LlamaIndex
Log in

Find answers from the community

Updated 2 days ago

Search-o1: Agentic Search-Enhanced Large...

OOmriNach
·
Has anyone figured out how to trigger RAG retrieval inside a reasoning models <think> steps? Similar to the Search-o1 paper here:https://search-o1.github.io

In my opinion, this would be the best way to use deepseek to optimize RAG since new retrieve knowledge can push the reasoning chain in a different direction.

I tried to make the system prompt of deepseek llama 8b amd qwen distil 32b output <search> tokens but it does not like to follow instructions !
L
O
5 comments
LLogan M
This is pretty tricky

You'd have to stream the response, interrupt when you see the tokens and do the search, insert the results back, and invoke a new stream. In the new stream, you'd have to prefill the assistant response
LLogan M
Pretty sure deepseek supports tool calls over their api though right? You could use the normal openai tools api?
OOmriNach
Yeah I’ll try tool calls but deepseek really doesn’t follow prompt instructions well inside its chain of thought… it’s very odd. Might need to be fine tuned for that.
OOmriNach
I see the perplexity search with deepseek only uses the model to reason over retrieved documents but I’m surprised no one is implementing it as a way to come up with the search path in the first place. Seems obvious to me as how humans search.
OOmriNach
@Logan M seems like that’s the approach of o3 deep researcher based on the announcement
Add a reply
Sign up and join the conversation on Discord