Add OrcaReplay

[xizhuomengcontin]

Adds integrations/orcareplay.md.

What it is

OrcaReplay (Apache-2.0, npm, Node 20+) records a Haystack run at the HTTP boundary to the model provider — from outside the process — and serves the recording back so the same run happens again with no provider contacted and no key needed.

There is nothing to register and no component to add, because OpenAIChatGenerator leaves api_base_url=None and the OpenAI SDK then falls back to OPENAI_BASE_URL — the variable the recorder sets for the child process it launches, and only for that process.

Verified before submitting

haystack-ai 3.1.1 on Python 3.12, both a standalone OpenAIChatGenerator and a full Pipeline (ChatPromptBuilder → OpenAIChatGenerator), each recorded and then replayed with the origin process killed:

$ orca replay last
info replaying exchanges=1 egress=blocked
info replay.done reused=1/1 exact=1 divergences=0 unmatched=0 exit=0

exact=1 is a byte-for-byte match against a deterministic local origin standing in for a provider, so it is a real comparison rather than a model happening to repeat itself.

I filed the same result as a Show and tell in the main repo — mentioning it here so the overlap is visible rather than discovered.

Type and limits

I filed it under type: Monitoring Tool as the closest existing category — it is really a record/replay debugger, so if you would rather see a different type (or a new one), say the word and I will change it.

The page has an explicit Limits section, because two things here are easy to overstate and one is a genuine gap:

• blocked egress is model-provider egress, not network isolation — retrievers and custom components still run for real, so it is not a sandbox;
• a matching replay is not a determinism result, since the model is not re-asked;
• /v1/embeddings is not captured. I measured this rather than assuming it: embedding traffic is passed through to the live provider and is absent from the recording, so a replay refuses it with a named error instead of serving a wrong vector. A pipeline that embeds at query time cannot be fully replayed today. I would rather say that on the integration page than let someone find out at replay time.

No PyPI package — it is an npm CLI that launches your Python process, so only repo is set, which the README allows. No logo included; happy to add one if you would like.

Disclosure: I maintain it.

[vercel]

@xizhuomengcontin is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.