Unsupported embedding model: bge-m3. Either add it to EMBEDDING_DIMENSIONS or set embedding.dimensions in config

[jszhang98]

Plugin Version

memory-lancedb-pro/beta

OpenClaw Version

2026.03.23

Bug Description

When I run the plugin firstly, If I didn't give the "dimentions", it gave following error.

20:52:26 [plugins] memory-lancedb-pro failed during register from C:\Users\Administrator.openclaw\extensions\memory-lancedb-pro\index.ts: Error: Unsupported embedding model: bge-m3. Either add it to EMBEDDING_DIMENSIONS or set embedding.dimensions in config.

or, if I give the "dimentions" in configuration file. the error gone. but I got the following error.

21:17:56 [gateway] memory-lancedb-pro: initialized successfully (embedding: FAIL, retrieval: FAIL, mode: hybrid, FTS: disabled)
21:17:56 [gateway] memory-lancedb-pro: embedding test failed: Failed to generate embedding from api.modelarts-maas.com: 400 ModelArts.81001: 400 Model "bge-m3" does not support matryoshka representation, changing output dimensions will lead to poor results.
21:17:56 [gateway] memory-lancedb-pro: retrieval test failed: Failed to generate embedding from api.modelarts-maas.com: 400 ModelArts.81001: 400 Model "bge-m3" does not support matryoshka representation, changing output dimensions will lead to poor results.

this mean that I config it it didn't work, I don't confit it, it can not start the openclaw gateway.
(I used Huawei Bge-m3 model for embeddings. windows operaton system, openclaw 3.23 )

`
"memory-lancedb-pro": {
"enabled": true,
"config": {
"embedding": {
"apiKey": "R-v6OoE9b65mgnZX-zgtkw42vZS2Ng0pM75nUJRPk2OLESx54El09CsKgBJpCPo_2hiQ0offPtHD_8SlHj1NkA",
"baseURL": "https://api.modelarts-maas.com/v1",
"model": "bge-m3",
"normalized": true
},
"autoCapture": true,
"autoRecall": true,
"smartExtraction": true,
"extractMinMessages": 2,
"extractMaxChars": 8000,
"sessionMemory": {
"enabled": false
}
}
},

`

Later, I had to comment the code:

`export function getVectorDimensions(model: string, overrideDims?: number): number {
if (overrideDims && overrideDims > 0) {
return overrideDims;
}

const dims = EMBEDDING_DIMENSIONS[model];
// if (!dims) {
// throw new Error(
// Unsupported embedding model: ${model}. Either add it to EMBEDDING_DIMENSIONS or set embedding.dimensions in config.
// );
// }

return dims;
}`

and change the following code also:
`const EMBEDDING_DIMENSIONS: Record<string, number> = {
"text-embedding-3-small": 1536,
"text-embedding-3-large": 3072,
"text-embedding-004": 768,
"gemini-embedding-001": 3072,
"nomic-embed-text": 768,
"mxbai-embed-large": 1024,
"bge-m3": 1024, // changed this row.
"all-MiniLM-L6-v2": 384,
"all-mpnet-base-v2": 512,

// Jina v5
"jina-embeddings-v5-text-small": 1024,
"jina-embeddings-v5-text-nano": 768,
};`

I think these two changes would lead not to pass the dimentions param to 'bge-m3', it worked.

Once all lancedb initially done. I found that I don't have to set the 'dimentions' param. it can start the gateway also.

I finally think that bge-m3 can not accept the 'dimentions' param.

Please follow this issues.

thanks.

Expected Behavior

system should support that there is no 'dimentions' param for some models.

Steps to Reproduce

1. install a memory-lancedb-pro firstly and newly.
2. use bge-m3 model of Huawei
3. start openclaw gateway.

Error Logs / Screenshots

20:52:26 [plugins] memory-lancedb-pro failed during register from C:\Users\Administrator\.openclaw\extensions\memory-lancedb-pro\index.ts: Error: Unsupported embedding model: bge-m3. Either add it to EMBEDDING_DIMENSIONS or set embedding.dimensions in config.


21:17:56 [gateway] memory-lancedb-pro: initialized successfully (embedding: FAIL, retrieval: FAIL, mode: hybrid, FTS: disabled)
21:17:56 [gateway] memory-lancedb-pro: embedding test failed: Failed to generate embedding from api.modelarts-maas.com: 400 ModelArts.81001: 400 Model "bge-m3" does not support matryoshka representation, changing output dimensions will lead to poor results.
21:17:56 [gateway] memory-lancedb-pro: retrieval test failed: Failed to generate embedding from api.modelarts-maas.com: 400 ModelArts.81001: 400 Model "bge-m3" does not support matryoshka representation, changing output dimensions will lead to poor results.

Embedding Provider

Other

OS / Platform

Windows 11 pro

[AliceLJY]

感谢反馈！bge-m3 不在内置的 EMBEDDING_DIMENSIONS 列表里，但可以通过配置解决：

在你的插件配置里手动指定 dimensions：

{
  "embedding": {
    "model": "bge-m3",
    "dimensions": 1024,
    "baseURL": "你的 embedding API 地址"
  }
}

bge-m3 默认输出 1024 维。设置 embedding.dimensions 后插件会跳过内置列表查找，直接使用你指定的值。

如果还有问题请贴一下完整的错误日志和配置（隐去 API key），方便排查。

[jszhang98]

在初始化的时候，如果我加上dimentions 这个参数，模型会返回。
21:17:56 [gateway] memory-lancedb-pro: embedding test failed: Failed to generate embedding from api.modelarts-maas.com: 400 ModelArts.81001: 400 Model "bge-m3" does not support matryoshka representation, changing output dimensions will lead to poor results.

如果我不加dimentions 参数它一开始就报错：
20:52:26 [plugins] memory-lancedb-pro failed during register from C:\Users\Administrator.openclaw\extensions\memory-lancedb-pro\index.ts: Error: Unsupported embedding model: bge-m3. Either add it to EMBEDDING_DIMENSIONS or set embedding.dimensions in config.

这形成了一个死循环。

我想bge-m3模型本身不支持可变的dimention,所以会报上面那个警告。

不过我现在已经解决这个问题了：
先注释那个无参报错。等lancedb 初始化完了，再取消注释，之后没有参数也可以正常启动了。我想可能是如果已经有lancedb的结构，它会自动从lancedb找维度

谢谢

[AliceLJY]

Thanks for the detailed follow-up — this is a real bug.

The problem is that the plugin conflates two different uses of dimensions:

1. Internal: telling LanceDB what vector size to use for the table schema
2. API request: passing dimensions to the embedding provider

For models like bge-m3 that don't support matryoshka (variable dimensions), passing dimensions in the API request triggers a 400 error, but omitting it makes the plugin fail at startup because it can't determine the vector size.

Your workaround confirms the fix direction — once the LanceDB table exists, the plugin reads dimensions from the existing schema and skips the API parameter. The startup path should do the same: use dimensions config only for internal schema creation, not as an API request parameter.

I haven't personally hit or tested this scenario, so I won't commit to a specific fix design yet. Would you be interested in submitting a PR? The relevant code is in src/embedder.ts — the goal would be to separate "schema dimensions" from "API request dimensions" so models like bge-m3 work out of the box.

If not, we'll pick it up when we can reproduce it locally.

[jszhang98]

OK. I will fork it

…

On Fri, Mar 27, 2026, 15:02 安闲静雅 ***@***.***> wrote: *AliceLJY* left a comment (CortexReach/memory-lancedb-pro#339) <#339 (comment)> Thanks for the detailed follow-up — this is a real bug. The problem is that the plugin conflates two different uses of dimensions: 1. *Internal*: telling LanceDB what vector size to use for the table schema 2. *API request*: passing dimensions to the embedding provider For models like bge-m3 that don't support matryoshka (variable dimensions), passing dimensions in the API request triggers a 400 error, but omitting it makes the plugin fail at startup because it can't determine the vector size. Your workaround confirms the fix direction — once the LanceDB table exists, the plugin reads dimensions from the existing schema and skips the API parameter. The startup path should do the same: use dimensions config only for internal schema creation, not as an API request parameter. I haven't personally hit or tested this scenario, so I won't commit to a specific fix design yet. Would you be interested in submitting a PR? The relevant code is in src/embedder.ts — the goal would be to separate "schema dimensions" from "API request dimensions" so models like bge-m3 work out of the box. If not, we'll pick it up when we can reproduce it locally. — Reply to this email directly, view it on GitHub <#339?email_source=notifications&email_token=AGF6SM372YA355FPGJET3YL4SYRPNA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIMJUGA3DCNZWGE3KM4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4140617616>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AGF6SM2DWZZ5YYFVR4V4L634SYRPNAVCNFSM6AAAAACW5QLC2KVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCNBQGYYTONRRGY> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[AliceLJY]

Thanks. The fix direction for this is now in PR #377 (separate internal schema dimensions from embedding API request dimensions), and that PR is already approved / assigned for merge.

I'm tracking implementation there now; once it lands, this issue should close.