How to convert an embeddings vector to be saved as %Vector

Question

Question

Ditmar Tybussek · Jun 23

#Embedded Python #Generative AI #Vector Search #InterSystems IRIS

I try to get a vector from calling GetEmbedding, but i failed to convert it into a vector

Here is a simplyfied sample class:

Class User.myclass Extends %Persistent
{ Property myVECTOR As %Vector(CAPTION = "Vector");

Property myProperty As %String(MAXLEN = 40) [ Required ];

}

here the GetEmbedding part from User.mymethods:

...
ClassMethod GetEmbedding(sentences As %String) As %String [ Language = python ]
{
  import sentence_transformers   model = sentence_transformers.SentenceTransformer('C:/InterSystems/IRIS/lib/python/Lib/site-packages/sentence_transformers/models/all-MiniLM-L6-v2')
  embeddings = model.encode(sentences)   embeddings_list = [str(embedding.tolist()) for embedding in embeddings]
  return embeddings_list
}.....

I try to save the VECTOR into myVECTOR, but that will fail because the vector is not in the right format:

set VECTOR=##class(User.mymethods).GetEmbedding("this is my text")

/// here I need to convert the VECTOR, but I do not know the right command

set data=##class(User.myclass).%New()
set data.myProperty ="anything"
set data.myVECTOR=VECTOR
set ok=data.%Save()

Question: is there a method to convert the vector into the right format to be saved as %Vector

$ZV: IRIS for Windows (x86-64) 2024.1 (Build 262U) Thu Mar 7 2024 15:55:25 EST

Discussion (4)1

Log in or sign up to continue

score 0 · Answer 1 · 2024-06-23T08:51:18-04:00

Hallo @Ditmar Tybussek lange nix mehr gehört von dir !

See my article Using VECTORs in ObjectScript.

related to your code:
Simple Set vector=.... doesn't work
it is Set $VECTOR(vectorname, pos,type) very clear example
and type can never be changed
position is less sensitive.

In your case using SQL functíon TO_VECTOR() might be quite comfortable
as it does all the checks for you under cover.
i also used all-MiniLM-L6-v2 (as PY method) starting from label sc3 in the example
in my OEX package Vector-inside-IRIS

LG. aus Wien

score 0 · Answer 2 · 2024-06-23T10:07:47-04:00

Danke Robert,

is the TO_VECTOR(SQL) the "only" way to convert an array, or is there an alternativ, if someone likes to see "what's under the cover" and set the vector while creating a record?

Ditmar

score 0 · Answer 3 · 2024-06-23T16:08:41-04:00

not really since as you know there is always ObjectScript under most SQL Functions.
BUT as mentioned: I found TO_VECTOR() the most comfortable one.
without SQL it might look like this: or worse

;; assume a $LB structure of %Integer as input
set intlist=$LB(...........)
;; fill vector
set maxvecsize=22    ; just an assumption
 for i=1:1:maxvecsize {
  set val=$li(intlist,i)
  if '(val\1=val)  continue
  set $vector(vec,i,"int")=val
 }
write $isvector(vec)
write vec
zwrite vec

I think the difference is obvious.

Servus

score 0 · Answer 4 · 2024-06-24T05:19:41-04:00

Try this example

Class User.myclass Extends %Persistent
{

Property myVECTOR As %Vector(CAPTION = "Vector", DATATYPE = "INTEGER");

Property myProperty As %String(MAXLEN = 40) [ Required ];

ClassMethod GetEmbedding1(sentences As %String) As %String
{
  q "2,4,6,8"
}

ClassMethod GetEmbedding2(sentences As %String) As %DynamicObject
{
  q [1,3,5,7,9]
}

ClassMethod Test()
{
   d ..%KillExtent()
   
  set data=##class(User.myclass).%New()
  set data.myProperty ="anything 1"
  set data.myVECTOR=##class(User.myclass).myVECTORDisplayToLogical(##class(User.myclass).GetEmbedding1("this is my text"))
  d $system.OBJ.DisplayError(data.%Save())

   ; OR
   
  set data=..%New()
  set data.myProperty ="anything 2"
  set data.myVECTOR=data.myVECTORDisplayToLogical(..GetEmbedding2("this is my text"))
  d $system.OBJ.DisplayError(data.%Save())

  zw ^User.myclassD
}

}


USER>d ##class(User.myclass).Test()
^User.myclassD=2
^User.myclassD(1)=$lb("",{"type":"integer", "count":4, "length":4, "vector":[2,4,6,8]}  ; <VECTOR>,"anything 1")
^User.myclassD(2)=$lb("",{"type":"integer", "count":5, "length":5, "vector":[1,3,5,7,9]}  ; <VECTOR>,"anything 2")