Prototypes
We have implemented several prototypes for advanced functionalities that are already useable, but haven't gotten the care that would be advisable for production use. That is, they might not yet be well integrated into the UI yet, or they would need some more optimization for production usage. But you are welcome to try them out, even now they might just be what you need. Please make sure to give us feedback on what you think about them and what you would like to see improved!
AI supported search / RAG assisted question answering
This is a prototype how to support search with AI techniques like embeddings and how to support question answering with RAG (retrieval augmented generation) techniques. It provides a servlet (RAGServlet at /bin/cpm/ai/rag) and two mini applications that let you access the servlet for demonstration purposes. There is not yet an actual component that uses the search since the way it is used is likely quite project specific and might be implemented in Javascript, anyway, so that the servlet could be used as a backend.
- a search that searches for all words that are in the query and rates the top n results using AI embeddings before displaying them
- a RAG supported query answering (no chat) that does this kind of search and then provides the AI with the texts of the top 5 found pages for answering the query.
The mini applications are available at the following URLs (for a local installation):
- Composum: http://localhost:9090/libs/composum/pages/options/ai/prototype.html
- AEM: http://localhost:4502/apps/composum-ai/prototype.html
Precondition: there has to be a lucene fulltext index for the searched repository parts. That is already there in Composum Pages (/oak:index/lucene) but possibly has to be set up at AEM - e.g. with: https://github.com/ist-dresden/composum-AI/blob/develop/backend/slingbase/src/test/resources/lucene/aemContentTextLucene/.content.xml
More discussion / ideas are in https://github.com/ist-dresden/composum-AI/blob/develop/featurespecs/9RAGvariants.md
AI Page Templating : AI supported page creation in one step
While the content creation dialog can support you when working on individual texts in a page, you might sometimes want support for creating whole pages at once. That's the idea for page templating: you create a template page that contains prompts (instructions) to create the individual texts in the page. For creating an actual page from that template you copy it and supply the AI with text(s) from which it can gather the information you want to present on the page (currently either in form of one or more URL(s) or a text). Since it'll likely take some work to get the prompts right, that will likely only be useful if you have quite a number of very similar pages to create about different topics / products etc.
For more details see the AI Page Templating.