	truncated response

Every argument is optional except the query strings. Omitted models record as "unknown" and omitted counts as 0, so a partially-wired pipeline still produces a readable trace.

Queries are cut to 500 characters; outputs and responses are JSON-serialised and truncated at 500 characters, then run through the client’s redactor. Every event carries metadata.framework: "llamaindex", and all of them share one trace for the life of the handler, build a handler per request, or seed traceId, if you want one trace per query rather than one per process.