Find answers from the community

Updated last year

Retrieval

i'm using a vector index as chat engine in context mode
L
s
43 comments
What you described is essentially how the context chat engine works πŸ€”
for the retrieval part?
i trace through code
the string gets sent in for subsequent questions are not llm enhanced
how do i reduce carbon
then what about water usage
so the second question doesn't search for the right thing,
@Logan M i think i will add that to the context chat implementation
do you agree with that approach?
or should it be a "memory retriever" so that it can be plugged into other places?
I think that makes sense! My only worry qould be if a conversation changes topics, now the chat history might retrieve unhelpful data πŸ€”
actually, there is a reset in the chat already
so when the chat engine is reset the retrieval should start over as well
that's how i'm using it right now
it is a bit of pain that sometimes the users don't reset
and no amount of ux prompting can get them to do it
until they realize the answers and starting to get a little weird
necessary evil i suppose
i assume you rather have it in context chat engine rather than a new retriever type, correct?
Yea I agree, some kind of new mode for it would be cool πŸ™‚
hi @Logan M fyi i have a working version of the followup chat retriever
but tuning the prompt to get a good combined question is tricky
especially given users are not really always asking follow up questions
i implemented in a custom chat engine
won't pr unless i get a good result. may go away from chat being the default mode and use query instead
in a mostly unrelated question
what's the future for composable graph vs route based query?
they seem to overlap a lot
i started off using a composable graph
to select which underlying index to query
now looking to move to route, as that seems more flexible
good question! Right now composable graph is mostly on life support lol it may get deprecated officially soon, but right now none of the core team is working on it
both fundamentally uses llm to look at the "description" of an index and decide which to query, right?
exactly why i asked
so definitely going over to router based
Yea exactly, the concepts are similar, composable graph is just hard to maintain for us haha
how big is the dev team anyways? you guys raised a good round right?
Yea we raised 8.5M for the seed round. Theres about 5 of us right now, plus a few contractors. Heavily interviewing right now too, want to expand to 8 full-time or so soon I think
i would be interesting having a conversation about how i could potentially be more involved.... google my name first though. jun ying
Add a reply
Sign up and join the conversation on Discord