Adding vector search #132

programmatix · 2023-12-20T11:42:37Z

No description provided.

programmatix · 2023-12-20T11:44:18Z

This is currently uploaded just for initial review by the SDK PM(s). No review is requested from the wider team at this time.

programmatix · 2024-01-02T15:13:04Z

Review from the wider SDK team is now welcome following initial PM review.

programmatix · 2024-01-05T16:58:28Z

rfc/0052-sdk3-full-text-search.md

+## Vector search
+This is a feature being added to Couchbase Server 7.6 in the FTS service.
+
+The feature will initially be released in a Private Preview, and all SDK additions related to it should be annotated with the platform equivalent of @Stability.Volatile as changes may well be required following user feedback. 


I appreciate the Private Preview may be no more, and that this proposed stability level of the API is now somewhat controversial as a result. Personally I don't feel we should be adding a brand-new API for a potentially volatile feature directly at committed level.

Agree. Feature is definitely not private (it is public but could be be Beta or GA). Regardless as per our discussions, we should still mark APIs as volatile.

programmatix · 2024-01-05T16:59:15Z

rfc/0052-sdk3-full-text-search.md

+
+Important SDK considerations:
+
+* There is some uncertainty on whether it is mandatory to provide a `query` field containing a 'normal' FTS query.  This is being followed up with the FTS team.  As it makes sense to be able to perform solely a vector search, we will assume for now that will be the case.


@jon-strabala can you confirm?

Raised this as a ticket as I've been trying to discover this info since 20th Dec and we now urgently need a response to this.

programmatix · 2024-01-05T17:00:21Z

rfc/0052-sdk3-full-text-search.md

+
+* Performing a vector search in isolation (without a traditional FTS query).
+* A SQL++ statement is used to also perform a vector search.
+  * This is just passed in the string statement and is expected to return the same query results JSON as at present, so from the SDK POV it's just a passthrough with no SDK modifications needed.


@jon-strabala can you confirm that the query results will be the same in this mode?

rfc/0052-sdk3-full-text-search.md

brett19 · 2024-01-06T00:12:13Z

rfc/0052-sdk3-full-text-search.md

+### VectorQuery
+`VectorQuery` creation is platform-idiomatic:
+
+`VectorQuery.create(String vectorFieldName, float[] vector)`


In order to be more specific and support potential integrations with vector libraries, why not VectorQuery.fromVector?

It's mentioned that construction is platform-idiomatic and this is intended to include the "VectorQuery.create(...)" naming. If new VectorQuery(...), or VectorQuery.fromX(...), or some other form, is more idiomatic and better fits the SDK's existing FTS creation methods, it should use such.

For SDKs which use them, is this float64 or float32?

brett19 · 2024-01-06T00:13:53Z

rfc/0052-sdk3-full-text-search.md

+
+The first parameter needs to be mandatory.
+
+`VectorSearch.create(VectorQuery vectorQuery, [VectorSearchOptions options])` can be added to make it easier to send a single `VectorQuery`.


Why does this accept VectorSearchOptions (which only has the operator, which would be unused with a single VectorQuery object)? Relatedly, in the case of multiple queries, should the operator being used between them not be a more high-level object rather than being part of general options? For instance:

VectorSearch.fromMultiple(VectorQueryCombination, List<VectorQuery>) VectorSearch.create(VectorQuery)

Why does this accept VectorSearchOptions (which only has the operator, which would be unused with a single VectorQuery object)

It's forward-looking to when more options will be present in future, but yes, it's not useful right now. Since discovering what a hard time some SDKs have with adding overloads later, I'm trying to anticipate these situations.
I don't have strong feelings on it - can certainly remove it from this optional sugar.

I'm not a big fan of the fromMultiple / VectorQueryCombination idea, I think, for similar reasons - there may well be future options.

rajagp · 2024-01-08T17:14:22Z

rfc/0052-sdk3-full-text-search.md

+SearchRequest request = SearchRequest
+        .searchQuery(SearchQuery.matchAll())
+        .vectorSearch(VectorSearch.create(VectorQuery.create("vector_field", vector)));
+


Are we attached to searchQuery for historical reasons (as in what we have today so less ).It would be ideal to use ftsSearch and ftsQuery to align with vectorSearch/vectorQuery respectively- no?
A new developer reading the API without any of the context would find it hard to understand that searchQuery is actually FTS.

Possibly - Ray and I have certainly wondered before why it's named "search" and not "fullTextSearch" or similar - but it potentially opens a rather large can of worms. .ftsSearch(SearchQuery.matchAll()) or .fullTextSearch(SearchQuery.matchAll()) both look weirdly inconsistent to me - would probably want something instead like .fullTextSearch(FullTextSearch[Query].matchAll()), which looks cleaner. But we have a lot of classes implementing that SearchQuery interface, so that's potentially an expensive addition - some SDKs may end up having to C&P a bunch.

+1 to what @rajagp said.n It would be great to have consistency in function names ftsQuery, geoDistance, vectorQuery, etc.

For the existing SearchQuery including geo I've used the existing names so the large corpus of existing SDK code based on it can be refused. Plus it is somewhat orthogonal to the vector search work, and I didn't want to unduly scope creep given the timescales required.

It's worth noting that this proposal was originally presented for PM review last year, and we are due to begin implementing tomorrow. I can still make small bug fixes but it is too late for feedback that will require such large ranging changes - unless everything is pushed back by a sprint. Let me know if you want to go that path; I'm happy to discuss renaming if given time to do it, though will need details on the desired names from PMs.

chvck · 2024-01-08T18:24:01Z

rfc/0052-sdk3-full-text-search.md

+### VectorSearch
+`VectorSearch` construction is platform-idiomatic:
+
+`VectorSearch.create(List<VectorQuery> vectorQueries, [VectorSearchOptions options])`


Why isn't this also a fluent style API? Where there would be no VectorSearchOptions but myVectorSearch.vectorQueryCombination(...) instead.

Ideally it would not use fluent-style anywhere: VectorQuery + VectorQueryOptions, and VectorSearch + VectorSearchOptions are the ideal, as they are more idiomatic to SDK3.

But, the Java SDK for some reason uses a fluent-style API for SearchQuerys, but an options block for SearchOptions. Don't ask me why! So I've tried to give enough latitude in the RFC for people to be able to implement something that ideally looks like an SDK3 API, but can instead fit the idiomatic style where needed.

PS for the creation of the SearchRequest, fluent-style is recommended. This is "A and/or B [and/or C, D...] major top-level feature" is a new concept to the SDKs AFAIK, and fluent-style seems the best way of handling it.

…ds with the same name

RiPont · 2024-02-21T07:00:22Z

rfc/0052-sdk3-full-text-search.md

+   "search": ["vectorSearch"]
+  }
+```
+If it is not present the SDK will raise `FeatureNotAvailableException` with a message along the lines of "Vector queries are available from Couchbase Server 7.6.0 and above".


Will the clusterCaps included vectorSearch and scopedSearchIndex on 7.6.0+ if there are no search nodes?

RiPont · 2024-02-21T19:25:00Z

rfc/0052-sdk3-full-text-search.md

+If it is not present the SDK will raise `FeatureNotAvailableException` with a message along the lines of "Vector queries are available from Couchbase Server 7.6.0 and above".
+
+The SDK will check for this capability in the most recent config it received.
+If it does not currently have one (in which case one should already be in the process of being asynchronously fetched), the SDK will wait for it to be available.


I feel like waiting for the config adds edge cases and potential performance problems for not much benefit.

If there is no config, chances are nothing is bootstrapped yet and the whole thing will fail anyways.

chvck · 2024-05-21T17:08:55Z

@programmatix could you tidy the commit history a bit please?

Adding vector search

055e679

programmatix mentioned this pull request Dec 20, 2023

Adding vector search #131

Closed

programmatix added 3 commits December 22, 2023 13:55

Major revision, proposing a new search API.

3cb6719

Naming tweaks

1a183b5

Iterating

6f9ebcb

Iterating

e96e176

programmatix commented Jan 5, 2024

View reviewed changes

Detailing SearchRequest

0cb6172

brett19 reviewed Jan 6, 2024

View reviewed changes

programmatix added 2 commits January 7, 2024 17:46

Fixes following feedback

da70c6b

Minor fixes

5ed5846

rajagp reviewed Jan 8, 2024

View reviewed changes

chvck reviewed Jan 8, 2024

View reviewed changes

programmatix added 6 commits January 9, 2024 15:52

Further fixes

d4122fd

Minor fix

4a4f740

Some languages (Java) do not allow having static and non-static metho…

225692e

…ds with the same name

Aligning APIs for scoped index and vector search

f555694

Fixes for aligning APIs for scoped index and vector search

f8d329f

Specify FeatureNotAvailable handling for scope and vector search

c7d89ef

RiPont reviewed Feb 21, 2024

View reviewed changes

chvck approved these changes May 20, 2024

View reviewed changes

DemetrisChr approved these changes May 20, 2024

View reviewed changes

dnault self-requested a review May 21, 2024 19:04

dnault approved these changes May 21, 2024

View reviewed changes

programmatix merged commit ea669e1 into couchbaselabs:master May 22, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adding vector search #132

Adding vector search #132

programmatix commented Dec 20, 2023

programmatix commented Dec 20, 2023

programmatix commented Jan 2, 2024

programmatix Jan 5, 2024

rajagp Jan 8, 2024

programmatix Jan 5, 2024

programmatix Jan 9, 2024 •

edited

programmatix Jan 5, 2024

brett19 Jan 6, 2024

programmatix Jan 7, 2024

chvck Jan 8, 2024

programmatix Jan 8, 2024

brett19 Jan 6, 2024

programmatix Jan 7, 2024

rajagp Jan 8, 2024

programmatix Jan 8, 2024

vishaldhiman22 Jan 9, 2024

programmatix Jan 9, 2024

chvck Jan 8, 2024

programmatix Jan 9, 2024

programmatix Jan 9, 2024 •

edited

RiPont Feb 21, 2024

RiPont Feb 21, 2024 •

edited

chvck commented May 21, 2024


		Important SDK considerations:

		* There is some uncertainty on whether it is mandatory to provide a `query` field containing a 'normal' FTS query. This is being followed up with the FTS team. As it makes sense to be able to perform solely a vector search, we will assume for now that will be the case.


		The first parameter needs to be mandatory.

		`VectorSearch.create(VectorQuery vectorQuery, [VectorSearchOptions options])` can be added to make it easier to send a single `VectorQuery`.

Adding vector search #132

Adding vector search #132

Conversation

programmatix commented Dec 20, 2023

programmatix commented Dec 20, 2023

programmatix commented Jan 2, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

programmatix Jan 9, 2024 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

programmatix Jan 9, 2024 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

RiPont Feb 21, 2024 • edited

Choose a reason for hiding this comment

chvck commented May 21, 2024

programmatix Jan 9, 2024 •

edited

programmatix Jan 9, 2024 •

edited

RiPont Feb 21, 2024 •

edited