[Add costomized metrics] how to implement a new metrics so that I can use for collection indexing？

[jayxio]

I want to implement some modification based on Internal Product metrics, how to contribution the code?

[yhmo]

So far we don't have a plan to support customized metric type. This feature requires a good design for how to define customized metric from client-side and how to convert the options to algorithms in search engine. So, I don't think a specific user should implement it. We intend to collect more requirements from more users, and summarize them before we can design it well. So far, we don't have enough input from users to understand what customized metrics they want.

It would be helpful if you could offer more details about what metrics you want.

[jayxio]

Hello yhmo, thx for the reply.
The customized metrics here is: two vectors do elementwise product, then do elementwise non-linear transformation and add all element's values together(a modified version of Internal Product.)
From my understanding, to implement such metric, will need us write this specific function, and corporate them in the collection framework, and use this metric to build the index when inserting new data for user search?
It does not have to be client-side feature. Is it feasiable for a new developer to develop such new metrics for buiding index and searching? If possible, how to get hands on doing it?

[yhmo]

In my opinion, customized metric should be able to cover more specific use cases. The client-side api provides some operators or functions so that user can define their metrics flexibly. The server organizes the user-define operators/functions and translates to runable algorithm, maybe JIT compiler is involved.
If we add a new metric type for each specific use case, there might be dozens of special metric types, I don't think it is a good design.

Wait @xiaofan-luan @liliu-z comment.

[xiaofan-luan]

To add new metrics, i think we need to follow the following stype

https://github.com/zilliztech/knowhere/blob/main/src/common/metric.h

at knowhere, add a special metrics named customize. The customize function need to read a config and load dynamic lib and user need to implement DistanceComputer interface with customize distance(This is performance related)

Then at milvus, add a wrapper of supporting to use customized function is all we need.

@jayxio let me know if this is clear to you

[jayxio]

Thx, It makes sense to me. Does milvus-lite and Zilliz use Knowhere as its backend engine as well?

[xiaofan-luan]

Yes, zilliz and milvus lite use Knowhere as well.
But it tailored a lot of use detailed index and even code. Knowhere is more like a wrapper defined metrics and many other details. Under the hood there is different implementations for index, for example, under opensource you need HNSW or IVF, on milvus lite you have FLAT index and on Zilliz it's called cardinal

[jayxio]

Sounds like it decoupled metrics and index implementation by zilliz, milvus, milvus-lite and Knowhere right? An exquisite design.