Hi, I'm using an instance of LiteLLM as a remote e...
# community-help
l
Hi, I'm using an instance of LiteLLM as a remote embedder. For some network misconfiguration the LiteLLM instance went down for a while. It seems that this event caused typesense to delete the embedding field. Was it just a coincidence? In any case, the embedding field disappeared from all collections at some point.
👀 1
a
Hey @Luca Lusso It shouldn't have deleted. It's likely that it's still in disk, but when trying to indexing it, it bumped into the network error. When you saw that it deleted the embedding field, how did you checked it? You looked at the collection schema? Could you check if it's present in your documents? The solution for that might be a quick restart (since your cluster is small). To do that, you can go to Configuration Change -> Modify and schedule it without changing any of the parameters.
l
yes, I look at
/collections
endpoint
and if I perform a semantic search it says that is impossibile because no embedding field exists
(this issue is not on typesense cloud, but in a customer self-hosted instance)
we have recreate the schema so the issue is fixes now. I just need to understand why it happened
a
and if I perform a semantic search it says that is impossibile because no embedding field exists
Yeah it verifies the schema collection as well. Recreating the schema does fix, but maybe restarting the cluster might be a better/faster approach. Just to confirm, did the self-hosted became unavailable when you were creating/altering the collection? I've seem this situations when, on the initial load, the cluster doesn't have connection to the embedder.
l
Just to confirm, did the self-hosted became unavailable when you were creating/altering the collection? I've seem this situations when, on the initial load, the cluster doesn't have connection to the embedder.
we haven't checked for this (we're not in production yet, so we don't have much load)
a
If this is frequent, you can also generate your embeddings, store somewhere, and then send the documents to Typesense with the generated embeddings: https://typesense.org/docs/29.0/api/vector-search.html#option-a-importing-externally-generated-embeddings-into-typesense
l
yes, this can be an option