github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-05 23:53:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #4128] Normalization of output from embedding model #2565

New Issue
Open
opened 2026-04-12 12:53:30 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-04-12 12:53:30 -05:00
Owner
Copy Link

Originally created by @hagemon on GitHub (May 3, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/4128

When I use Ollama Embedding together with Langchain Retriever's get_relevant_documents, I always get a score that around 200. However, when I use HuggingFaceEmbedding, this value is between 0 and 1.

So I continued to explore the reason and, according to the official documentation, used OllamaEmbedding to vectorize both query and documents. I found that their dot product still exceeds 100:

from langchain_community.embeddings import OllamaEmbeddings
import numpy as np

ollama_emb = OllamaEmbeddings(
    model="mxbai-embed-large:latest",
)
r1 = ollama_emb.embed_documents(
    [
        "Alpha is the first letter of Greek alphabet",
        "Beta is the second letter of Greek alphabet",
    ]
)
r2 = ollama_emb.embed_query(
    "What is the second letter of Greek alphabet"
)
print(np.dot(r1, r2))  
# Output: array([196.91232687, 198.68434774])

Therefore, I assume that they are not normalized.

Some vector databases, such as Milvus, suggest normalizing the vectors before inserting them into the database. So, I wonder if OllamaEmbedding has plans to (or had already) support an option like HuggingFaceEmbedding's encode_kwargs = {"normalize_embeddings": True}, which allows the output vectors to be normalized without the need for me to manually implement this process.

Originally created by @hagemon on GitHub (May 3, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/4128 When I use Ollama Embedding together with Langchain Retriever's `get_relevant_documents`, I always get a score that around 200. However, when I use HuggingFaceEmbedding, this value is between 0 and 1. So I continued to explore the reason and, according to the official documentation, used OllamaEmbedding to vectorize both query and documents. I found that their dot product still exceeds 100: ```python from langchain_community.embeddings import OllamaEmbeddings import numpy as np ollama_emb = OllamaEmbeddings( model="mxbai-embed-large:latest", ) r1 = ollama_emb.embed_documents( [ "Alpha is the first letter of Greek alphabet", "Beta is the second letter of Greek alphabet", ] ) r2 = ollama_emb.embed_query( "What is the second letter of Greek alphabet" ) print(np.dot(r1, r2)) # Output: array([196.91232687, 198.68434774]) ``` Therefore, I assume that they are not normalized. Some vector databases, such as Milvus, suggest normalizing the vectors before inserting them into the database. So, I wonder if OllamaEmbedding has plans to (or had already) support an option like HuggingFaceEmbedding's `encode_kwargs = {"normalize_embeddings": True}`, which allows the output vectors to be normalized without the need for me to manually implement this process.
GiteaMirror added the feature request label 2026-04-12 12:53:30 -05:00
GiteaMirror commented 2026-04-12 12:53:31 -05:00
Author
Owner
Copy Link

@cristoslc commented on GitHub (Jun 30, 2024):

That's an issue for me as well. Trying to use Ollama as a low-effort local embedding server to pair with FileMaker 2024, but the built-in functions in FileMaker require normalized embeddings (as of 2024-06-30).

<!-- gh-comment-id:2198695839 --> @cristoslc commented on GitHub (Jun 30, 2024): That's an issue for me as well. Trying to use Ollama as a low-effort local embedding server to pair with FileMaker 2024, but the built-in functions in FileMaker require normalized embeddings (as of 2024-06-30).
GiteaMirror commented 2026-04-12 12:53:33 -05:00
Author
Owner
Copy Link

@Susensio commented on GitHub (Jul 2, 2024):

I run into the same problem with ollama not normalizing embeddings. I ended up extending OllamaEmbeddings class:

import numpy as np
from langchain_community.embeddings import OllamaEmbeddings

class OllamaEmbeddingsNormalized(OllamaEmbeddings):
    def _process_emb_response(self, input: str) -> list[float]:
        emb = super()._process_emb_response(input)
        return (np.array(emb) / np.linalg.norm(emb)).tolist()

# The rest of your code
ollama_emb = OllamaEmbeddingsNormalized(
    model="mxbai-embed-large:latest",
)
r1 = ollama_emb.embed_documents(
    [
        "Alpha is the first letter of Greek alphabet",
        "Beta is the second letter of Greek alphabet",
    ]
)
r2 = ollama_emb.embed_query(
    "What is the second letter of Greek alphabet"
)
print(np.dot(r1, r2))  
# Output: array([0.67319929 0.75340838])
<!-- gh-comment-id:2203403086 --> @Susensio commented on GitHub (Jul 2, 2024): I run into the same problem with ollama not normalizing embeddings. I ended up extending `OllamaEmbeddings` class: ```python import numpy as np from langchain_community.embeddings import OllamaEmbeddings class OllamaEmbeddingsNormalized(OllamaEmbeddings): def _process_emb_response(self, input: str) -> list[float]: emb = super()._process_emb_response(input) return (np.array(emb) / np.linalg.norm(emb)).tolist() # The rest of your code ollama_emb = OllamaEmbeddingsNormalized( model="mxbai-embed-large:latest", ) r1 = ollama_emb.embed_documents( [ "Alpha is the first letter of Greek alphabet", "Beta is the second letter of Greek alphabet", ] ) r2 = ollama_emb.embed_query( "What is the second letter of Greek alphabet" ) print(np.dot(r1, r2)) # Output: array([0.67319929 0.75340838]) ```
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 feature request
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#2565
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?