Skip to content

Latest commit

 

History

History
64 lines (43 loc) · 2.64 KB

9RAGvariants.md

File metadata and controls

64 lines (43 loc) · 2.64 KB

Retrieval Augmented Generation (RAG) Variants

Status: prototypical mini implementation - a servlet and experimental UI
Composum: http://localhost:9090/libs/composum/pages/options/ai/prototype.html
AEM: http://localhost:4502/apps/composum-ai/prototype.html

Background

We assume that we have an informational website with static content.

Basic idea

We want to combine a text search with LLM RAG, possibly with vector (embedding) search, possibly not. Ideas for that:

  • Intelligent search that uses a LLM to improve search results
  • Answer questions in a chat better by supporting the LLM with a RAG search

Basic implementation decisions

  • Instead of a vector database we calculate embeddings on demand, possibly with a cache.

Out of scope

We do not want to employ a separate vector database for RAG, but see what we can do without that.

Implementation

Search variants

The JCR repository has an integrated lucene search that can be used to search for words in a query and rate the results.

  • The top N results can either be used directly, or
  • their embeddings can be compared with the query embeddings to filter the best out, or
  • the LLM can rate them directly.

Search trigger variants

  • The search can be triggered directly from the users query, or
  • the LLM can be asked to preprocess it, or
  • in a chat the LLM can trigger search actions on its own when it sees that it needs more information (or always after a user message).

Generated content variants

  • Rating of links for a search result
  • Answer with related links
  • Just an answer, supported by a search

More ideas for later

  • Generate recommendations for related content for a page, either at editing time or on the fly
  • Generate links from a page for the CMS editor, to generate teasers or links to related content
  • Find related content to support the CMS editor in creating a page: additional source for content creation dialog?
  • Possibly: find related assets (embedding of description?)
  • ??? "Real time content adaption" : create pages about something on the fly? "View" according to a given topic?
  • Content analysis: clustering of pages by content, through embeddings
  • Personalized Content Recommendations based on interaction history, using embeddings for finding related content.
  • clustering of user feedback (comments)?
  • adaption of search results by integrating user comments
  • ??? custom content feed for a user, based on previous interactions and description of intent

PROMPT: create 20 more variants how RAG can support either the end user or the content creator. Describe the features from the users / content creators perspective, not the implementation.