L2 distance computation misunderstanding in documentation about redisearch

[Voveka98]

Ok, thanks a lot!

[meiravgri]

Hi @Voveka98
When passing a binary blob to the CLI, enclosing it in quotes is essential. Otherwise, each character is interpreted as one byte. You can find more information about this in the Redis documentation.

The binary blob "\x01\x01\x01\x01" you provided is only 4 bytes long, but a 16-byte blob is expected (dim * sizeof(float)).
Ensure that you're passing the correct blob size for your data.

In addition to that, we don't apply sqrt on the distance value. Therefore, the expected distance is

(sum((x_i - y_i)**2))

Good luck :)

[Voveka98]

Hi @meiravgri!
Thanks for explanation, maybe update documentation (https://redis.io/docs/interact/search-and-query/advanced-concepts/vectors/)?

And another question:
I am trying to find most similar vector in type float32 dim 512.
I do next:

func embeddingToBytes(embedding []float32) []byte {
    bytes := make([]byte, 0, len(embedding)*4)

    for _, val := range embedding {
        bits := math.Float32bits(val)
        buf := make([]byte, 4)
        binary.LittleEndian.PutUint32(buf, bits)
        bytes = append(bytes, buf...)
    }

    return bytes
}

Is it right way to do this?
Thanks in advance!

[meiravgri]

Regarding documentation, we acknowledge that while it might be a bit confusing, the decision to skip sqrt operation is made to save computation time during the query process.

For the second question, I'm not personally familiar with Go. it's worth noting that the Go client doesn't directly support Redisearch integration. If you're looking for an alternative method, you might consider referring to the Python client documentation, where they provide an example of converting vectors into binary blobs using the tobytes() function.

Additionally, you can explore the RedisVL repository, which offers an easy start Python client dedicated to using Redis as a vector database.