Instrument LlamaIndex

Recipe — instrument a LlamaIndex pipeline. SDK adapter + manual trace upload patterns.

Private preview. This surface is not yet generally available. Adapters live in the Python SDK today and the rendering / configuration UX is under iteration. Contact [email protected] to request access to the private preview.

When to use: LlamaIndex pipeline (RAG, query engines) and you want traces in Stratix.

Approach 1 — SDK adapter (when available)

The Stratix SDK ships LlamaIndex adapter coverage as an optional extras group:

pip install 'layerlens[frameworks-llamaindex]'

Wire the LayerLens callback into your CallbackManager. For canonical patterns (auto-instrumentation primitives), see the samples/instrument/ directory and the LangChain instrumented sample (same pattern).

Approach 2 — manual trace upload from your pipeline

Capture per-query input + output, upload as a trace file:

from layerlens import Stratix
client = Stratix()

# Run your LlamaIndex query
response = query_engine.query("What did the report say about Q3?")

# Build a trace record (any flexible Dict[str, Any] for `data`)
trace_record = {
 "input": "What did the report say about Q3?",
 "output": str(response),
 "spans": [
 {"name": "retrieval", "chunks": [n.node.text for n in response.source_nodes]},
 {"name": "synthesis", "text": str(response)},
 ],
}

# Write to JSONL and upload (≤50 MB per file)
import json
from pathlib import Path
Path("trace.jsonl").write_text(json.dumps(trace_record))
result = client.traces.upload("trace.jsonl")

# Score with a faithfulness judge
client.trace_evaluations.create(trace_id=result.trace_ids[0], judge_id=FAITHFULNESS_JUDGE_ID)

See also

• Workflow: Instrument
• Integrations
• SDK reference: traces
• SDK adapter samples