ideas from meeting with David and more !

[agaldemas]

• introduce Flowise in the loop

• use local llm (ollama + compatible model)

• add vector storage to store retrieved POIs as embeddings, then use as retriever to augment prompt instead of giving the context-broker results...

• add additional agent that can search on the web, and add some results to the response

• format better the output (display markdown return transformed in html)

[dncampo]

Alain. Please, create a branch from feature/fiware_integration in order the try out the local model using the CB

[agaldemas]

sorry @dncampo, @javicond3,

I did not saw your message, I forked the repo @ mine: https://github.com/agaldemas/Poc_LLM_CB_FIWARE
and made a pull request...
may be it wasn't the best way to do...
but you can merge on the feature/fiware_integration branch !!!

Tell me if you want I create the branch anyway...

[agaldemas]

I will do a new branch for flowise integration...

[agaldemas]

• Still have to fix the display of markdown response transformed in html (for the moment the component display a string, not html, I have to find a way to display as html !

[agaldemas]

@javicond3

Your idea by using openai.beta.threads& openai.beta.assistants from openai api, was to be able to keep the context in a thread and using an assistant to tune the behavior of the chat, provided by openai, but unfortunately it's not supported by ollama...

I had to use openai.chat.completions, which is working, but we loose the context of a conversation :( ...

=> this is a feature we can introduce using flowise (which use langchain under the hood)

• For ollama version use a local context to keep conversation, through Flowise

[javicond3]

Yes, threads are only supported now by OpenAI. You can send the context (or a summary) in every iteration but is not efficient... Does flowise support it?

[agaldemas]

Yes, threads are only supported now by OpenAI. You can send the context (or a summary) in every iteration but is not efficient... Does flowise support it?

you can do this in many different ways with flowise, manage prompt, store context, inject embeddings, prepared in vector db, etc....