# Atlas Stream Processing per-event vector search enrichment > Source-anchored research for Atlas Stream Processing per-event vector search enrichment. Run status: COMPLETED. Parent: [MongoDB Atlas Stream Processing](https://llms-explorer.com/tree/mongodb-atlas-stream-processing/) · 4 facets · 18 facts · page: https://llms-explorer.com/tree/atlas-stream-processing-per-event-vector-search-enrichment/ ## Definitions - $vectorSearch is a middle stage: it queries per input document and attaches an array under as; no matches produce an empty array. (medium confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) ## Problems, failure modes and limitations - A missing target index causes MongoServerError when search first executes; the processor enters an error state and stops. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - limit exceeding numCandidates, or unsupported filter operators, produces PlanExecutor errors for every document at runtime. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - Non-Atlas from.connectionName connections fail parsing. Empty as or metadata-name collisions cause StreamProcessorInvalidOptions at startup. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - QuerySentinel documents ASP write-rate aggregation separately from SimilarIncidentAgent vector retrieval. Its reported vector-dimension mismatch is an application report, not evidence of native ASP-stage failure semantics. (low confidence) (verified as of 2026-10-27) — [source](https://github.com/pulkit6732/QuerySentinel) ## Comparisons and alternatives - The media architecture enriches clicks with article metadata, summarizes sessions, and serves semantic recommendations afterward; this differs from querying an index for each incoming event. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/architecture/current/solutions-library/stream-media-personalization.md) - Gorman's March 28, 2025 example calls an LLM through $https; its vector-memory discussion proposes an extension rather than demonstrating the native $vectorSearch stage. (low confidence) — [source](https://kennygorman.dev/2025/03/28/adding-probabilistic-reasoning-to-mongodb-atlas-stream-proce/) - Confluent warns that external stream enrichment can overload dependencies during replay, propagate latency/failures, and retrieve current rather than historical state. These are useful ASP validation questions, not established native-stage guarantees. (low confidence) — [source](https://docs.confluent.io/cloud/current/ai/external-tables/overview.html) ## Facts and statements - queryVector and query.text are mutually exclusive; field expressions enable event-specific vectors or automatically embedded text. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - from requires an Atlas connection. No matches produce an empty results array. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - Wrong runtime queryVector/query.text types send only that event to the configured DLQ, or drop it; later events continue. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - An empty as or metadata-name collision raises StreamProcessorInvalidOptions at startup. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - let binds event values for the forwarded result pipeline; Native Reranking must be enabled for $rerank. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - The continuously updating RAG architecture uses external embedding workers and a separate query client. It does not demonstrate native per-event vector retrieval inside ASP. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/architecture/current/solutions-library/rag-applications.md) - QuerySentinel documents ASP write-rate aggregation and separate agent vector retrieval. Its README does not demonstrate the native ASP vector-search stage. (low confidence) (verified as of 2026-10-27) — [source](https://github.com/pulkit6732/QuerySentinel) - MongoDB announced the ASP vector-search stage as generally available on August 13, 2026. (low confidence) (verified as of 2026-08-13) — [source](https://www.mongodb.com/company/blog/product-release-announcements/closing-gap-between-agentic-promise-impact-industry-leading-retrieval) - pipeline processes retrieved results before attachment; it is forwarded verbatim, with per-event values passed through let variables. (low confidence) (verified as of 2026-10-27) — [source](https://www.mongodb.com/docs/atlas/atlas-stream-processing/sp-agg-vectorsearch/) - The MongoDB partner streaming example writes ASP rollups to stream_events, then uses a separate change-stream ingestion worker to populate knowledge_base. It illustrates searchable ingestion, not native per-event retrieval enrichment. (low confidence) (verified as of 2026-10-27) — [source](https://github.com/mongodb-partners/mongodb-langchain-agentic-ai/blob/main/docs/streaming.md)