feat: support semantic search in AI chat and embedding ability by hgaol · Pull Request #1510 · apache/answer

hgaol · 2026-03-02T10:43:55Z

Implementation of #1468 , add a semantic search tool to keep user experience consistent.

Settings

Calling semantic search tool

logs

LinkinStars · 2026-03-05T12:30:08Z

internal/repo/embedding/embedding_repo.go

+}
+
+// SearchSimilar performs brute-force cosine similarity search in Go.
+func (r *embeddingRepo) SearchSimilar(ctx context.Context, queryVector []float32, topK int) ([]SimilarResult, error) {


The biggest issue with this approach lies here: all data queries are performed in-memory. While this works for small datasets, it is undoubtedly unacceptable for large-scale data.

Therefore, my suggestion is that if users are using PostgreSQL, they could directly utilize PostgreSQL + pgvector to store vector data and perform searches within the database.

Furthermore, a more 'ideal' approach would be to expose this component as a plugin, allowing user Q&A data to be synchronized to external systems. This design is similar to how search plugins operate. For instance, an Elasticsearch (ES) plugin synchronizes Q&A data to ES for retrieval. The benefit of a plugin-based implementation is extensibility: users aren't restricted to the built-in database and can use their own custom vector databases. The required interfaces would likely mirror those of a search plugin, such as a data synchronization interface and a search interface.

What do you think?

First of all, thank you @LinkinStars for the careful review! The goal of this PR is also to make it clear, so feel free to raise your concern and suggestions!

The biggest issue with this approach lies here: all data queries are performed in-memory. While this works for small datasets, it is undoubtedly unacceptable for large-scale data.

Therefore, my suggestion is that if users are using PostgreSQL, they could directly utilize PostgreSQL + pgvector to store vector data and perform searches within the database.

It makes sense. It's better to save vectors in vector DB. And I think it could support multiple vector DBs, while still support saving in main database or memory for test purpose. WDYT?

Furthermore, a more 'ideal' approach would be to expose this component as a plugin, allowing user Q&A data to be synchronized to external systems. This design is similar to how search plugins operate. For instance, an Elasticsearch (ES) plugin synchronizes Q&A data to ES for retrieval. The benefit of a plugin-based implementation is extensibility: users aren't restricted to the built-in database and can use their own custom vector databases. The required interfaces would likely mirror those of a search plugin, such as a data synchronization interface and a search interface.

About plugin, do you mean make all the functionalities as a plugin, or just the storage for vectors? Per your statement, I assume the later one?

And I think it could support multiple vector DBs
About plugin, do you mean make all the functionalities as a plugin, or just the storage for vectors? Per your statement, I assume the later one?

Yes, if implemented via plugins, it can support different vector databases. I think we can omit the in-memory implementation since it won't be used in most cases.

You're right—it's just a matter of providing vector storage functionality as a plugin. Of course, vector search capabilities are also included.

Agreed, I love this idea! I'll update the PR per our discussions.

hgaol · 2026-03-20T07:26:55Z

Sorry for late update and I'm kind of busy recently, will start this feature at earliest convenience! Most likely I can start it in next week! Actually it's already WIP.

LinkinStars · 2026-03-20T07:32:53Z

Sorry for late update and I'm kind of busy recently, will start this feature at earliest convenience! Most likely I can start it in next week! Actually it's already WIP.

@hgaol Don't worry. We have other features in the works as well. So take your time. ♥️

…mbeddings - Added a new vector search syncer to aggregate questions and answers with comments for vector embedding. - Introduced a new VectorSearch interface and related structures for managing vector storage and similarity search. - Refactored embedding service to delegate semantic search to the new vector search plugin. - Removed embedding-related fields from SiteAIProvider and UI forms as part of the transition to the new vector search architecture. - Updated plugin registration to include vector search capabilities. - Cleaned up embedding service methods and removed unused dependencies.

hgaol · 2026-04-12T06:08:09Z

Updated to involve new vector search plugin. See demo below.

vector-search-plugin.mp4

Here's summary of the design and implementation.

Vector Search & Semantic Search Design

Architecture Layers

┌──────────────────────────────────────────────┐
│  AI Chat / MCP Tool ("semantic_search")      │  ← Controller layer
├──────────────────────────────────────────────┤
│  EmbeddingService                            │  ← Service layer (thin facade)
├──────────────────────────────────────────────┤
│  plugin.VectorSearch interface               │  ← Plugin abstraction
├──────────────────────────────────────────────┤
│  pgvector / elasticsearch / weaviate / ...   │  ← Plugin implementations
└──────────────────────────────────────────────┘

Plugin Interface (`plugin/vector_search.go`)

RegisterSyncer(ctx, syncer) — core provides a syncer for bulk data pull
SearchSimilar(ctx, query, topK) — returns []VectorSearchResult{ObjectID, ObjectType, Metadata, Score}
UpdateContent(ctx, content) — upserts a document with embedding
DeleteContent(ctx, objectID) — removes a document
ConfigReceiver(config) / ConfigFields() — plugin config lifecycle

GenerateEmbedding() is the shared embedding utility used by plugins.

Content Syncing (`vector_search_sync/syncer.go`)

Core implements VectorSearchSyncer with:

GetQuestionsPage(page, pageSize)
GetAnswersPage(page, pageSize)

Each indexed document aggregates question/answer/comment text. Metadata stores deshortened IDs for reconstruction at query time.

Sync is triggered from RegisterSyncer() (startup + config update flow).

Startup & Activation Flow

initPluginData():
  1. Load plugin status from DB
  2. Call ConfigReceiver for configured plugins
     -> parse config always
     -> if active: run heavy init (probe embedding + connect/schema checks)
     -> if inactive: skip heavy init (IsEnabled guard)
  3. Call RegisterSyncer for vector search plugins
     -> if active/initialized: trigger full sync
     -> if inactive/uninitialized: skip sync

On admin config save:

ConfigReceiver
UpdatePluginConfig -> RegisterSyncer -> full sync

Current Behavior Summary

Active plugin on startup: does one probe embedding call, then full sync to vector storage.
Inactive plugin on startup: parses config only; no probe embedding and no sync.
Config save for active plugin: re-runs init path and full sync.

Semantic Search Query Flow

User query -> MCP tool "semantic_search"
  -> EmbeddingService.SearchSimilar(query, topK)
  -> plugin.SearchSimilar() returns scored IDs + metadata
  -> handler fetches full DB content (question/answers/comments)
  -> returns structured semantic search response

hgaol · 2026-04-12T06:09:07Z

There're also some follow ups I don't have clear answer for now. Post here for discussion.

Follow-up: When to calc the embeddings and sync to vector storage

Comparison with Search Plugin

Aspect	Search Plugin	VectorSearch Plugin
Bulk sync	Yes	Yes
Real-time sync	Yes (create/update/delete hooks)	No
Trigger	Event-driven + startup/config	Startup/config only
Consistency	Near real-time	Eventually consistent

Current Gap

UpdateContent() / DeleteContent() exist in plugin.VectorSearch, but are not called from question/answer service events. So after initial sync, content changes are not reflected until next full re-sync.

Options

Manual (current)
- Re-sync only on plugin config save/update
- Simple, but stale results between syncs
Real-time
- Add event hooks to call vector search update/delete
- Can be async (goroutine / queue) to avoid write-path latency
- Higher embedding API call volume
Scheduled (cron)
- Periodic bulk sync via cron expression
- Good for off-peak syncing
- Delayed freshness until next run

hgaol · 2026-04-12T06:12:34Z

I'll create another PR in plugin repo later for memory vector search plugin. I feel that for test purpose, memory has less settings compared to database option. I also implemented some other 3rd party vector store plugins like weaviate etc, but some of them haven't validated yet. I'll submit PR for them once done the validations.

hgaol · 2026-04-12T09:47:49Z

Here's the PR for memory vector search plugin. apache/answer-plugins#315

feat: support semantic search in AI chat and embedding ability

edb87cd

hgaol mentioned this pull request Mar 2, 2026

Add copilot chat plugin to leverage AI to answer user queries based on questions and answers #1468

Open

hgaol added 2 commits March 2, 2026 19:02

fix lint issue

851c778

update init tables

899638b

LinkinStars self-requested a review March 5, 2026 12:22

LinkinStars requested changes Mar 5, 2026

View reviewed changes

LinkinStars mentioned this pull request Mar 10, 2026

Webhooks #1509

Closed

hgaol added 3 commits April 12, 2026 13:22

fix lint

3c1aa87

fix lint

ebd7c7f

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: support semantic search in AI chat and embedding ability#1510

feat: support semantic search in AI chat and embedding ability#1510
hgaol wants to merge 6 commits intoapache:devfrom
hgaol:1468

hgaol commented Mar 2, 2026

Uh oh!

LinkinStars Mar 5, 2026

Uh oh!

hgaol Mar 5, 2026

Uh oh!

LinkinStars Mar 6, 2026

Uh oh!

hgaol Mar 6, 2026

Uh oh!

hgaol commented Mar 20, 2026

Uh oh!

LinkinStars commented Mar 20, 2026

Uh oh!

hgaol commented Apr 12, 2026

Uh oh!

hgaol commented Apr 12, 2026

Uh oh!

hgaol commented Apr 12, 2026

Uh oh!

hgaol commented Apr 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

hgaol commented Mar 2, 2026

Uh oh!

LinkinStars Mar 5, 2026

Choose a reason for hiding this comment

Uh oh!

hgaol Mar 5, 2026

Choose a reason for hiding this comment

Uh oh!

LinkinStars Mar 6, 2026

Choose a reason for hiding this comment

Uh oh!

hgaol Mar 6, 2026

Choose a reason for hiding this comment

Uh oh!

hgaol commented Mar 20, 2026

Uh oh!

LinkinStars commented Mar 20, 2026

Uh oh!

hgaol commented Apr 12, 2026

Vector Search & Semantic Search Design

Architecture Layers

Plugin Interface (plugin/vector_search.go)

Content Syncing (vector_search_sync/syncer.go)

Startup & Activation Flow

Current Behavior Summary

Semantic Search Query Flow

Uh oh!

hgaol commented Apr 12, 2026

Follow-up: When to calc the embeddings and sync to vector storage

Comparison with Search Plugin

Current Gap

Options

Uh oh!

hgaol commented Apr 12, 2026

Uh oh!

hgaol commented Apr 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Plugin Interface (`plugin/vector_search.go`)

Content Syncing (`vector_search_sync/syncer.go`)