Vector Search and RAG Implementation for Enhanced User Search Experience

couchbase · Finance · 2023 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Semantic search, RAG, and vector-search education

Couchbase's presentation argues that many application search experiences fail because keyword search cannot understand semantic intent, while production teams still need to know when vector search is worth its extra cost.

Constraints:

• The source is a vendor educational talk, not a first-party deep dive from Revolut or Scen.it.
• Search quality matters quickly because users decide whether a search experience works within seconds.
• Vector search is more computationally expensive and harder to maintain than keyword search.
• RAG can reduce hallucinations but does not eliminate the need to verify generated outputs.
• The case-study metrics are compelling but should be treated as cited examples rather than independently audited measurements.

Architecture

The talk presents vector search as a reusable semantic-retrieval layer: user queries or business events are embedded into vectors, compared against embedded documents or transactions, and then used either directly for semantic search/anomaly detection or as retrieved context for a RAG-style generative AI response.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  user-query-or-business-event["User query or business event"] --> embedding-model["Embedding model"]
  embedding-model --> vector-index["Vector index"]
  vector-index --> similarity-search["Similarity search"]
  similarity-search --> rag-context-builder["RAG context builder"]
  rag-context-builder --> generative-ai-agent["Generative AI agent"]
  vector-index --> anomaly-detector["Anomaly detector"]

Components

Data flow

A query, document, video clip, or transaction is transformed into an embedding; vector similarity finds semantically related or anomalous items; the retrieved items either become search results, fraud signals, or context for a generative AI agent; and teams reserve keyword search for simple exact-match cases.

• Vector search finds similar documents • Those documents provide context • The context augments AI agents for more accurate responses"

— summary, stated

Why this architecture

The architecture separates semantic matching from generation so teams can use vector search where language, recommendations, anomaly detection, or cross-modal search need meaning-aware retrieval, while avoiding it where simple keyword search is cheaper and sufficient.

Models used: Embedding models · Vector search · Retrieval-augmented generation · Generative AI agents

Scale

• Bad search experiences can lose nearly 90% of users.
  “The speaker notes that nearly 90% of users will not return to an application after a bad experience, and users typically decide whether a search experience is effective within 12-14 seconds.”summary, stated
• Revolut's Sherlock system is described as serving 12 million users.

  "Production Performance:

• Serves 12 million users • Processes fraud detection in under 50 milliseconds • Catches 96% of fraudulent transactions • Saved customers over $3 million in one year"

— summary, stated

• Scen.it applies search over more than 500,000 video clips.
  “Companies need to find relevant video clips from a library of over 500,000 clips for their marketing campaigns.”summary, stated

Alternatives considered (roads not taken)

Use vector search for every search problem

Rejected because: The talk explicitly recommends keyword search for exact-match and simple-dataset cases because it is cheaper and easier to maintain.

Let generative AI answer without retrieval context

Rejected because: The RAG framing uses vector search to supply grounding context and reduce hallucinations.

Patterns

rag-retrieval (retrieval)

Vector search retrieves semantically similar documents that become grounding context for generative AI answers.

"RAG is described as a three-step process:

• Vector search finds similar documents • Those documents provide context • The context augments AI agents for more accurate responses"

— summary, stated

retrieval-ranking (search)

The system ranks candidates by vector similarity, or dissimilarity for anomaly detection, rather than exact keyword overlap.

semantic-caching (optimization)

The case studies show embeddings as a reusable semantic representation across text, images, audio, and transactions, though the source does not describe cache invalidation mechanics.

Tradeoffs

Reserve vector search for semantic and anomaly-detection use cases

Chose Vector search where meaning-aware retrieval is needed over Universal vector search for all queries.

Rationale: Keyword search is cheaper and easier when exact matches or simple datasets satisfy the need.

Use RAG context to reduce hallucinations

Chose Retrieved documents augmenting AI agents over Unretrieved model responses.

Rationale: RAG grounds answers in similar documents while still requiring verification.

Gotchas (what bit them)

Vector search can be overkill for exact-match search.

Root cause: Semantic vector retrieval carries higher compute and maintenance costs than keyword search.

Resolution: Use vector search for recommendations, anomaly detection, and semantic language cases, while keeping keyword search for simple exact-match datasets.

• Exact match retrieval where keyword search suffices • Simple datasets"

— summary, stated

RAG does not remove hallucination risk.

Root cause: Retrieved context improves grounding, but generative systems can still produce incorrect outputs.

Resolution: Verify AI-generated outputs before production use and treat RAG as hallucination reduction rather than elimination.

Vendor educational case studies provide limited architecture depth.

Root cause: The talk references production examples but does not present the full internal designs or measurement methodology.

Resolution: Use the case metrics as directional evidence and avoid inferring unshown implementation details.

Evaluation

The talk relies on production case-study metrics for Revolut and Scen.it plus qualitative guidance about when vector search is appropriate.

Metrics: 12 million Revolut users served · Under 50 milliseconds fraud detection latency · 96% fraudulent transactions caught · Over $3 million saved in one year · 500,000+ Scen.it video clips searchable by text

Outcomes

• The talk shows vector search as useful for semantic search, RAG grounding, fraud anomaly detection, and multimedia search.
• Revolut's cited Sherlock metrics include 96% fraud catch rate, sub-50ms detection, and over $3 million saved in one year.
• Scen.it's cited use case enables natural-language search across a large video-clip library.

The production examples span strict real-time fraud latency for 12 million users and search across more than 500,000 video clips.