Modifying Sliced Documents

[xuzeyu91]

Context / Scenario

I hope to add modifications to the sliced document after Import. I have encapsulated the Handle implementation of QA slicing based on step.

The problem

For example, the following are my prompt words:

I will give you a paragraph of text, study them, and organize the learning results. The requirements are:
1. Ask up to 25 questions.
2. Provide answers to each question.
3. The answer should be detailed and complete, and can include markup elements such as regular text, links, codes, tables, announcements, media links, etc.
4. Return multiple questions and answers in format:
Q1: Question.
A1: Answer.
Q2:
A2:
My text: "" {$input}} ""

In this case, after importing the document, I may need to manually modify the QA extracted by llm. I hope this can be implemented from Kernel Memory because it is compatible with multiple different vector databases

Proposed solution

Add the ability to modify and re vector sliced documents in MemoryServerless

Importance

would be great to have

[xuzeyu91]

I don't know if this is in the plan, but my friends and I urgently need this feature. We are unable to make modifications to a certain segment, which leads to the need to re import the entire document

[dluc]

Documents can be updated using the document ID returned by the memory API. As for internal storage, you should consider using something like Entity Framework or the proper library for the chosen storage, in order to manipulate records following your requirements. Note that internal details can change over time.

About the prompt and how chunks are injected into the LLM prompts, if that's needed, the prompt template is customizable.

[xuzeyu91]

Documents can be updated using the document ID returned by the memory API. As for internal storage, you should consider using something like Entity Framework or the proper library for the chosen storage, in order to manipulate records following your requirements. Note that internal details can change over time.

About the prompt and how chunks are injected into the LLM prompts, if that's needed, the prompt template is customizable.

For example, when I import a document, what I want is finer fragments within the document, rather than re uploading the entire document

[dluc]

I'm afraid that feature is not supported. Memories are updated only via documents upload. If you need finer records management, you'll need to access the storage directly using bespoke code.