github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #7085] Embedding discrepancies vs SentenceTransformers and between Ollama Versions #66554

New Issue
Open
opened 2026-05-04 07:24:49 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-05-04 07:24:49 -05:00
Owner
Copy Link

Originally created by @nicksrusso on GitHub (Oct 3, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/7085

What is the issue?

I'm working on a RAG app, I'd like to simplify my stack by using ollama instead of sentence transformers but I'm observing some odd behavior from ollama embeddings. Here's what I'm seeing:

  1. Output embeddings don't match sentence transformers
  2. Hitting the "embd" endpoint from the api docs here returns different results than the "embeddings" endpoint described in the blog here
  3. Embeddings returned by different versions of ollama are different.

To test this, I embedded the same prompt with sentence transformers, ollama/embed, and ollama/embeddings. I printed out the first 5 elements in the output vector to inspect, and calculated a cosine similarity matrix between each pair of embedding methods (results in a symmetrical matrix with 1s along the diagonal). I repeated this using all-mini-lm and nomic-embed-text. Here's the code and outputs for ollama 0.3.3 and 0.3.12

code

ollama033_output

ollama_0312

The embeddings are pretty clearly different, both when compared to sentence transformers and between ollama 0.3.3 and 0.3.12.

The cosine similarity between ollama and sentence transformers is pretty close though. I also created a 3500 chunk vectorstore, once with ollama and once with sentence transformers. When I query them, I get the same chunks returned albeit in a different order. That makes me think every vector that is getting embedded is getting transformed somehow (when compared to sentence transformers), but its happening to all of them more or less equally.

Is this expected behavior? If so, is there anywhere I can read up on what ollama is doing under the hood?

OS

Linux

GPU

AMD

CPU

AMD

Ollama version

0.3.3, 0.3.12

Originally created by @nicksrusso on GitHub (Oct 3, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/7085 ### What is the issue? I'm working on a RAG app, I'd like to simplify my stack by using ollama instead of sentence transformers but I'm observing some odd behavior from ollama embeddings. Here's what I'm seeing: 1. Output embeddings don't match sentence transformers 2. Hitting the "embd" endpoint from the api docs [here](https://github.com/ollama/ollama/blob/79d3b1e2bdfc97542a7259b0c839520d39578514/docs/api.md#generate-embeddings) returns different results than the "embeddings" endpoint described in the blog [here](https://ollama.com/blog/embedding-models) 3. Embeddings returned by different versions of ollama are different. To test this, I embedded the same prompt with sentence transformers, ollama/embed, and ollama/embeddings. I printed out the first 5 elements in the output vector to inspect, and calculated a cosine similarity matrix between each pair of embedding methods (results in a symmetrical matrix with 1s along the diagonal). I repeated this using all-mini-lm and nomic-embed-text. Here's the code and outputs for ollama 0.3.3 and 0.3.12 ![code](https://github.com/user-attachments/assets/f0d3620f-4511-49d9-8e87-10a520bbeb62) ![ollama033_output](https://github.com/user-attachments/assets/1870367b-85b6-4fb7-9f07-adfaf80e7cba) ![ollama_0312](https://github.com/user-attachments/assets/81d61fcb-9749-4ed5-ba4a-ac4959587567) The embeddings are pretty clearly different, both when compared to sentence transformers and between ollama 0.3.3 and 0.3.12. The cosine similarity between ollama and sentence transformers is pretty close though. I also created a 3500 chunk vectorstore, once with ollama and once with sentence transformers. When I query them, I get the same chunks returned albeit in a different order. That makes me think every vector that is getting embedded is getting transformed somehow (when compared to sentence transformers), but its happening to all of them more or less equally. Is this expected behavior? If so, is there anywhere I can read up on what ollama is doing under the hood? ### OS Linux ### GPU AMD ### CPU AMD ### Ollama version 0.3.3, 0.3.12
GiteaMirror added the bug label 2026-05-04 07:24:49 -05:00
GiteaMirror commented 2026-05-04 07:24:51 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Oct 3, 2024):

I have no input to the rest of your document, but /api/embeddings is deprecated.

<!-- gh-comment-id:2391514065 --> @rick-github commented on GitHub (Oct 3, 2024): I have no input to the rest of your document, but `/api/embeddings` [is deprecated](https://github.com/ollama/ollama/blob/main/docs/api.md#:~:text=Note%3A%20this%20endpoint%20has%20been%20superseded%20by%20/api/embed).
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#66554
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?