Import from the OpenAI dashboard

Do the shared setup first: an ingest key, pip install metergraphrelay, and a .env holding METERGRAPH_APP_TOKEN.

This path needs no observability tool at all. If your application calls OpenAI with store=True, OpenAI keeps the completion server side, and the relay reads it back. It is the same storage OpenAI’s own dashboard and evals features use.

What the relay reads

Stored chat completions, listed through OpenAI’s List Chat Completions API. Only completions created with store=True exist to be read, and an API key only ever sees the completions made with that key.

What you need

Credentials	The OPENAI_API_KEY your application uses
Prerequisite	store=True on the calls you want to import

.env

METERGRAPH_APP_TOKEN=<your ingest key>
OPENAI_API_KEY=sk-...

If your calls do not set store=True yet, add it:

app/llm.py

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "..."}],
    store=True,
)

Storage is a decision, not a detail

A stored completion holds the full request and response content on OpenAI’s side, and nothing expires it. You delete stored completions through OpenAI’s own API. Decide that before you turn store=True on in production, not after.

Backfill

1. Pull the stored completions

   metergraphrelay pull openai -n 100 --route my-app/support-bot --output openai.jsonl

   Flag	Meaning
   -n, --count	How many of the most recent stored completions to read. Defaults to 10
   --route	The route every row lands on. Defaults to openai/backfill
   --include-content	Copy the prompt and response text. Off by default
   --output	Where to write the JSONL. Defaults to ./traces.jsonl
   --stdout	Echo each row as it is written

   There are no time selectors here. The relay reads the most recent stored completions in descending order, so -n is the only bound.

   Add --include-content only if the prompt and response text may be copied. This is the one import path where content is opt in rather than automatic.

2. Read the file, then upload it

   head -n 2 openai.jsonl
   metergraphrelay push openai.jsonl

   metergraphrelay sync openai runs both steps in one command and takes the same flags. It is a convenience, not a scheduled mode.

3. Check the first rows

   Open Traces in Metergraph. Every row carries the route you passed, because OpenAI stores no workflow name for the relay to read. One pull is one route.

No scheduled mode

Every other source can run hourly from cron, holding a checkpoint on the Metergraph side. OpenAI cannot. The server’s list of sources that may take a sync lease is portkey, langfuse, braintrust, phoenix and langsmith (app/src/metergraph_app/import_identity.py:30), and OpenAI is deliberately not on it.

That means OpenAI rows carry no import identity and are not deduplicated. Two pulls over overlapping stored completions import the same call twice. Track what you have loaded, or use -n deliberately.

If you want ongoing capture of OpenAI traffic, instrument the application instead: see OpenAI.

What lands in Metergraph

Field	Where it comes from
Route	Whatever you pass with --route, or openai/backfill
Timestamp	The completion’s created time
Provider	Always openai
Model	The completion’s model
Endpoint	Always chat.completions
Input and output tokens	The completion’s prompt_tokens and completion_tokens
Status	Always success. A stored completion is one that succeeded
Request id	The completion’s id
Tags	The completion’s metadata object
Content	Only with --include-content. The request messages and the first choice’s reply

Cost is not carried. Metergraph prices these rows from its own catalog.

What does not carry over

• Cache, cache write and reasoning token buckets. The stored completion exposes only the prompt and completion totals.
• Latency. Nothing in stored completions records it.
• Trace, span and session ids. Each row stands alone, so imported OpenAI calls do not assemble into traces.
• Failed calls. Only successful completions are stored, so a failed call and the input tokens it spent are invisible on this path.
• Function attribution, streaming detail, outcomes and repository identity, as on every import path.

Troubleshooting

Symptom	Likely cause	What to do
No stored completions found	The calls were not made with store=True, or were made with a different key	Add store=True and use that application’s key
Content columns are empty	--include-content was not passed	Pass it, if the content may be copied
Every route is openai/backfill	No --route was passed	Pass one. OpenAI stores no workflow name
Rows counted twice	Two pulls covered the same completions	This path does not deduplicate