[BUG] Issue with Ollama and context.

[mc9625]

Describe the bug
I have installed the latest version (1.8.1) on two brand new setup: ubuntu 24.0.1 and Raspian OS. On both I have installed Ollama as local service (not a docker container). Everything works fine when I chat (eg I get a proper reply from the model) but there is no context information passed to the prompt. The # Context part of the SystemMessage is missing. I have checked in the memory page and I am able to find declarative message.

To Reproduce
Steps to reproduce the behavior (as exampple`):

1. Install Ollama as local service
2. Upload a document or Upload via URL
3. Ask anything related to the uploaded content
4. Check the system message

[AlessandroSpallina]

what embedder do you have configured?
if you have a proper embedder, check the similarity score you find in the memory tab, the default threshold is 0.7

[AlessandroSpallina]

any update @mc9625 ?

[digitARTI]

Hi all, same problem here! can we help debug?

[pieroit]

Set a proper embedder and test it in the memory page in the admin

[LucaRainone]

I had the same problem, and the reason was the score being below 0.7—even though, to a "human" eye, the question was perfectly relevant to the document (and got a score of 0.6).

By adding more details to the input/question, the score increased, and the system correctly included it in the context.

So, it was simply an embedding model issue in my case. (easily debugged in memory panel). I tried mxbai-embed-large and nomic-embed-text.

I could work on a plugin that can include some context regardless the threshold (something like: "include always at least N chunks of declarative memories". High risk of hallucinations, but at least in-topic, i think. Useful if we have few documents as I did in my playground and first tests).

In addition: does it make sense to allow the modification of the threshold for declarative memories in embedding settings? (I can work on it if you agree)

[lucagobbi]

I had the same problem, and the reason was the score being below 0.7—even though, to a "human" eye, the question was perfectly relevant to the document (and got a score of 0.6).

By adding more details to the input/question, the score increased, and the system correctly included it in the context.

So, it was simply an embedding model issue in my case. (easily debugged in memory panel). I tried mxbai-embed-large and nomic-embed-text.

I could work on a plugin that can include some context regardless the threshold (something like: "include always at least N chunks of declarative memories". High risk of hallucinations, but at least in-topic, i think. Useful if we have few documents as I did in my playground and first tests).

In addition: does it make sense to allow the modification of the threshold for declarative memories in embedding settings? (I can work on it if you agree)

Have you checked the C.A.T. — Cat Advanced Tools plugin? It allows you to configure thresholds and K results for the three main memory collections, plus other useful settings for this type of implementation.

P.S: are you the guy from Gitbar? Big fan of you guys! Keep it up! 🚀

[LucaRainone]

Have you checked the C.A.T. — Cat Advanced Tools plugin? It allows you to configure thresholds and K results for the three main memory collections, plus other useful settings for this type of implementation.

Yo'h. I was looking for this. Thank you!

P.S: are you the guy from Gitbar? Big fan of you guys! Keep it up! 🚀

❤️