Question(s) about RAG

HCL Domino Domino Forum
1

Please complete the details below then remove this line:

Domino/Notes Version: 14.5.1
Operating System: win server
Client Notes 14.5.1

Recently I’ve tried some features of the RAQ functionalities in Domino but the results are sometime misleading or incomplete. Just to add an additional level of info: yes, I know Domino and no, I am absolutely a newbie on LLM /AI /RAG.

The last try was with a ticketing applications already in place: something like date, customer, problem, calssification of the problem, hours spent.

I’ve tried to use RAG on this application , the basic idea is to interrogate these entries with questions like “what problem occur more often” or similar but after load updall -w on my db the domino console showed me several error message like:

LLMVectorCollectionEmbedText> Skipping embedding for item Subject in NoteId 2B2E. Chunk size 27 is below minimum 30

It seems that a small piece of information (like username o classification) could not be embedded , just because is too small

I’ve read some docs about embedding and I understand rhere are a cuople of parameters which could be useful in RAG like chunk_size / chunk_overlap but I’ve no idea where to put them in Domino (or if they’re really useful).

I’m a little concerned : a lot of domino applications own small pieces of information and loosing RAG capalities on these could move DominoIQ RAG to be not 100% useful.

BTW: I’m currently using Meta-Llama-3.1-8B-Instruct-Q5_K_L as LLM model and qwen3-embedding-8b-q8_0 as embedding model. No guard model

2

Hello Fabdpl,

Here are the most useful HCL docs that explain what you’re seeing and how RAG works in Domino:

HCL Domino IQ documentation

Domino Query Results Processor (RAQ) documentation

Domino AI and embedding configuration

Domino updall task documentation

3

Links to the documentation are broken.

4

Hello,

The current RAG based searches are semantic or dense searches. We cannot lower the limit(minimum 30 chunks) further down.

You can open a case with HCL Support as needed. Hope it helps.

Best regards,

Xiaoyun

5

What I found would work for these scenarios is to add an additional Field onto your Form that captures the values from ALL the Fields you want to use in the vectorized data.

You can name the field whatever you would like, but for this example let’s call it “DocOverview”. In that Field, you would use an “Input Translation” value, or better yet, make the Field “Computed”. Then make the value compute to all the Fields that you want to use in Domino IQ RAG, similar to the following:

fld1:=”Customer: “+Customer;

fld2:=”Problem: “+Problem;

fld3:=”Classification: “+Classification;

fld4:=”Request Date: “+@Text(Date);

fld5:=”Hours Spent: “+@Text(HoursSpent);

cr:=@Char(10);

fld1+cr+fld2+cr+fld3+cr+fld4+cr+fld5

This value will be computed when the Document is saved.

And now you can change your Domino IQ Command document to use just that single Field. It will capture all the data from the other Fields and it will also overcome the “Chunk size” challenge that you are facing.

And, Yes, I understand that this would need to be applied to all your existing Document to make it work on historical data. For that, you can create an @Formula agent that runs on all Selected Documents, with the following @Formula:

@Command([ToolsRefreshSelectedDocs])

1 Like