feat: add configurable rerank timeout (rerankTimeoutMs)

[ggzeng]

Summary

The cross-encoder rerank request timeout was hardcoded at 5 seconds, which is too aggressive for self-hosted rerank services (e.g. HuggingFace TEI via Infinity) that may need 6-7 seconds to respond.

Changes

• Add rerankTimeoutMs optional field to RetrievalConfig interface
• Replace hardcoded 5000 with config.rerankTimeoutMs ?? 10_000
• Default raised from 5s to 10s for better compatibility with self-hosted endpoints

Usage

{
  "retrieval": {
    "rerank": "cross-encoder",
    "rerankTimeoutMs": 20000
  }
}

Backwards Compatibility

Fully backwards compatible. Existing configs without rerankTimeoutMs will get the new 10s default.

[AliceLJY]

Hey @ggzeng, thanks for the contribution! Just a heads-up — PR #356 by @jlin53882 was submitted about 11 hours earlier and implements the same rerankTimeoutMs feature. It has already gone through a round of review feedback and includes additional changes (schema in openclaw.plugin.json, PluginConfig type in index.ts, explicit default in DEFAULT_RETRIEVAL_CONFIG, and timeout cleanup in the finally block).

One notable difference: your PR raises the default timeout from 5s to 10s, while #356 keeps the existing 5s default. The 10s default is arguably a better choice for self-hosted endpoints, but that's a design decision worth discussing in one place.

Could you coordinate with #356? If the higher default is important to you, it might be simplest to suggest that change on #356 rather than maintaining two parallel PRs for the same feature. If your approach differs in a meaningful way I'm missing, happy to hear more!

Thanks!

[AliceLJY]

Hi @ggzeng — thank you for taking the time to submit this PR! Making rerankTimeoutMs configurable is a real need, and we appreciate you jumping in to address it.

After reviewing both contributions, we're going to close this in favor of #356, which was submitted a bit earlier and covers the same feature more comprehensively — it adds the JSON Schema definition in openclaw.plugin.json, updates the PluginConfig type in index.ts, wires up the config in retriever.ts, and includes timeout cleanup in a finally block.

That said, your PR raises a valid point that #356 doesn't address: the default timeout value. A 10s default may be more appropriate for self-hosted reranker setups where latency is higher. We'd love it if you could leave a comment on #356 suggesting they consider raising the default — that feedback would be genuinely useful there.

You can follow along here: #356

Thanks again for the contribution, and we hope to see you around the project!