Open
Conversation
📝 WalkthroughWalkthroughThis pull request introduces an event feed system with post-event tag associations, background similarity processing, and corresponding API endpoints. The implementation includes database schema, repository methods, service layers for API and background processing, API endpoints, and comprehensive integration tests covering all workflows. Changes
Sequence DiagramssequenceDiagram
actor User
participant API as EventController
participant EventService
participant EventPostRepo as EventPostRepository
participant EventTagRepo as EventTagRepository
participant PostRepo as PostRepository
participant DB as Database
User->>API: GET /event/{eventTagId}/posts?page=1&limit=10&source=user
API->>EventService: getEventPosts(user, eventTagId, 1, 10, source)
EventService->>EventTagRepo: findOne(eventTagId)
EventTagRepo->>DB: Query EventTag
DB-->>EventTagRepo: EventTag | null
EventTagRepo-->>EventService: EventTag or throw NotFound
par Parallel Fetch
EventService->>EventPostRepo: getPostsForEvent(eventTagId, source, skip, limit)
EventPostRepo->>DB: Query posts with joins (user, categories, eventTags)
DB-->>EventPostRepo: EventPostModel[] with relations
EventPostRepo-->>EventService: EventPostModel[]
and
EventService->>EventPostRepo: getPostCountForEvent(eventTagId, source)
EventPostRepo->>DB: COUNT relationships
DB-->>EventPostRepo: total count
EventPostRepo-->>EventService: number
end
EventService->>PostRepo: getBlockedUserIds(user)
PostRepo->>DB: Query UserBlock
DB-->>PostRepo: blocked IDs
PostRepo-->>EventService: string[]
EventService->>EventService: Filter inactive & blocked users
EventService->>EventService: Enrich posts with source & relevanceScore
EventService-->>API: GetEventPostsResponse
API-->>User: { posts[], total, page, limit }
sequenceDiagram
participant Cron as EventSimilarityCron
participant EventPostService
participant EventTagRepo as EventTagRepository
participant PostRepo as PostRepository
participant EventPostRepo as EventPostRepository
participant DB as Database
Cron->>Cron: 0 * * * * (hourly trigger)
Cron->>EventTagRepo: getAllEventTags()
EventTagRepo->>DB: SELECT * FROM event_tags ORDER BY name
DB-->>EventTagRepo: EventTagModel[]
EventTagRepo-->>Cron: tags
loop For each EventTag
Cron->>EventPostService: processEventSimilarity(tagId)
EventPostService->>EventPostRepo: getUserTaggedPostsForEvent(tagId)
EventPostRepo->>DB: Query user-sourced eventPosts
DB-->>EventPostRepo: EventPostModel[]
EventPostRepo-->>EventPostService: relationships
Note over EventPostService: Exit if no user posts or no embeddings
EventPostService->>EventPostService: computeCentroid(embeddings[])
EventPostService->>PostRepo: findSimilarPostsForEvent(centroid, threshold, limit)
PostRepo->>DB: pgvector similarity search (1 - (embedding <=> centroid))
DB-->>PostRepo: { post, score }[]
PostRepo-->>EventPostService: similar posts with scores
EventPostService->>EventPostRepo: deleteRelationshipsBySourceForEvent(tagId, "similarity")
EventPostRepo->>DB: DELETE eventPosts WHERE source='similarity'
DB-->>EventPostRepo: ✓
loop For each similar post
EventPostService->>EventPostRepo: upsertRelationship(postId, tagId, "similarity", score)
EventPostRepo->>DB: INSERT/UPDATE eventPost
DB-->>EventPostRepo: EventPostModel
end
EventPostService-->>Cron: ✓ Processed
end
Cron->>Cron: Log completion
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Poem
🚥 Pre-merge checks | ✅ 1 | ❌ 2❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
🧹 Nitpick comments (7)
src/app.ts (1)
29-29: Consider single-run protection for event similarity cron startup.Line 203 starts the cron unconditionally on each app instance. In multi-replica deployments, this can trigger duplicate hourly jobs and unnecessary DB load. Consider gating with an env flag or leader/lock-based scheduling.
Also applies to: 203-203
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/app.ts` at line 29, The cron is being started unconditionally via startEventSimilarityCron which causes duplicate jobs in multi-replica deployments; modify the startup to guard the call to startEventSimilarityCron by either (A) honoring an environment flag (e.g., ENABLE_EVENT_SIMILARITY_CRON) and only calling startEventSimilarityCron when that flag is true, or (B) acquiring a leader/distributed lock (Redis/DB advisory lock) at startup and only calling startEventSimilarityCron if the lock is obtained; update startup logic around startEventSimilarityCron to check the chosen guard and log whether the instance is running the cron or skipping it.src/repositories/EventPostRepository.ts (2)
41-48: Hardcoded string'user'should use theEventPostSourcetype.Line 46 uses a hardcoded string
'user'instead of referencing a constant. While functionally correct, using a typed constant would be more maintainable.♻️ Suggested fix
- .andWhere("epr.source = 'user'") + .andWhere("epr.source = :source", { source: 'user' as EventPostSource })🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/repositories/EventPostRepository.ts` around lines 41 - 48, The query in getUserTaggedPostsForEvent uses the hardcoded string 'user'; replace it with the typed constant from EventPostSource to improve maintainability: import EventPostSource and change the .andWhere clause to use a parameterized value (e.g. .andWhere("epr.source = :source", { source: EventPostSource.User })) so the code references EventPostSource instead of the literal string.
79-84: Remove redundant type annotations.ESLint correctly flags that
skip: number = 0andlimit: number = 10have types that can be trivially inferred from the default values.♻️ Proposed fix
public async getPostsForEvent( eventTagId: string, source?: EventPostSource, - skip: number = 0, - limit: number = 10, + skip = 0, + limit = 10, ): Promise<EventPostModel[]> {🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/repositories/EventPostRepository.ts` around lines 79 - 84, The method signature for getPostsForEvent in EventPostRepository.ts has redundant type annotations on parameters skip: number = 0 and limit: number = 10; remove the explicit ": number" annotations so the types are inferred from the default values (i.e., change to skip = 0, limit = 10) while keeping the parameter names, defaults, return type Promise<EventPostModel[]>, and the rest of the method intact; locate the signature in the getPostsForEvent method and update it accordingly.src/services/EventPostService.ts (3)
87-98: Missing dimension validation incomputeCentroid.If embeddings have mismatched dimensions, the loop at line 92-94 will silently produce incorrect results (shorter embeddings won't contribute to higher indices). While the caller filters to posts with embeddings, there's no guarantee they all have the same dimension.
🛡️ Proposed defensive check
private computeCentroid(embeddings: number[][]): number[] { const dim = embeddings[0].length; + if (embeddings.some(e => e.length !== dim)) { + throw new Error(`Embedding dimension mismatch: expected ${dim}`); + } const centroid = new Array(dim).fill(0);🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/services/EventPostService.ts` around lines 87 - 98, computeCentroid lacks validation for embedding dimensionality which allows mismatched-length arrays to silently produce wrong results; update computeCentroid to first assert embeddings is non-empty, capture const dim = embeddings[0].length, then iterate embeddings and throw a descriptive Error (or return a safe failure) if any embedding.length !== dim (include the offending index and length in the message), only then proceed to sum and average so all vectors contribute correctly.
24-26: Test environment skip pattern may hide integration bugs.Skipping the entire method when
NODE_ENV === "test"prevents unit testing of this logic. The test file works around this by temporarily overridingNODE_ENV, but this pattern is fragile.Consider injecting a configuration flag or making the skip behavior configurable to allow proper testing without environment manipulation.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/services/EventPostService.ts` around lines 24 - 26, The method processEventSimilarity in EventPostService currently returns early based on process.env.NODE_ENV === "test"; change this to use an injectable/configurable flag so tests can control the behavior without mutating NODE_ENV: add a constructor parameter (e.g., skipEventSimilarityInTest: boolean or a ConfigService boolean like config.skipEventSimilarity) stored on the EventPostService instance (or read from ConfigService), defaulting to the existing NODE_ENV check for backwards compatibility, then replace the direct process.env check in processEventSimilarity with a check against that injected flag and update callers/tests to pass the desired value.
71-78: Sequential upserts could be batched for performance.The loop issues individual
upsertRelationshipcalls sequentially. For large result sets (up toMAX_RESULTS = 100), this creates many round trips. Consider batching these operations.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/services/EventPostService.ts` around lines 71 - 78, The loop in EventPostService that iterates over similarPosts and calls eventPostRepo.upsertRelationship sequentially (up to MAX_RESULTS) causes many round trips; change this to batch the operations by adding a repository method (e.g., eventPostRepo.upsertRelationshipsBulk or upsertRelationshipBatch) that accepts an array of {postId, eventTagId, type, score} and performs a single bulk upsert or transaction, or if DB/binding lacks bulk support, run the existing upserts in parallel with controlled concurrency (Promise.all with a p-limit) to avoid sequential awaits; update the code in the similarPosts handling block to prepare the array of relationships from similarPosts and call the new bulk/parallel mechanism instead of awaiting each eventPostRepo.upsertRelationship inside the for loop.src/tests/EventTest.test.ts (1)
479-481: Type assertion suggests a type mismatch.The
source: "similarity" as anycast indicates TypeORM'sfindoptions may not properly recognize theEventPostSourceliteral type. This works but is a code smell.Consider defining a type-safe query interface or using query builder for better type safety.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/tests/EventTest.test.ts` around lines 479 - 481, The test is using a loose cast "similarity" as any when calling conn.manager.find for EventPostModel; replace the unsafe any cast with a type-safe value by using the actual EventPostSource literal/enum or a proper typed query: ensure the string "similarity" is typed as EventPostSource (or import/construct the enum/union used by EventPostModel) when calling conn.manager.find (the simRows lookup), or switch to conn.createQueryBuilder/EventPostModel with an explicit where clause to avoid the type assertion.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/repositories/PostRepository.ts`:
- Around line 555-572: In findSimilarPostsForEvent add a guard to exclude empty
embeddings before casting by ensuring array_length(post.embedding,1) = 512 (or
>0 / =512 as appropriate) in the WHERE clauses that use CAST(... AS vector(512))
so empty arrays ([]) are filtered out; update the scoreExpr usage and the
.andWhere chain in the method to include this array_length check alongside
"post.embedding IS NOT NULL". Apply the same array_length guard to the other
vector-distance queries: PostRepository.findSimilarPosts (both places it
builds/uses the vector cast) and the second ordering clause in
PostRepository.findSimilarPostsForEvent, and to
RequestRepository.findSimilarRequests so pgvector never receives an empty array
and avoids the dimension mismatch error.
In `@src/services/EventService.ts`:
- Around line 41-68: The current code applies active/blocked filtering after
calling eventPostRepo.getPostsForEvent and getPostCountForEvent, which causes
pagination and total to include rows that should be hidden; modify the repo
calls so filtering happens in the DB layer: retrieve blocked UIDs from
userRepo.getUserWithBlockedInfo(user.firebaseUid) first, then call
eventPostRepo.getPostsForEvent(eventTagId, source, skip, limit, {excludeUids:
blockedUids, requireActiveUser: true}) and
eventPostRepo.getPostCountForEvent(eventTagId, source, {excludeUids:
blockedUids, requireActiveUser: true}) (or add equivalent parameters/overloads)
so both the SELECT and COUNT apply the same predicates; remove the post-level
isActive/blocked filtering in EventService and keep only mapping to
PostWithSource for attaching source/relevanceScore.
In `@src/services/PostService.ts`:
- Around line 553-562: When removing the manual event tags in PostService (use
removeEventTagsRequest.eventTags, removedTags,
eventPostRepository.deleteRelationship and post.eventTags), ensure you also
clear any leftover similarity rows when the last user-tagged seed for that event
is gone: after deleting each (post.id, tag.id) and updating post.eventTags,
check if there are no remaining user-tagged posts or embeddings for the event
and then call the repository method that removes similarity rows
(eventPostRepository.deleteRelationshipsBySourceForEvent or the EventPostService
equivalent) for that event/source before saving the post; this mirrors the
cleanup logic found in EventPostService.ts (deleteRelationshipsBySourceForEvent)
to prevent stale event matches.
---
Nitpick comments:
In `@src/app.ts`:
- Line 29: The cron is being started unconditionally via
startEventSimilarityCron which causes duplicate jobs in multi-replica
deployments; modify the startup to guard the call to startEventSimilarityCron by
either (A) honoring an environment flag (e.g., ENABLE_EVENT_SIMILARITY_CRON) and
only calling startEventSimilarityCron when that flag is true, or (B) acquiring a
leader/distributed lock (Redis/DB advisory lock) at startup and only calling
startEventSimilarityCron if the lock is obtained; update startup logic around
startEventSimilarityCron to check the chosen guard and log whether the instance
is running the cron or skipping it.
In `@src/repositories/EventPostRepository.ts`:
- Around line 41-48: The query in getUserTaggedPostsForEvent uses the hardcoded
string 'user'; replace it with the typed constant from EventPostSource to
improve maintainability: import EventPostSource and change the .andWhere clause
to use a parameterized value (e.g. .andWhere("epr.source = :source", { source:
EventPostSource.User })) so the code references EventPostSource instead of the
literal string.
- Around line 79-84: The method signature for getPostsForEvent in
EventPostRepository.ts has redundant type annotations on parameters skip: number
= 0 and limit: number = 10; remove the explicit ": number" annotations so the
types are inferred from the default values (i.e., change to skip = 0, limit =
10) while keeping the parameter names, defaults, return type
Promise<EventPostModel[]>, and the rest of the method intact; locate the
signature in the getPostsForEvent method and update it accordingly.
In `@src/services/EventPostService.ts`:
- Around line 87-98: computeCentroid lacks validation for embedding
dimensionality which allows mismatched-length arrays to silently produce wrong
results; update computeCentroid to first assert embeddings is non-empty, capture
const dim = embeddings[0].length, then iterate embeddings and throw a
descriptive Error (or return a safe failure) if any embedding.length !== dim
(include the offending index and length in the message), only then proceed to
sum and average so all vectors contribute correctly.
- Around line 24-26: The method processEventSimilarity in EventPostService
currently returns early based on process.env.NODE_ENV === "test"; change this to
use an injectable/configurable flag so tests can control the behavior without
mutating NODE_ENV: add a constructor parameter (e.g., skipEventSimilarityInTest:
boolean or a ConfigService boolean like config.skipEventSimilarity) stored on
the EventPostService instance (or read from ConfigService), defaulting to the
existing NODE_ENV check for backwards compatibility, then replace the direct
process.env check in processEventSimilarity with a check against that injected
flag and update callers/tests to pass the desired value.
- Around line 71-78: The loop in EventPostService that iterates over
similarPosts and calls eventPostRepo.upsertRelationship sequentially (up to
MAX_RESULTS) causes many round trips; change this to batch the operations by
adding a repository method (e.g., eventPostRepo.upsertRelationshipsBulk or
upsertRelationshipBatch) that accepts an array of {postId, eventTagId, type,
score} and performs a single bulk upsert or transaction, or if DB/binding lacks
bulk support, run the existing upserts in parallel with controlled concurrency
(Promise.all with a p-limit) to avoid sequential awaits; update the code in the
similarPosts handling block to prepare the array of relationships from
similarPosts and call the new bulk/parallel mechanism instead of awaiting each
eventPostRepo.upsertRelationship inside the for loop.
In `@src/tests/EventTest.test.ts`:
- Around line 479-481: The test is using a loose cast "similarity" as any when
calling conn.manager.find for EventPostModel; replace the unsafe any cast with a
type-safe value by using the actual EventPostSource literal/enum or a proper
typed query: ensure the string "similarity" is typed as EventPostSource (or
import/construct the enum/union used by EventPostModel) when calling
conn.manager.find (the simRows lookup), or switch to
conn.createQueryBuilder/EventPostModel with an explicit where clause to avoid
the type assertion.
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 08335a9d-5164-42c9-bf02-4aacc68f66fc
📒 Files selected for processing (22)
.gitignorejest.config.jssrc/api/controllers/EventController.tssrc/api/controllers/index.tssrc/app.tssrc/cron/eventSimilarityCron.tssrc/migrations/1769700000000-CreateEventPost.tssrc/models/EventPostModel.tssrc/models/EventTagModel.tssrc/models/index.tssrc/repositories/EventPostRepository.tssrc/repositories/EventTagRepository.tssrc/repositories/PostRepository.tssrc/repositories/index.tssrc/services/EventPostService.tssrc/services/EventService.tssrc/services/PostService.tssrc/tests/EventTest.test.tssrc/tests/controllers/ControllerFactory.tssrc/tests/data/DatabaseConnection.tssrc/types/ApiResponses.tsswagger.json
Comment on lines
+555
to
+572
| public async findSimilarPostsForEvent( | ||
| centroidEmbedding: number[], | ||
| excludePostIds: string[], | ||
| threshold: number, | ||
| limit: number = 100, | ||
| ): Promise<{ post: PostModel; score: number }[]> { | ||
| const lit = `[${centroidEmbedding.join(",")}]`; | ||
| const scoreExpr = `(1 - (post.embedding::vector <=> CAST('${lit}' AS vector(512))))`; | ||
|
|
||
| // get IDs and scores with threshold filter + pagination | ||
| const qb = this.repository | ||
| .createQueryBuilder("post") | ||
| .select("post.id", "id") | ||
| .addSelect(scoreExpr, "score") | ||
| .where("post.embedding IS NOT NULL") | ||
| .andWhere("post.archive = false") | ||
| .andWhere("post.sold = false") | ||
| .andWhere(`${scoreExpr} >= :threshold`, { threshold }); |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
# First, let's check the PostService.ts file around line 117
head -n 130 src/services/PostService.ts | tail -n 25Repository: cuappdev/resell-backend
Length of output: 1079
🏁 Script executed:
# Let's also check the full PostRepository method
sed -n '555,610p' src/repositories/PostRepository.tsRepository: cuappdev/resell-backend
Length of output: 2059
🏁 Script executed:
# Check if there are any other references to failed embeddings
rg -A 3 -B 3 "embedding.*\[\]" src/services/PostService.tsRepository: cuappdev/resell-backend
Length of output: 1088
🏁 Script executed:
# Let me verify pgvector behavior with empty arrays and dimension mismatches
# First, check if there's any pgvector-related documentation or version info
rg -i "pgvector|vector" package.jsonRepository: cuappdev/resell-backend
Length of output: 92
🏁 Script executed:
# Check if there are database migrations or schema definitions
fd -t f "migration\|schema" --type f | head -20Repository: cuappdev/resell-backend
Length of output: 49
🏁 Script executed:
# Look for any test files or examples showing pgvector usage
rg -l "embedding.*vector" --type ts --type js | head -10Repository: cuappdev/resell-backend
Length of output: 208
🏁 Script executed:
# Check the pgvector migration to see the schema
cat src/migrations/1740007049335-pgvector.tsRepository: cuappdev/resell-backend
Length of output: 1328
🏁 Script executed:
# Look for RequestRepository to see how it uses embeddings
grep -A 20 "embedding" src/repositories/RequestRepository.ts | head -40Repository: cuappdev/resell-backend
Length of output: 1286
🏁 Script executed:
# Get the full findSimilarRequests method from RequestRepository
sed -n '/findSimilarRequests/,/^ public async/p' src/repositories/RequestRepository.ts | head -40Repository: cuappdev/resell-backend
Length of output: 1054
🏁 Script executed:
# Check the exact dimensions being used - search for vector casting patterns
rg "vector\([0-9]+\)" src/Repository: cuappdev/resell-backend
Length of output: 719
🏁 Script executed:
# Let's verify the exact behavior - search for how embeddings are used elsewhere
rg -B 5 -A 5 "embedding.*vector" src/ | head -50Repository: cuappdev/resell-backend
Length of output: 3780
🏁 Script executed:
# Let me verify the array_length behavior and confirm the exact behavior
# Check if there are any array operations on embeddings
rg "array_length" src/Repository: cuappdev/resell-backend
Length of output: 49
🏁 Script executed:
# Let's check if there are any other places where empty embeddings might cause issues
rg -B 3 -A 3 "embedding.*\[\]|\\[\\].*embedding" src/ | head -60Repository: cuappdev/resell-backend
Length of output: 3069
🏁 Script executed:
# Let me search for how the embedding centroid is computed to understand the dimensions
rg -B 5 "centroidEmbedding" src/ | head -50Repository: cuappdev/resell-backend
Length of output: 873
🏁 Script executed:
# Check SearchService to see how it handles empty embeddings
sed -n '/let embedding = "\[\]"/,/catch/p' src/services/SearchService.tsRepository: cuappdev/resell-backend
Length of output: 1744
🏁 Script executed:
# Let me check where findSimilarPostsForEvent is called to understand the centroid computation
rg "findSimilarPostsForEvent" src/ -B 2 -A 2Repository: cuappdev/resell-backend
Length of output: 763
🏁 Script executed:
# Let me also check if centroidEmbedding could have variable dimensions
rg -B 10 "findSimilarPostsForEvent" src/services/ | head -40Repository: cuappdev/resell-backend
Length of output: 839
🏁 Script executed:
# Find the computeCentroid method to understand dimensions
rg -A 15 "computeCentroid" src/services/EventPostService.tsRepository: cuappdev/resell-backend
Length of output: 963
🏁 Script executed:
# Also let's check the postsWithEmbeddings filtering
rg -B 10 "postsWithEmbeddings.length === 0" src/services/EventPostService.ts | head -30Repository: cuappdev/resell-backend
Length of output: 450
🏁 Script executed:
# Let me verify the exact error that would occur with empty embeddings
# Check if there are any error handling or logs related to dimension mismatches
rg "dimension|mismatch" src/ -iRepository: cuappdev/resell-backend
Length of output: 241
🏁 Script executed:
# Let me also check if other similar vector operations handle empty arrays
rg -B 3 -A 3 "::vector" src/repositories/ | head -60Repository: cuappdev/resell-backend
Length of output: 2816
🌐 Web query:
pgvector empty array cast vector dimension mismatch PostgreSQL error
💡 Result:
The PostgreSQL error "different vector dimensions" in pgvector occurs when attempting operations (e.g., distance calculations like <->, <=> ) between vectors of mismatched dimensions, such as a stored vector in a fixed-dimension column like vector(1536) and a query vector with fewer elements. Casting an empty array like '[]' or ARRAY[]::vector(N) to a typed vector(N) fails with a dimension mismatch (0 vs N) because pgvector requires exactly N finite float elements; empty arrays have 0 dimensions. To represent a zero vector (all elements 0.0), use '[0:1536]' or generate an array of N zeros, e.g., array_fill(0.0::real, ARRAY[N])::vector(N). For variable-dimension columns (vector without typmod), filter by vector_dims(vec) = N before distance ops to avoid mismatches. Use NULL for no vector instead of empty arrays, as pgvector vectors cannot contain NULL/NaN/Inf elements.
Citations:
- 1: PGVectorStore: Error: different vector dimensions 4 and 1 langchain-ai/langchainjs#3509
- 2: https://stackoverflow.com/questions/77883843/how-to-deal-with-different-vector-dimensions-for-embeddings-and-search-with-pgve
- 3: 'Error: column does not have dimensions' when creating index pgvector/pgvector#375
- 4: vector_eq raises ERROR instead of returning FALSE. Is that intended behavior? pgvector/pgvector#451
- 5: https://github.com/pgvector/pgvector
- 6: Querying variable dimension vector with pgvector 0.5.1 pgvector/pgvector#426
Add array_length guard to exclude empty embeddings before vector cast.
Empty embeddings are stored as [] (e.g., from failed computations in PostService.ts), and PostgreSQL's IS NOT NULL check includes empty arrays. When the query attempts to cast an empty array to vector(512), pgvector fails with a dimension mismatch error (0 vs 512), crashing the entire query.
Required fix
public async findSimilarPostsForEvent(
centroidEmbedding: number[],
excludePostIds: string[],
threshold: number,
limit: number = 100,
): Promise<{ post: PostModel; score: number }[]> {
const lit = `[${centroidEmbedding.join(",")}]`;
const scoreExpr = `(1 - (post.embedding::vector <=> CAST('${lit}' AS vector(512))))`;
const qb = this.repository
.createQueryBuilder("post")
.select("post.id", "id")
.addSelect(scoreExpr, "score")
.where("post.embedding IS NOT NULL")
+ .andWhere("array_length(post.embedding, 1) = 512")
.andWhere("post.archive = false")
.andWhere("post.sold = false")
.andWhere(`${scoreExpr} >= :threshold`, { threshold });Also apply this same guard to other vector distance queries in PostRepository (findSimilarPosts, findSimilarPostsForEvent in the second ordering clause) and RequestRepository (findSimilarRequests).
🧰 Tools
🪛 ESLint
[error] 559-559: Type number trivially inferred from a number literal, remove type annotation.
(@typescript-eslint/no-inferrable-types)
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@src/repositories/PostRepository.ts` around lines 555 - 572, In
findSimilarPostsForEvent add a guard to exclude empty embeddings before casting
by ensuring array_length(post.embedding,1) = 512 (or >0 / =512 as appropriate)
in the WHERE clauses that use CAST(... AS vector(512)) so empty arrays ([]) are
filtered out; update the scoreExpr usage and the .andWhere chain in the method
to include this array_length check alongside "post.embedding IS NOT NULL". Apply
the same array_length guard to the other vector-distance queries:
PostRepository.findSimilarPosts (both places it builds/uses the vector cast) and
the second ordering clause in PostRepository.findSimilarPostsForEvent, and to
RequestRepository.findSimilarRequests so pgvector never receives an empty array
and avoids the dimension mismatch error.
Comment on lines
+41
to
+68
| const skip = (page - 1) * limit; | ||
|
|
||
| // Fetch paginated relationships and total count in parallel | ||
| const [relationships, total] = await Promise.all([ | ||
| eventPostRepo.getPostsForEvent(eventTagId, source, skip, limit), | ||
| eventPostRepo.getPostCountForEvent(eventTagId, source), | ||
| ]); | ||
|
|
||
| // Filter inactive and blocked users | ||
| const userWithBlockedInfo = await userRepo.getUserWithBlockedInfo(user.firebaseUid); | ||
| const blockedUids = new Set( | ||
| userWithBlockedInfo?.blocking?.map(u => u.firebaseUid) ?? [], | ||
| ); | ||
|
|
||
| const filtered = relationships.filter(r => | ||
| r.post?.user?.isActive && | ||
| !blockedUids.has(r.post.user.firebaseUid), | ||
| ); | ||
|
|
||
| // Attach source + relevanceScore to each post | ||
| const posts: PostWithSource[] = filtered.map(r => | ||
| Object.assign(r.post, { | ||
| source: r.source, | ||
| relevanceScore: r.relevanceScore, | ||
| }), | ||
| ); | ||
|
|
||
| return { posts, total, page, limit }; |
There was a problem hiding this comment.
Apply visibility filters before paginating and counting.
getPostsForEvent() and getPostCountForEvent() run before blocked and inactive authors are removed, so hidden rows still consume skip/limit slots and total can advertise pages the caller can never fill. Push the active/block predicates into the repository query/count instead of filtering after pagination here.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@src/services/EventService.ts` around lines 41 - 68, The current code applies
active/blocked filtering after calling eventPostRepo.getPostsForEvent and
getPostCountForEvent, which causes pagination and total to include rows that
should be hidden; modify the repo calls so filtering happens in the DB layer:
retrieve blocked UIDs from userRepo.getUserWithBlockedInfo(user.firebaseUid)
first, then call eventPostRepo.getPostsForEvent(eventTagId, source, skip, limit,
{excludeUids: blockedUids, requireActiveUser: true}) and
eventPostRepo.getPostCountForEvent(eventTagId, source, {excludeUids:
blockedUids, requireActiveUser: true}) (or add equivalent parameters/overloads)
so both the SELECT and COUNT apply the same predicates; remove the post-level
isActive/blocked filtering in EventService and keep only mapping to
PostWithSource for attaching source/relevanceScore.
Comment on lines
+553
to
562
| const tagsToRemoveNames = new Set(removeEventTagsRequest.eventTags); | ||
| const removedTags = post.eventTags?.filter(tag => tagsToRemoveNames.has(tag.name)) || []; | ||
|
|
||
| const eventPostRepository = Repositories.eventPost(transactionalEntityManager); | ||
| for (const tag of removedTags) { | ||
| await eventPostRepository.deleteRelationship(post.id, tag.id); | ||
| } | ||
|
|
||
| post.eventTags = post.eventTags?.filter(tag => !tagsToRemoveNames.has(tag.name)) || []; | ||
| return await postRepository.savePost(post); |
There was a problem hiding this comment.
Clear similarity rows when the last manual tag disappears.
This only deletes the direct (postId, eventTagId) row. If that was the last user-tagged seed for the event, the remaining similarity rows stay behind; src/services/EventPostService.ts Lines 29-41 and 56-60 currently return before deleteRelationshipsBySourceForEvent(...) when no user-tagged posts or embeddings remain. The feed can keep showing stale event matches indefinitely.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@src/services/PostService.ts` around lines 553 - 562, When removing the manual
event tags in PostService (use removeEventTagsRequest.eventTags, removedTags,
eventPostRepository.deleteRelationship and post.eventTags), ensure you also
clear any leftover similarity rows when the last user-tagged seed for that event
is gone: after deleting each (post.id, tag.id) and updating post.eventTags,
check if there are no remaining user-tagged posts or embeddings for the event
and then call the repository method that removes similarity rows
(eventPostRepository.deleteRelationshipsBySourceForEvent or the EventPostService
equivalent) for that event/source before saving the post; this mirrors the
cleanup logic found in EventPostService.ts (deleteRelationshipsBySourceForEvent)
to prevent stale event matches.
JoshD94
requested changes
Mar 26, 2026
JoshD94
left a comment
JoshD94
left a comment
There was a problem hiding this comment.
lgtm after coderabbit fixes are added
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
EventPostModelto track post-event relationships with source (user/similarity/nlp_context) and relevanceScoreGET /events/available-for-tagging/andGET /events/:eventId/posts/endpointsTest Coverage
EventTest.test.tscovering event feed reads, pagination, source filtering, and user blockingSummary by CodeRabbit
Release Notes
New Features
Documentation