fix: local embedding model compatibility (Qwen3/vLLM/Ollama) + configurable rerank timeout

[yjjheizhu]

Summary

Fixes compatibility issues when using local embedding models (Qwen3, vLLM, Ollama) that don't support the dimensions parameter, and adds configurable rerank timeout for slow local rerank models.

Changes

1. Add omitDimensions config switch

• New embedding.omitDimensions: boolean option (default: false)
• When enabled, stops sending the dimensions parameter in embedding requests
• Users opt-in explicitly, avoiding false positives for local models that do support matryoshka

2. Friendly error hint for matryoshka errors

• Detects errors matching matryoshka, dimensions not supported, unknown param: dimensions
• Returns actionable message: Set "embedding.omitDimensions": true in your config
• Users can self-diagnose without reading documentation

3. Configurable rerank timeout

• New retrieval.rerankTimeoutMs: number option (default: 5000)
• Replaces hardcoded 5s timeout that fails for slow local rerank models
• Fully backward compatible

Testing

• All existing tests pass
• New test cases for matryoshka error detection (3 variants covered)

Breaking Changes

None — all changes are opt-in with sensible defaults.

[chenjiyong]

I validated PR #264 locally with the OpenClaw harness. The main omitDimensions change works as expected, and both the host functional test and the real agent E2E passed. However, I found one issue in the same PR: rerankTimeoutMs was added in src/retriever.ts, but it is not declared in openclaw.plugin.json. Because retrieval uses additionalProperties: false, adding retrieval.rerankTimeoutMs to plugin config currently fails openclaw config validate with “must NOT have additional properties”.

[AliceLJY]

Thanks for the contribution! 🙏

There are currently merge conflicts with main. Could you please rebase your branch against main so we can move forward with the merge?

If you need any help with the rebase, feel free to ask!

[AliceLJY]

Hey @yjjheizhu, thanks for addressing the rerankTimeoutMs feedback!

However, the latest commit (b60b58d) seems to have picked up unrelated changes during the rebase:

• New tagPrefixes config + extractTagTokens / bm25OnlyRetrieval methods (~60 lines of new retrieval logic) — this is a standalone feature, not part of this PR's scope
• json5 and proper-lockfile dependencies added to package-lock.json — these belong to feat: cross-process file lock + CLAWTEAM_MEMORY_SCOPE support #311

Could you clean up the branch so only the changes related to this PR (omitDimensions + error hints) remain? A quick git rebase -onto or interactive rebase to drop the extra changes should do it.

Let me know if you need a hand!

[yjjheizhu]

Superseded by #315, which is the clean replacement PR containing only the omitDimensions fix and related schema/test updates. Closing this one to avoid merging the mixed-in unrelated changes and current merge conflicts.