[Blog Post] asymmetric model support in neural search by fen-qin · Pull Request #4058 · opensearch-project/project-website

fen-qin · 2026-01-12T21:45:35Z

Description

This PR is for asymmetric model support in neural search blog post.

Issues Resolved

Resolves [BLOG] Asymmetric Models Support in Neural Search #4057

Check List

Commits are signed per the DCO using --signoff

By submitting this pull request, I confirm that my contribution is made under the terms of the BSD-3-Clause License.

github-actions · 2026-01-12T21:45:44Z

Thank you for submitting a blog post!

The blog post review process is: Submit a PR -> (Optional) Peer review -> Doc review -> Editorial review -> Marketing review -> Published.

github-actions · 2026-01-12T21:45:45Z

Hi @fen-qin,

It looks like you're adding a new blog post but don't have an issue mentioned. Please link this PR to an open issue using one of these keywords in the PR description:

Closes #issue-number
Fixes #issue-number
Resolves #issue-number

If an issue hasn't been created yet, please create one and then link it to this PR.

heemin32 · 2026-02-27T21:37:44Z

+
+
+```json
+PUT /_ingest/pipeline/asymmetric_embedding_pipeline


Can we use either semantic text field or text embedding processor instead of ml_inference processor?

Signed-off-by: Fen Qin <mfenqin@amazon.com>

bzhangam · 2026-03-03T19:43:11Z

+```
+cd opensearch-py-ml/docs/source/example/common
+
+## deploy


I think we need to pip install requirement first?

yes. that's true. this is the setup process for host asymmetric model in sagemaker endpoint. I'm thinking if we should only focus on local model for opensearch. so, it had less dependencies with the aws resources

In blog, we don't need to focus on local model. We can use sage maker.

bzhangam · 2026-03-03T19:44:16Z

+Follow these steps to implement asymmetric neural search in your OpenSearch cluster. This example uses a remote SageMaker endpoint, but you can also deploy models locally.
+
+1. Prerequisites: Deploy a sagemaker endpoint
+check out the deployment scripts: https://github.com/opensearch-project/opensearch-py-ml/pull/587


What blocks this PR from being merged? It's a little bit weird that we point people to a PR. Should we get it merged and point people to the README?

no, these PRs will be merged before blog. blog review usually takes longer time. let me connect with ml-commons team to get it merge before the blog

bzhangam · 2026-03-03T19:46:11Z

+    "region": "<YOUR_AWS_REGION>",
+    "service_name": "sagemaker"
+  },
+  "credential": {


Should we call out in the AOS we need to follow this doc https://docs.aws.amazon.com/opensearch-service/latest/developerguide/ml-amazon-connector.html to create the connector?

~~The blog should be about using opensource OpenSearch not AWS OpenSearch.~~

here is the current how local asymmetric model is setup - https://docs.opensearch.org/latest/tutorials/vector-search/semantic-search/semantic-search-asymmetric/

it requires hosting model to local running endpoint.

I think most of users won't setup local model. Maybe, you can share how you set it up in sage maker so that users can reference it.

heemin32 · 2026-03-03T21:38:23Z

+      },
+      "passage_text": {
+        "type": "semantic",
+        "model_id": "<YOUR_MODEL_ID>",


Can we add actual model id you used during testing?

And tell how to get the model id as well?

you mean getModel api to get the model id ?, the model_id will change everytime we create a new model

It does not matter if model_id change or not; they will not use the same model id anyway. It is more about showing readers how to get the model id.

heemin32 · 2026-03-03T21:38:50Z

+OpenSearch returns response:
+
+```json
+{"took":317,"timed_out":false,"_shards":{"total":1,"successful":1,"skipped":0,"failed":0},"hits":{"total":{"value":1,"relation":"eq"},"max_score":0.25255635,"hits":[{"_index":"my-nlp-index","_id":"1","_score":0.25255635,"_source":{"passage_text":"Hello world","id":"s1"}}]}}%     


% is added at the end. Is this intentional?

got it. will remove the %

heemin32 · 2026-03-03T21:39:26Z

+
+## Next steps
+
+- Review the [asymmetric model documentation](https://opensearch.org/docs/latest/tutorials/vector-search/semantic-search/semantic-search-asymmetric/) for detailed configuration options


We might need to update the document before publishing this blog.

Maybe, let's just remove this line. There is no detailed configuration options in the linked page.

heemin32 · 2026-03-03T21:41:11Z

+
+This distinction allows the model to learn specialized representations. For example, the E5 model internally processes "What are some parks in NYC?" as `query: What are some parks in NYC?` during search, while indexing "Central Park is a large public park..." as `passage: Central Park is a large public park...`. This asymmetry helps the model better match short queries to longer documents.
+
+## Why asymmetric models outperform symmetric models


Suggested change

## Why asymmetric models outperform symmetric models

## When asymmetric models outperform symmetric models

heemin32 · 2026-03-03T21:44:42Z

+
+Neural search in OpenSearch has traditionally used symmetric embedding models, where queries and documents are encoded identically. While effective, this approach doesn't reflect how search actually works: queries are typically short and question-like, while documents are longer and information-rich. Asymmetric embedding models address this mismatch by optimizing embeddings differently for queries versus documents, leading to significant improvements in search relevance.
+
+OpenSearch now supports asymmetric embedding models, including state-of-the-art models like E5 that dominate the [MTEB leaderboard](https://huggingface.co/spaces/mteb/leaderboard). In this post, you'll learn how asymmetric models work, see comprehensive benchmark results, and follow a step-by-step guide to implement asymmetric neural search in your OpenSearch cluster.


Suggested change

OpenSearch now supports asymmetric embedding models, including state-of-the-art models like E5 that dominate the [MTEB leaderboard](https://huggingface.co/spaces/mteb/leaderboard). In this post, you'll learn how asymmetric models work, see comprehensive benchmark results, and follow a step-by-step guide to implement asymmetric neural search in your OpenSearch cluster.

Semantic text field now supports asymmetric embedding models, including state-of-the-art models like E5 that dominate the [MTEB leaderboard](https://huggingface.co/spaces/mteb/leaderboard). In this post, you'll learn how asymmetric models work, see comprehensive benchmark results, and follow a step-by-step guide to implement asymmetric neural search in your OpenSearch cluster.

pajuric · 2026-03-05T22:18:08Z

@fen-qin - Please reach out to me when you are ready to move this forward.

fen-qin requested review from AMoo-Miki, CEHENKLE, elfisher, kolchfa-aws, krisfreedain, natebower, nateynateynate, nknize and peterzhuamazon as code owners January 12, 2026 21:45

heemin32 reviewed Feb 27, 2026

View reviewed changes

fen-qin force-pushed the asymmetric_model_support branch from 050d85b to f640500 Compare March 3, 2026 01:12

asymmetric model support in neural search blog

5da7c0b

Signed-off-by: Fen Qin <mfenqin@amazon.com>

fen-qin force-pushed the asymmetric_model_support branch from f640500 to 5da7c0b Compare March 3, 2026 01:17

bzhangam reviewed Mar 3, 2026

View reviewed changes

Comment thread _posts/2026-01-15-asymmetric-model-support-neural-search.md

heemin32 reviewed Mar 3, 2026

View reviewed changes


		## Next steps

		- Review the [asymmetric model documentation](https://opensearch.org/docs/latest/tutorials/vector-search/semantic-search/semantic-search-asymmetric/) for detailed configuration options


		This distinction allows the model to learn specialized representations. For example, the E5 model internally processes "What are some parks in NYC?" as `query: What are some parks in NYC?` during search, while indexing "Central Park is a large public park..." as `passage: Central Park is a large public park...`. This asymmetry helps the model better match short queries to longer documents.

		## Why asymmetric models outperform symmetric models

	## Why asymmetric models outperform symmetric models
	## When asymmetric models outperform symmetric models


		Neural search in OpenSearch has traditionally used symmetric embedding models, where queries and documents are encoded identically. While effective, this approach doesn't reflect how search actually works: queries are typically short and question-like, while documents are longer and information-rich. Asymmetric embedding models address this mismatch by optimizing embeddings differently for queries versus documents, leading to significant improvements in search relevance.

		OpenSearch now supports asymmetric embedding models, including state-of-the-art models like E5 that dominate the [MTEB leaderboard](https://huggingface.co/spaces/mteb/leaderboard). In this post, you'll learn how asymmetric models work, see comprehensive benchmark results, and follow a step-by-step guide to implement asymmetric neural search in your OpenSearch cluster.

	OpenSearch now supports asymmetric embedding models, including state-of-the-art models like E5 that dominate the [MTEB leaderboard](https://huggingface.co/spaces/mteb/leaderboard). In this post, you'll learn how asymmetric models work, see comprehensive benchmark results, and follow a step-by-step guide to implement asymmetric neural search in your OpenSearch cluster.
	Semantic text field now supports asymmetric embedding models, including state-of-the-art models like E5 that dominate the [MTEB leaderboard](https://huggingface.co/spaces/mteb/leaderboard). In this post, you'll learn how asymmetric models work, see comprehensive benchmark results, and follow a step-by-step guide to implement asymmetric neural search in your OpenSearch cluster.



		```json
		PUT /_ingest/pipeline/asymmetric_embedding_pipeline

Conversation

fen-qin commented Jan 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Issues Resolved

Check List

Uh oh!

github-actions Bot commented Jan 12, 2026

Uh oh!

github-actions Bot commented Jan 12, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

heemin32 Mar 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pajuric commented Mar 5, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

fen-qin commented Jan 12, 2026 •

edited

Loading

heemin32 Mar 3, 2026 •

edited

Loading