Skip to content

Embedding & Vector Search Settings

This chapter covers the embedding model, reranker, and vector database settings — the core of knowledge search.

Components

Knowledge search relies on three components working together.

Component Korean Role
Embedding Model 임베딩 모델 Convert documents and queries into vectors
Vector Database 벡터 데이터베이스 Store vectors and perform similarity search
Reranker 리랭커 Re-rank initial retrieval results with finer precision

Embedding / Search Settings — three tabs (Embedding Model / Reranker / Vector DB) and registered models

Registering an Embedding Model

Select Admin → Environment → Search / Embedding in the left sidebar, then switch to the Embedding Model tab.

  1. Click + Add Model
  2. Enter:
    • Model name
    • Provider (OpenAI / vLLM, etc.)
    • Model identifier (e.g., text-embedding-3-large, internal model name)
    • Dimension — the vector dimensionality output by the model (e.g., 1536)
  3. Click Test Connection → confirm an actual embedding call succeeds
  4. Save

Re-embedding Required When Changing Dimension

If you switch to a model with a different dimension than an already-embedded collection, the existing vectors cannot be used. The whole collection must be re-embedded (consumes time and cost).

Registering a Reranker

A reranker re-orders initial retrieval results (e.g., top 50) for higher accuracy. Effective for accuracy improvements but increases response latency.

  1. Reranker tab → + Add Reranker
  2. Enter model name, provider, identifier
  3. Test Connection → Save

Reranker tab — *Active Provider* card at the top with *Available Reranker Providers* (Sentence Transformers / vLLM Server) below, and configuration items (Reranker Model / Device) for the active provider

The reranker is optional. Whether to use it and the threshold are configured per-collection or per-agentflow.

Connecting a Vector Database

The default supported engine is Qdrant.

  1. Vector Database tab → Connection Settings
  2. Enter:
    • Host: e.g., qdrant.internal.example.com
    • Port: default 6333 (HTTP), 6334 (gRPC)
    • API key (if authentication is enabled)
  3. Test Connection → Save

Vector Database tab — *Active Provider* card at the top and Qdrant connection settings (vector dimension, host, port, API key, gRPC toggle)

Disk Monitoring

The vector database consumes disk quickly as collections grow. Periodically check disk usage in System Monitor and configure threshold alerts.

Operational Checks

Item Frequency Method
Vector DB disk usage Weekly System Monitor
Embedding call failure rate Weekly Audit log or LLM provider console
Reranker response time Monthly Chat response time stats
Similarity search quality Quarterly Sample query result review

Contact

For questions about embedding and vector search, please contact the Xgen Solution Administrator.