metergraphrelay 0.1.1

Move LLM trace data between systems: pull from providers, push to metergraph.

metergraphrelay

Try metergraph without installing its SDK: export the chat completions OpenAI already stores for you, and push them into metergraph.

1. Your app calls OpenAI with store=True — a one-line addition if it doesn't already (see below). OpenAI keeps the completion server-side.
2. metergraphrelay pull openai lists those stored completions via OpenAI's own API and writes them as JSONL, already shaped to metergraph's native trace schema.
3. metergraphrelay push uploads that file to metergraph.

No SDK, no instrumentation beyond the store=True flag — which is also what OpenAI's own dashboard and evals features use.

This only reads what's associated with the API key you provide; it can't see or export anyone else's data. Built on OpenAI's List Chat Completions API — see the API reference for the underlying data model.

Setup

pip install metergraphrelay
cp .env.example .env
# edit .env and set OPENAI_API_KEY

Quickstart

metergraphrelay pull openai -n 25 --output traces.jsonl
metergraphrelay push traces.jsonl

No stored completions yet? Generate a couple first:

metergraphrelay demo openai

Enabling storage on your own calls

pull openai only finds completions created with store=True. Add it to calls you're already making:

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "..."}],
    store=True,
    metadata={"source": "my-app"},  # optional, filterable later
)

Before enabling this in production:

• Stored completions include full request/response content by default.
• There's no automatic expiry from OpenAI's side — delete via their Delete chat completion API; metergraphrelay doesn't do this for you.

Commands

metergraphrelay pull openai -n 25 --output my-traces.jsonl --stdout --include-content --route my-app/support-bot
metergraphrelay demo openai --model gpt-4o-mini
metergraphrelay push traces.jsonl

pull anthropic / pull langfuse accept the same shape but aren't implemented yet — they check for ANTHROPIC_API_KEY / LANGFUSE_PUBLIC_KEY+LANGFUSE_SECRET_KEY and report accordingly.

All subcommands accept --env-file PATH.

Trace record shape

Each line of pull openai's output is a JSON object already shaped for metergraph's ingest API:

{
  "ts": "2026-07-30T12:00:00+00:00",
  "provider": "openai",
  "model": "gpt-4o-mini",
  "status": "success",
  "endpoint": "chat.completions",
  "input_tokens": 12,
  "output_tokens": 34,
  "error": false,
  "error_type": null,
  "request_id": "chatcmpl-...",
  "tags": {},
  "route": "openai/backfill",
  "content_opted_in": false,
  "request_json": null,
  "response_text": null,
  "sdk": "metergraphrelay",
  "sdk_version": "0.1.1"
}

request_json/response_text are populated only when --include-content is passed.

Every completion returned by the stored-completions list already succeeded, so status is always "success". error/error_type flag a partial record: --include-content was requested but the follow-up message fetch failed, so token counts are still real while the content is missing.

Development

git clone https://github.com/VasiliyRad/metergraphrelay
cd metergraphrelay
pip install -e ".[dev]"
pytest