DOC-5457: add the new SVS-VAMANA vector algorithm type #1822

dwdougherty · 2025-07-08T16:09:57Z

Note to reviewers: there were no applicable changes to the FT.CREATE command page; it refers to the main vector indexing page for all vector-related parameters. This is where the lion's share of the changes were made.

github-actions · 2025-07-08T16:10:15Z

Staging links:
https://redis.io/docs/staging/DOC-5457/develop/ai/search-and-query/indexing/field-and-type-options
https://redis.io/docs/staging/DOC-5457/develop/ai/search-and-query/vectors

mich-elle-luna

thank you! A couple suggestions to remove dups, but maybe this is incorrect? and the file to update for the uni link.

content/develop/ai/search-and-query/vectors.md

@mich-elle-luna

Thank you, @mich-elle-luna! Co-authored-by: mich-elle-luna <[email protected]>

meiravgri

Thanks!

meiravgri · 2025-07-09T04:49:08Z

content/develop/ai/search-and-query/vectors.md

+
+**Required attributes**
+
+| Attribute          | Description                              |


Consider adding "Default value" column. Clearer IMO

meiravgri · 2025-07-09T05:01:14Z

content/develop/ai/search-and-query/vectors.md

+Scalable Vector Search (SVS) is an Intel project in which a new vector search library, VAMANA graph index, was created. SVS-VAMANA supports highly accurate compressed vector indexes. You can read more about the project [here](https://intel.github.io/ScalableVectorSearch/intro.html). Support for `SVS-VAMANA` indexing was added in Redis 8.2.
+
+Choose the `SVS-VAMANA` index type when you need vector search
+
+- on billions of high-dimensional vectors,
+- at high accuracy and state-of-the-art speed,
+- using less memory than alternatives.


We should include a more comprehensive explanation that highlights the differences between HNSW and VAMANA, as well as when to choose each.
I suggest starting with a short description of the vanilla SVS-VAMANA algorithm, then emphasizing that its main advantage is the ability to use it with the COMPRESSION parameter to reduce memory usage.
That’s mentioned here, but I think it could be clearer.
@alonre24 WDYT? Also, can we ask them for a simple guideline table that outlines when to choose each compression type, depending on the data characteristics?

I would say here that you should choose the SVS-VAMANA index when you need a vector search

search performance and scalability are more important than perfect search accuracy (same as HNSW)

using less memory than alternatives.

to be optimized for running on Intel hardware

Regarding the compression type, I would add a clarification in the appropriate section

meiravgri · 2025-07-09T05:03:21Z

content/develop/ai/search-and-query/vectors.md

+| `COMPRESSION`              | Compression algorithm (`LVQ8`, `LVQ8`, `LVQ4`, `LVQ4x4`, `LVQ4x8`, `LeanVec4x8`, or `LeanVec8x8`). Vectors will be compressed during indexing. Note: On non-Intel platforms, `SVS-VAMANA` with `COMPRESSION` will fall back to Intel’s basic quantization implementation. |
+| `CONSTRUCTION_WINDOW_SIZE` | The search window size (the default is 200) to use during graph construction. A higher search window size will yield a higher quality graph since more overall vertexes are considered, but will increase construction time. |


the note should be placed outside the table in big bold letters :)

Note: On non-Intel platforms, SVS-VAMANA with COMPRESSION will fall back to Intel’s basic quantization implementation. |

this note

meiravgri · 2025-07-09T05:03:36Z

content/develop/ai/search-and-query/vectors.md

+| `COMPRESSION`              | Compression algorithm (`LVQ8`, `LVQ8`, `LVQ4`, `LVQ4x4`, `LVQ4x8`, `LeanVec4x8`, or `LeanVec8x8`). Vectors will be compressed during indexing. Note: On non-Intel platforms, `SVS-VAMANA` with `COMPRESSION` will fall back to Intel’s basic quantization implementation. |
+| `CONSTRUCTION_WINDOW_SIZE` | The search window size (the default is 200) to use during graph construction. A higher search window size will yield a higher quality graph since more overall vertexes are considered, but will increase construction time. |


Suggested change

| `COMPRESSION` | Compression algorithm (`LVQ8`, `LVQ8`, `LVQ4`, `LVQ4x4`, `LVQ4x8`, `LeanVec4x8`, or `LeanVec8x8`). Vectors will be compressed during indexing. Note: On non-Intel platforms, `SVS-VAMANA` with `COMPRESSION` will fall back to Intel’s basic quantization implementation. |

| `CONSTRUCTION_WINDOW_SIZE` | The search window size (the default is 200) to use during graph construction. A higher search window size will yield a higher quality graph since more overall vertexes are considered, but will increase construction time. |

| `COMPRESSION` | Compression algorithm (`LVQ8`, `LVQ8`, `LVQ4`, `LVQ4x4`, `LVQ4x8`, `LeanVec4x8`, or `LeanVec8x8`). Vectors will be compressed during indexing. Note: On non-Intel platforms, `SVS-VAMANA` with `COMPRESSION` will fall back to Intel’s basic scalar quantization implementation. |

| `CONSTRUCTION_WINDOW_SIZE` | The search window size (the default is 200) to use during graph construction. A higher search window size will yield a higher quality graph since more overall vertexes are considered, but will increase construction time. |

meiravgri · 2025-07-09T05:07:12Z

content/develop/ai/search-and-query/indexing/field-and-type-options.md

@@ -157,6 +157,7 @@ Where:

    - `FLAT`: brute force algorithm.
    - `HNSW`: hierarchical, navigable, small world algorithm.
+    - `SVS-VAMANA`: a graph-based nearest neighbor search algorithm.


Suggested change

- `SVS-VAMANA`: a graph-based nearest neighbor search algorithm.

- `SVS-VAMANA`: a graph-based nearest neighbor search algorithm, optimized for use with compression methods to reduce memory footprint.

meiravgri · 2025-07-09T05:11:59Z

content/develop/ai/search-and-query/vectors.md

+|:----------|:------------|:--------------|
+| `SEARCH_WINDOW_SIZE` | The size of the search window (applies only to KNN searches). | 10 or the value that was passed upon index creation. |
+| `GRAPH_MAX_DEGREE` | The maximum node degree in the graph. | 32 or the value that was passed upon index creation. |
+| `SEARCH_WINDOW_SIZE` | The size of the search window. | 10 or the value that was passed upon index creation. |


dup SEARCH_WINDOW_SIZE?

and missing USE_SEARCH_HISTORY

content/develop/ai/search-and-query/vectors.md

meiravgri · 2025-07-09T05:15:54Z

content/develop/ai/search-and-query/vectors.md

+| `SEARCH_WINDOW_SIZE` | The size of the search window (applies only to KNN searches). | 10 or the value that was passed upon index creation. |
+| `GRAPH_MAX_DEGREE` | The maximum node degree in the graph. | 32 or the value that was passed upon index creation. |
+| `SEARCH_WINDOW_SIZE` | The size of the search window. | 10 or the value that was passed upon index creation. |
+| `EPSILON` | The range search approximation factor. | 0.01 or the value that was passed upon index creation. |


GRAPH_MAX_DEGREE is not a runtime param

alonre24

few comments

alonre24 · 2025-07-09T07:10:34Z

content/develop/ai/search-and-query/vectors.md

+Scalable Vector Search (SVS) is an Intel project in which a new vector search library, VAMANA graph index, was created. SVS-VAMANA supports highly accurate compressed vector indexes. You can read more about the project [here](https://intel.github.io/ScalableVectorSearch/intro.html). Support for `SVS-VAMANA` indexing was added in Redis 8.2.
+
+Choose the `SVS-VAMANA` index type when you need vector search
+
+- on billions of high-dimensional vectors,
+- at high accuracy and state-of-the-art speed,
+- using less memory than alternatives.


I would say here that you should choose the SVS-VAMANA index when you need a vector search

search performance and scalability are more important than perfect search accuracy (same as HNSW)

using less memory than alternatives.

to be optimized for running on Intel hardware

Regarding the compression type, I would add a clarification in the appropriate section

alonre24 · 2025-07-09T07:25:07Z

content/develop/ai/search-and-query/vectors.md

+
+| Attribute                  | Description                              |
+|:---------------------------|:-----------------------------------------|
+| `COMPRESSION`              | Compression algorithm (`LVQ8`, `LVQ8`, `LVQ4`, `LVQ4x4`, `LVQ4x8`, `LeanVec4x8`, or `LeanVec8x8`). Vectors will be compressed during indexing. Note: On non-Intel platforms, `SVS-VAMANA` with `COMPRESSION` will fall back to Intel’s basic quantization implementation. |


typo - LVQ8 is here twice.

I would suggest reffering here - https://intel.github.io/ScalableVectorSearch/performance/using_compression.html to help choosing the appropriate compression type (+ we will ask Intel to write here about LeanVec as well, which currently available here https://intel.github.io/ScalableVectorSearch/python/experimental/leanvec.html)

"SVS-VAMANA with COMPRESSION will fall back to Intel’s basic quantization implementation." - I would rephrase to SVS-VAMANA with COMPRESSION will fall back to SVS implementation of global scalar quantisation using 8 bits.

content/develop/ai/search-and-query/vectors.md

alonre24 · 2025-07-09T07:26:40Z

content/develop/ai/search-and-query/vectors.md

+|:----------|:------------|:--------------|
+| `SEARCH_WINDOW_SIZE` | The size of the search window (applies only to KNN searches). | 10 or the value that was passed upon index creation. |
+| `GRAPH_MAX_DEGREE` | The maximum node degree in the graph. | 32 or the value that was passed upon index creation. |
+| `SEARCH_WINDOW_SIZE` | The size of the search window. | 10 or the value that was passed upon index creation. |


and missing USE_SEARCH_HISTORY

dwdougherty requested review from alonre24, meiravgri and a team July 8, 2025 16:09

dwdougherty self-assigned this Jul 8, 2025

dwdougherty added dev ros 8.2 labels Jul 8, 2025

DOC-5457: add the new SVS-VAMANA vector algorithm type

2440a42

dwdougherty force-pushed the DOC-5457 branch from 892cdde to 2440a42 Compare July 8, 2025 17:15

mich-elle-luna approved these changes Jul 8, 2025

View reviewed changes

dwdougherty and others added 3 commits July 8, 2025 11:12

Apply suggestions from code review

c91ec3f

Thank you, @mich-elle-luna! Co-authored-by: mich-elle-luna <[email protected]>

Apply more suggestions from code review

3c32f10

Apply misspelling fix

88fe66d

dwdougherty added the do not merge yet label Jul 8, 2025

meiravgri reviewed Jul 9, 2025

View reviewed changes

alonre24 reviewed Jul 9, 2025

View reviewed changes

		\| `COMPRESSION` \| Compression algorithm (`LVQ8`, `LVQ8`, `LVQ4`, `LVQ4x4`, `LVQ4x8`, `LeanVec4x8`, or `LeanVec8x8`). Vectors will be compressed during indexing. Note: On non-Intel platforms, `SVS-VAMANA` with `COMPRESSION` will fall back to Intel’s basic quantization implementation. \|
		\| `CONSTRUCTION_WINDOW_SIZE` \| The search window size (the default is 200) to use during graph construction. A higher search window size will yield a higher quality graph since more overall vertexes are considered, but will increase construction time. \|

	- `SVS-VAMANA`: a graph-based nearest neighbor search algorithm.
	- `SVS-VAMANA`: a graph-based nearest neighbor search algorithm, optimized for use with compression methods to reduce memory footprint.

DOC-5457: add the new SVS-VAMANA vector algorithm type #1822

Are you sure you want to change the base?

DOC-5457: add the new SVS-VAMANA vector algorithm type #1822

Conversation

dwdougherty commented Jul 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Jul 8, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

mich-elle-luna left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

meiravgri left a comment

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!

meiravgri Jul 9, 2025 • 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!

meiravgri Jul 9, 2025 • 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!

Uh oh!

meiravgri Jul 9, 2025 • 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!

alonre24 left a comment

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!

Uh oh!

dwdougherty commented Jul 8, 2025 •

edited

Loading

github-actions bot commented Jul 8, 2025 •

edited

Loading

meiravgri Jul 9, 2025 •

edited

Loading

meiravgri Jul 9, 2025 •

edited

Loading

meiravgri Jul 9, 2025 •

edited

Loading